diff --git a/sleep_assessment/cases/case10/case10_01.json b/sleep_assessment/cases/case10/case10_01.json new file mode 100644 index 0000000000000000000000000000000000000000..24217384dbbd5bd185bce2e3dec6a52675d8cd84 --- /dev/null +++ b/sleep_assessment/cases/case10/case10_01.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Please use the provided sleep-stage label file to calculate two N3-related sleep metrics for the whole night: the number of N3 bouts, and the mean N3 bout length in minutes. Here, a bout is defined as a contiguous N3 segment with length greater than or equal to 2 epochs. Please clearly report both final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two values from the agent's report:\n1) the number of N3 bouts across the night\n2) the mean N3 bout length in minutes\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"n3_bout_count\" and \"n3_mean_bout_length_min\".\n5. \"n3_bout_count\" must be an INTEGER or null.\n6. \"n3_mean_bout_length_min\" must be a FLOAT or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"n3_bout_count\": , \"n3_mean_bout_length_min\": }", + "metrics": [ + { + "metric_id": "n3_bout_count_accuracy", + "type": "numeric_check", + "target_key": "n3_bout_count", + "weight": 50, + "params": { + "gt_value": 14, + "tolerance": 0.7000000000000001 + } + }, + { + "metric_id": "n3_mean_bout_length_accuracy", + "type": "numeric_check", + "target_key": "n3_mean_bout_length_min", + "weight": 50, + "params": { + "gt_value": 8.142857142857142, + "tolerance": 0.40714285714285714 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case10/case10_02.json b/sleep_assessment/cases/case10/case10_02.json new file mode 100644 index 0000000000000000000000000000000000000000..eceda8edf330fd37794d32dfca303a2a4422a915 --- /dev/null +++ b/sleep_assessment/cases/case10/case10_02.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_02.npy", + "instruction": "Please use the provided sleep-stage label file to calculate two N3-related sleep metrics for the whole night: the number of N3 bouts, and the mean N3 bout length in minutes. Here, a bout is defined as a contiguous N3 segment with length greater than or equal to 2 epochs. Please clearly report both final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two values from the agent's report:\n1) the number of N3 bouts across the night\n2) the mean N3 bout length in minutes\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"n3_bout_count\" and \"n3_mean_bout_length_min\".\n5. \"n3_bout_count\" must be an INTEGER or null.\n6. \"n3_mean_bout_length_min\" must be a FLOAT or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"n3_bout_count\": , \"n3_mean_bout_length_min\": }", + "metrics": [ + { + "metric_id": "n3_bout_count_accuracy", + "type": "numeric_check", + "target_key": "n3_bout_count", + "weight": 50, + "params": { + "gt_value": 7, + "tolerance": 0.35000000000000003 + } + }, + { + "metric_id": "n3_mean_bout_length_accuracy", + "type": "numeric_check", + "target_key": "n3_mean_bout_length_min", + "weight": 50, + "params": { + "gt_value": 10.928571428571429, + "tolerance": 0.5464285714285715 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case10/case10_03.json b/sleep_assessment/cases/case10/case10_03.json new file mode 100644 index 0000000000000000000000000000000000000000..6b52c388d79815b64b6254c2e45c342890eed809 --- /dev/null +++ b/sleep_assessment/cases/case10/case10_03.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_03.npy", + "instruction": "Please use the provided sleep-stage label file to calculate two N3-related sleep metrics for the whole night: the number of N3 bouts, and the mean N3 bout length in minutes. Here, a bout is defined as a contiguous N3 segment with length greater than or equal to 2 epochs. Please clearly report both final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two values from the agent's report:\n1) the number of N3 bouts across the night\n2) the mean N3 bout length in minutes\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"n3_bout_count\" and \"n3_mean_bout_length_min\".\n5. \"n3_bout_count\" must be an INTEGER or null.\n6. \"n3_mean_bout_length_min\" must be a FLOAT or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"n3_bout_count\": , \"n3_mean_bout_length_min\": }", + "metrics": [ + { + "metric_id": "n3_bout_count_accuracy", + "type": "numeric_check", + "target_key": "n3_bout_count", + "weight": 50, + "params": { + "gt_value": 5, + "tolerance": 0.25 + } + }, + { + "metric_id": "n3_mean_bout_length_accuracy", + "type": "numeric_check", + "target_key": "n3_mean_bout_length_min", + "weight": 50, + "params": { + "gt_value": 17.2, + "tolerance": 0.86 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case10/case10_04.json b/sleep_assessment/cases/case10/case10_04.json new file mode 100644 index 0000000000000000000000000000000000000000..57dd95a0a0619a0d775ac4ced20c4ef5335ec2bb --- /dev/null +++ b/sleep_assessment/cases/case10/case10_04.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "ISRUC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_04.npy", + "instruction": "Please use the provided sleep-stage label file to calculate two N3-related sleep metrics for the whole night: the number of N3 bouts, and the mean N3 bout length in minutes. Here, a bout is defined as a contiguous N3 segment with length greater than or equal to 2 epochs. Please clearly report both final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two values from the agent's report:\n1) the number of N3 bouts across the night\n2) the mean N3 bout length in minutes\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"n3_bout_count\" and \"n3_mean_bout_length_min\".\n5. \"n3_bout_count\" must be an INTEGER or null.\n6. \"n3_mean_bout_length_min\" must be a FLOAT or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"n3_bout_count\": , \"n3_mean_bout_length_min\": }", + "metrics": [ + { + "metric_id": "n3_bout_count_accuracy", + "type": "numeric_check", + "target_key": "n3_bout_count", + "weight": 50, + "params": { + "gt_value": 8, + "tolerance": 0.4 + } + }, + { + "metric_id": "n3_mean_bout_length_accuracy", + "type": "numeric_check", + "target_key": "n3_mean_bout_length_min", + "weight": 50, + "params": { + "gt_value": 13.375, + "tolerance": 0.6687500000000001 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case10/case10_05.json b/sleep_assessment/cases/case10/case10_05.json new file mode 100644 index 0000000000000000000000000000000000000000..5ba4764acf525a53d9c3df8bcac7b36bfb7385c0 --- /dev/null +++ b/sleep_assessment/cases/case10/case10_05.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_05.npy", + "instruction": "Please use the provided sleep-stage label file to calculate two N3-related sleep metrics for the whole night: the number of N3 bouts, and the mean N3 bout length in minutes. Here, a bout is defined as a contiguous N3 segment with length greater than or equal to 2 epochs. Please clearly report both final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two values from the agent's report:\n1) the number of N3 bouts across the night\n2) the mean N3 bout length in minutes\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"n3_bout_count\" and \"n3_mean_bout_length_min\".\n5. \"n3_bout_count\" must be an INTEGER or null.\n6. \"n3_mean_bout_length_min\" must be a FLOAT or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"n3_bout_count\": , \"n3_mean_bout_length_min\": }", + "metrics": [ + { + "metric_id": "n3_bout_count_accuracy", + "type": "numeric_check", + "target_key": "n3_bout_count", + "weight": 50, + "params": { + "gt_value": 5, + "tolerance": 0.25 + } + }, + { + "metric_id": "n3_mean_bout_length_accuracy", + "type": "numeric_check", + "target_key": "n3_mean_bout_length_min", + "weight": 50, + "params": { + "gt_value": 16.3, + "tolerance": 0.8150000000000001 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case10/case10_06.json b/sleep_assessment/cases/case10/case10_06.json new file mode 100644 index 0000000000000000000000000000000000000000..b7d2b03bc331a15f75f52c9dafe580d472b94cbd --- /dev/null +++ b/sleep_assessment/cases/case10/case10_06.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "HMC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_01.npy", + "instruction": "Please use the provided sleep-stage label file to calculate two N3-related sleep metrics for the whole night: the number of N3 bouts, and the mean N3 bout length in minutes. Here, a bout is defined as a contiguous N3 segment with length greater than or equal to 2 epochs. Please clearly report both final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two values from the agent's report:\n1) the number of N3 bouts across the night\n2) the mean N3 bout length in minutes\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"n3_bout_count\" and \"n3_mean_bout_length_min\".\n5. \"n3_bout_count\" must be an INTEGER or null.\n6. \"n3_mean_bout_length_min\" must be a FLOAT or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"n3_bout_count\": , \"n3_mean_bout_length_min\": }", + "metrics": [ + { + "metric_id": "n3_bout_count_accuracy", + "type": "numeric_check", + "target_key": "n3_bout_count", + "weight": 50, + "params": { + "gt_value": 6, + "tolerance": 0.30000000000000004 + } + }, + { + "metric_id": "n3_mean_bout_length_accuracy", + "type": "numeric_check", + "target_key": "n3_mean_bout_length_min", + "weight": 50, + "params": { + "gt_value": 1.75, + "tolerance": 0.08750000000000001 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case10/case10_07.json b/sleep_assessment/cases/case10/case10_07.json new file mode 100644 index 0000000000000000000000000000000000000000..9da0b796f2e57fba1797120b03661bb694c05ab2 --- /dev/null +++ b/sleep_assessment/cases/case10/case10_07.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "HMC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_02.npy", + "instruction": "Please use the provided sleep-stage label file to calculate two N3-related sleep metrics for the whole night: the number of N3 bouts, and the mean N3 bout length in minutes. Here, a bout is defined as a contiguous N3 segment with length greater than or equal to 2 epochs. Please clearly report both final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two values from the agent's report:\n1) the number of N3 bouts across the night\n2) the mean N3 bout length in minutes\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"n3_bout_count\" and \"n3_mean_bout_length_min\".\n5. \"n3_bout_count\" must be an INTEGER or null.\n6. \"n3_mean_bout_length_min\" must be a FLOAT or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"n3_bout_count\": , \"n3_mean_bout_length_min\": }", + "metrics": [ + { + "metric_id": "n3_bout_count_accuracy", + "type": "numeric_check", + "target_key": "n3_bout_count", + "weight": 50, + "params": { + "gt_value": 14, + "tolerance": 0.7000000000000001 + } + }, + { + "metric_id": "n3_mean_bout_length_accuracy", + "type": "numeric_check", + "target_key": "n3_mean_bout_length_min", + "weight": 50, + "params": { + "gt_value": 7.821428571428571, + "tolerance": 0.39107142857142857 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case10/case10_08.json b/sleep_assessment/cases/case10/case10_08.json new file mode 100644 index 0000000000000000000000000000000000000000..26cb2739386fbd0ac8f4d86d86ab7cd0b5392c55 --- /dev/null +++ b/sleep_assessment/cases/case10/case10_08.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "HMC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_03.npy", + "instruction": "Please use the provided sleep-stage label file to calculate two N3-related sleep metrics for the whole night: the number of N3 bouts, and the mean N3 bout length in minutes. Here, a bout is defined as a contiguous N3 segment with length greater than or equal to 2 epochs. Please clearly report both final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two values from the agent's report:\n1) the number of N3 bouts across the night\n2) the mean N3 bout length in minutes\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"n3_bout_count\" and \"n3_mean_bout_length_min\".\n5. \"n3_bout_count\" must be an INTEGER or null.\n6. \"n3_mean_bout_length_min\" must be a FLOAT or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"n3_bout_count\": , \"n3_mean_bout_length_min\": }", + "metrics": [ + { + "metric_id": "n3_bout_count_accuracy", + "type": "numeric_check", + "target_key": "n3_bout_count", + "weight": 50, + "params": { + "gt_value": 14, + "tolerance": 0.7000000000000001 + } + }, + { + "metric_id": "n3_mean_bout_length_accuracy", + "type": "numeric_check", + "target_key": "n3_mean_bout_length_min", + "weight": 50, + "params": { + "gt_value": 5.321428571428571, + "tolerance": 0.26607142857142857 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case10/case10_09.json b/sleep_assessment/cases/case10/case10_09.json new file mode 100644 index 0000000000000000000000000000000000000000..3045b40d08e9a956a6ecff4c2b50980b0c5e4b9d --- /dev/null +++ b/sleep_assessment/cases/case10/case10_09.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "HMC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_04.npy", + "instruction": "Please use the provided sleep-stage label file to calculate two N3-related sleep metrics for the whole night: the number of N3 bouts, and the mean N3 bout length in minutes. Here, a bout is defined as a contiguous N3 segment with length greater than or equal to 2 epochs. Please clearly report both final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two values from the agent's report:\n1) the number of N3 bouts across the night\n2) the mean N3 bout length in minutes\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"n3_bout_count\" and \"n3_mean_bout_length_min\".\n5. \"n3_bout_count\" must be an INTEGER or null.\n6. \"n3_mean_bout_length_min\" must be a FLOAT or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"n3_bout_count\": , \"n3_mean_bout_length_min\": }", + "metrics": [ + { + "metric_id": "n3_bout_count_accuracy", + "type": "numeric_check", + "target_key": "n3_bout_count", + "weight": 50, + "params": { + "gt_value": 8, + "tolerance": 0.4 + } + }, + { + "metric_id": "n3_mean_bout_length_accuracy", + "type": "numeric_check", + "target_key": "n3_mean_bout_length_min", + "weight": 50, + "params": { + "gt_value": 3.9375, + "tolerance": 0.19687500000000002 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case10/case10_10.json b/sleep_assessment/cases/case10/case10_10.json new file mode 100644 index 0000000000000000000000000000000000000000..7e4f037fa0c0e8342e0cc0a61a078525ae26cfe7 --- /dev/null +++ b/sleep_assessment/cases/case10/case10_10.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "HMC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_05.npy", + "instruction": "Please use the provided sleep-stage label file to calculate two N3-related sleep metrics for the whole night: the number of N3 bouts, and the mean N3 bout length in minutes. Here, a bout is defined as a contiguous N3 segment with length greater than or equal to 2 epochs. Please clearly report both final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two values from the agent's report:\n1) the number of N3 bouts across the night\n2) the mean N3 bout length in minutes\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"n3_bout_count\" and \"n3_mean_bout_length_min\".\n5. \"n3_bout_count\" must be an INTEGER or null.\n6. \"n3_mean_bout_length_min\" must be a FLOAT or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"n3_bout_count\": , \"n3_mean_bout_length_min\": }", + "metrics": [ + { + "metric_id": "n3_bout_count_accuracy", + "type": "numeric_check", + "target_key": "n3_bout_count", + "weight": 50, + "params": { + "gt_value": 10, + "tolerance": 0.5 + } + }, + { + "metric_id": "n3_mean_bout_length_accuracy", + "type": "numeric_check", + "target_key": "n3_mean_bout_length_min", + "weight": 50, + "params": { + "gt_value": 7.15, + "tolerance": 0.35750000000000004 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case10/case10_11.json b/sleep_assessment/cases/case10/case10_11.json new file mode 100644 index 0000000000000000000000000000000000000000..b1d3b5e09e0c64393a9ad7a1e3509983a81151ea --- /dev/null +++ b/sleep_assessment/cases/case10/case10_11.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_01.npy", + "instruction": "Please use the provided sleep-stage label file to calculate two N3-related sleep metrics for the whole night: the number of N3 bouts, and the mean N3 bout length in minutes. Here, a bout is defined as a contiguous N3 segment with length greater than or equal to 2 epochs. Please clearly report both final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two values from the agent's report:\n1) the number of N3 bouts across the night\n2) the mean N3 bout length in minutes\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"n3_bout_count\" and \"n3_mean_bout_length_min\".\n5. \"n3_bout_count\" must be an INTEGER or null.\n6. \"n3_mean_bout_length_min\" must be a FLOAT or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"n3_bout_count\": , \"n3_mean_bout_length_min\": }", + "metrics": [ + { + "metric_id": "n3_bout_count_accuracy", + "type": "numeric_check", + "target_key": "n3_bout_count", + "weight": 50, + "params": { + "gt_value": 15, + "tolerance": 0.75 + } + }, + { + "metric_id": "n3_mean_bout_length_accuracy", + "type": "numeric_check", + "target_key": "n3_mean_bout_length_min", + "weight": 50, + "params": { + "gt_value": 4.366666666666666, + "tolerance": 0.21833333333333332 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case10/case10_12.json b/sleep_assessment/cases/case10/case10_12.json new file mode 100644 index 0000000000000000000000000000000000000000..ffe41d72b8fc8cbeda9635f65eb0ad5b04498bdb --- /dev/null +++ b/sleep_assessment/cases/case10/case10_12.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_02.npy", + "instruction": "Please use the provided sleep-stage label file to calculate two N3-related sleep metrics for the whole night: the number of N3 bouts, and the mean N3 bout length in minutes. Here, a bout is defined as a contiguous N3 segment with length greater than or equal to 2 epochs. Please clearly report both final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two values from the agent's report:\n1) the number of N3 bouts across the night\n2) the mean N3 bout length in minutes\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"n3_bout_count\" and \"n3_mean_bout_length_min\".\n5. \"n3_bout_count\" must be an INTEGER or null.\n6. \"n3_mean_bout_length_min\" must be a FLOAT or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"n3_bout_count\": , \"n3_mean_bout_length_min\": }", + "metrics": [ + { + "metric_id": "n3_bout_count_accuracy", + "type": "numeric_check", + "target_key": "n3_bout_count", + "weight": 50, + "params": { + "gt_value": 12, + "tolerance": 0.6000000000000001 + } + }, + { + "metric_id": "n3_mean_bout_length_accuracy", + "type": "numeric_check", + "target_key": "n3_mean_bout_length_min", + "weight": 50, + "params": { + "gt_value": 1.7916666666666667, + "tolerance": 0.08958333333333335 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case10/case10_13.json b/sleep_assessment/cases/case10/case10_13.json new file mode 100644 index 0000000000000000000000000000000000000000..ffab9541eed4b242a9cb1e8c7961d91f46b97908 --- /dev/null +++ b/sleep_assessment/cases/case10/case10_13.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_03.npy", + "instruction": "Please use the provided sleep-stage label file to calculate two N3-related sleep metrics for the whole night: the number of N3 bouts, and the mean N3 bout length in minutes. Here, a bout is defined as a contiguous N3 segment with length greater than or equal to 2 epochs. Please clearly report both final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two values from the agent's report:\n1) the number of N3 bouts across the night\n2) the mean N3 bout length in minutes\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"n3_bout_count\" and \"n3_mean_bout_length_min\".\n5. \"n3_bout_count\" must be an INTEGER or null.\n6. \"n3_mean_bout_length_min\" must be a FLOAT or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"n3_bout_count\": , \"n3_mean_bout_length_min\": }", + "metrics": [ + { + "metric_id": "n3_bout_count_accuracy", + "type": "numeric_check", + "target_key": "n3_bout_count", + "weight": 50, + "params": { + "gt_value": 18, + "tolerance": 0.9 + } + }, + { + "metric_id": "n3_mean_bout_length_accuracy", + "type": "numeric_check", + "target_key": "n3_mean_bout_length_min", + "weight": 50, + "params": { + "gt_value": 8.222222222222221, + "tolerance": 0.4111111111111111 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case10/case10_14.json b/sleep_assessment/cases/case10/case10_14.json new file mode 100644 index 0000000000000000000000000000000000000000..67b65546e4dcdfc099e5522c62630141204b5127 --- /dev/null +++ b/sleep_assessment/cases/case10/case10_14.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_04.npy", + "instruction": "Please use the provided sleep-stage label file to calculate two N3-related sleep metrics for the whole night: the number of N3 bouts, and the mean N3 bout length in minutes. Here, a bout is defined as a contiguous N3 segment with length greater than or equal to 2 epochs. Please clearly report both final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two values from the agent's report:\n1) the number of N3 bouts across the night\n2) the mean N3 bout length in minutes\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"n3_bout_count\" and \"n3_mean_bout_length_min\".\n5. \"n3_bout_count\" must be an INTEGER or null.\n6. \"n3_mean_bout_length_min\" must be a FLOAT or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"n3_bout_count\": , \"n3_mean_bout_length_min\": }", + "metrics": [ + { + "metric_id": "n3_bout_count_accuracy", + "type": "numeric_check", + "target_key": "n3_bout_count", + "weight": 50, + "params": { + "gt_value": 17, + "tolerance": 0.8500000000000001 + } + }, + { + "metric_id": "n3_mean_bout_length_accuracy", + "type": "numeric_check", + "target_key": "n3_mean_bout_length_min", + "weight": 50, + "params": { + "gt_value": 8.588235294117647, + "tolerance": 0.4294117647058824 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case10/case10_15.json b/sleep_assessment/cases/case10/case10_15.json new file mode 100644 index 0000000000000000000000000000000000000000..eb84379289a48565ed84c99a6352054ba9601ecc --- /dev/null +++ b/sleep_assessment/cases/case10/case10_15.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_05.npy", + "instruction": "Please use the provided sleep-stage label file to calculate two N3-related sleep metrics for the whole night: the number of N3 bouts, and the mean N3 bout length in minutes. Here, a bout is defined as a contiguous N3 segment with length greater than or equal to 2 epochs. Please clearly report both final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two values from the agent's report:\n1) the number of N3 bouts across the night\n2) the mean N3 bout length in minutes\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"n3_bout_count\" and \"n3_mean_bout_length_min\".\n5. \"n3_bout_count\" must be an INTEGER or null.\n6. \"n3_mean_bout_length_min\" must be a FLOAT or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"n3_bout_count\": , \"n3_mean_bout_length_min\": }", + "metrics": [ + { + "metric_id": "n3_bout_count_accuracy", + "type": "numeric_check", + "target_key": "n3_bout_count", + "weight": 50, + "params": { + "gt_value": 13, + "tolerance": 0.65 + } + }, + { + "metric_id": "n3_mean_bout_length_accuracy", + "type": "numeric_check", + "target_key": "n3_mean_bout_length_min", + "weight": 50, + "params": { + "gt_value": 3.4615384615384617, + "tolerance": 0.1730769230769231 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case10/case10_16.json b/sleep_assessment/cases/case10/case10_16.json new file mode 100644 index 0000000000000000000000000000000000000000..ad7244442f494477edea581b5a8a3dfe4d1f4c9e --- /dev/null +++ b/sleep_assessment/cases/case10/case10_16.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "MASSSS3_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_01.npy", + "instruction": "Please use the provided sleep-stage label file to calculate two N3-related sleep metrics for the whole night: the number of N3 bouts, and the mean N3 bout length in minutes. Here, a bout is defined as a contiguous N3 segment with length greater than or equal to 2 epochs. Please clearly report both final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two values from the agent's report:\n1) the number of N3 bouts across the night\n2) the mean N3 bout length in minutes\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"n3_bout_count\" and \"n3_mean_bout_length_min\".\n5. \"n3_bout_count\" must be an INTEGER or null.\n6. \"n3_mean_bout_length_min\" must be a FLOAT or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"n3_bout_count\": , \"n3_mean_bout_length_min\": }", + "metrics": [ + { + "metric_id": "n3_bout_count_accuracy", + "type": "numeric_check", + "target_key": "n3_bout_count", + "weight": 50, + "params": { + "gt_value": 26, + "tolerance": 1.3 + } + }, + { + "metric_id": "n3_mean_bout_length_accuracy", + "type": "numeric_check", + "target_key": "n3_mean_bout_length_min", + "weight": 50, + "params": { + "gt_value": 2.3846153846153846, + "tolerance": 0.11923076923076924 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case10/case10_17.json b/sleep_assessment/cases/case10/case10_17.json new file mode 100644 index 0000000000000000000000000000000000000000..a56412bd2f5c5d5c2b1619f922cd9110bd3d0c01 --- /dev/null +++ b/sleep_assessment/cases/case10/case10_17.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "MASSSS3_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_02.npy", + "instruction": "Please use the provided sleep-stage label file to calculate two N3-related sleep metrics for the whole night: the number of N3 bouts, and the mean N3 bout length in minutes. Here, a bout is defined as a contiguous N3 segment with length greater than or equal to 2 epochs. Please clearly report both final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two values from the agent's report:\n1) the number of N3 bouts across the night\n2) the mean N3 bout length in minutes\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"n3_bout_count\" and \"n3_mean_bout_length_min\".\n5. \"n3_bout_count\" must be an INTEGER or null.\n6. \"n3_mean_bout_length_min\" must be a FLOAT or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"n3_bout_count\": , \"n3_mean_bout_length_min\": }", + "metrics": [ + { + "metric_id": "n3_bout_count_accuracy", + "type": "numeric_check", + "target_key": "n3_bout_count", + "weight": 50, + "params": { + "gt_value": 19, + "tolerance": 0.9500000000000001 + } + }, + { + "metric_id": "n3_mean_bout_length_accuracy", + "type": "numeric_check", + "target_key": "n3_mean_bout_length_min", + "weight": 50, + "params": { + "gt_value": 4.052631578947368, + "tolerance": 0.2026315789473684 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case10/case10_18.json b/sleep_assessment/cases/case10/case10_18.json new file mode 100644 index 0000000000000000000000000000000000000000..016b7942d06ea3f5b2f0c0373507f40cb408e491 --- /dev/null +++ b/sleep_assessment/cases/case10/case10_18.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "MASSSS3_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_03.npy", + "instruction": "Please use the provided sleep-stage label file to calculate two N3-related sleep metrics for the whole night: the number of N3 bouts, and the mean N3 bout length in minutes. Here, a bout is defined as a contiguous N3 segment with length greater than or equal to 2 epochs. Please clearly report both final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two values from the agent's report:\n1) the number of N3 bouts across the night\n2) the mean N3 bout length in minutes\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"n3_bout_count\" and \"n3_mean_bout_length_min\".\n5. \"n3_bout_count\" must be an INTEGER or null.\n6. \"n3_mean_bout_length_min\" must be a FLOAT or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"n3_bout_count\": , \"n3_mean_bout_length_min\": }", + "metrics": [ + { + "metric_id": "n3_bout_count_accuracy", + "type": "numeric_check", + "target_key": "n3_bout_count", + "weight": 50, + "params": { + "gt_value": 19, + "tolerance": 0.9500000000000001 + } + }, + { + "metric_id": "n3_mean_bout_length_accuracy", + "type": "numeric_check", + "target_key": "n3_mean_bout_length_min", + "weight": 50, + "params": { + "gt_value": 2.8421052631578947, + "tolerance": 0.14210526315789473 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case10/case10_19.json b/sleep_assessment/cases/case10/case10_19.json new file mode 100644 index 0000000000000000000000000000000000000000..3f27eb812985aa40338adbb307d95c1b7fc3256c --- /dev/null +++ b/sleep_assessment/cases/case10/case10_19.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "MASSSS3_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_04.npy", + "instruction": "Please use the provided sleep-stage label file to calculate two N3-related sleep metrics for the whole night: the number of N3 bouts, and the mean N3 bout length in minutes. Here, a bout is defined as a contiguous N3 segment with length greater than or equal to 2 epochs. Please clearly report both final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two values from the agent's report:\n1) the number of N3 bouts across the night\n2) the mean N3 bout length in minutes\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"n3_bout_count\" and \"n3_mean_bout_length_min\".\n5. \"n3_bout_count\" must be an INTEGER or null.\n6. \"n3_mean_bout_length_min\" must be a FLOAT or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"n3_bout_count\": , \"n3_mean_bout_length_min\": }", + "metrics": [ + { + "metric_id": "n3_bout_count_accuracy", + "type": "numeric_check", + "target_key": "n3_bout_count", + "weight": 50, + "params": { + "gt_value": 10, + "tolerance": 0.5 + } + }, + { + "metric_id": "n3_mean_bout_length_accuracy", + "type": "numeric_check", + "target_key": "n3_mean_bout_length_min", + "weight": 50, + "params": { + "gt_value": 2.65, + "tolerance": 0.1325 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case10/case10_20.json b/sleep_assessment/cases/case10/case10_20.json new file mode 100644 index 0000000000000000000000000000000000000000..3f9a7f8f6e9e53d88aab2277fd841f988cfde6c8 --- /dev/null +++ b/sleep_assessment/cases/case10/case10_20.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "MASSSS3_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_05.npy", + "instruction": "Please use the provided sleep-stage label file to calculate two N3-related sleep metrics for the whole night: the number of N3 bouts, and the mean N3 bout length in minutes. Here, a bout is defined as a contiguous N3 segment with length greater than or equal to 2 epochs. Please clearly report both final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two values from the agent's report:\n1) the number of N3 bouts across the night\n2) the mean N3 bout length in minutes\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"n3_bout_count\" and \"n3_mean_bout_length_min\".\n5. \"n3_bout_count\" must be an INTEGER or null.\n6. \"n3_mean_bout_length_min\" must be a FLOAT or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"n3_bout_count\": , \"n3_mean_bout_length_min\": }", + "metrics": [ + { + "metric_id": "n3_bout_count_accuracy", + "type": "numeric_check", + "target_key": "n3_bout_count", + "weight": 50, + "params": { + "gt_value": 6, + "tolerance": 0.30000000000000004 + } + }, + { + "metric_id": "n3_mean_bout_length_accuracy", + "type": "numeric_check", + "target_key": "n3_mean_bout_length_min", + "weight": 50, + "params": { + "gt_value": 15.833333333333334, + "tolerance": 0.7916666666666667 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case10/case10_21.json b/sleep_assessment/cases/case10/case10_21.json new file mode 100644 index 0000000000000000000000000000000000000000..06a89c79ce06208718f0662ece5844a856408597 --- /dev/null +++ b/sleep_assessment/cases/case10/case10_21.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_01.npy", + "instruction": "Please use the provided sleep-stage label file to calculate two N3-related sleep metrics for the whole night: the number of N3 bouts, and the mean N3 bout length in minutes. Here, a bout is defined as a contiguous N3 segment with length greater than or equal to 2 epochs. Please clearly report both final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two values from the agent's report:\n1) the number of N3 bouts across the night\n2) the mean N3 bout length in minutes\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"n3_bout_count\" and \"n3_mean_bout_length_min\".\n5. \"n3_bout_count\" must be an INTEGER or null.\n6. \"n3_mean_bout_length_min\" must be a FLOAT or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"n3_bout_count\": , \"n3_mean_bout_length_min\": }", + "metrics": [ + { + "metric_id": "n3_bout_count_accuracy", + "type": "numeric_check", + "target_key": "n3_bout_count", + "weight": 50, + "params": { + "gt_value": 4, + "tolerance": 0.2 + } + }, + { + "metric_id": "n3_mean_bout_length_accuracy", + "type": "numeric_check", + "target_key": "n3_mean_bout_length_min", + "weight": 50, + "params": { + "gt_value": 8.875, + "tolerance": 0.44375000000000003 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case10/case10_22.json b/sleep_assessment/cases/case10/case10_22.json new file mode 100644 index 0000000000000000000000000000000000000000..149657f5d9eff318b5d31ef9530c634aed26368f --- /dev/null +++ b/sleep_assessment/cases/case10/case10_22.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_02.npy", + "instruction": "Please use the provided sleep-stage label file to calculate two N3-related sleep metrics for the whole night: the number of N3 bouts, and the mean N3 bout length in minutes. Here, a bout is defined as a contiguous N3 segment with length greater than or equal to 2 epochs. Please clearly report both final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two values from the agent's report:\n1) the number of N3 bouts across the night\n2) the mean N3 bout length in minutes\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"n3_bout_count\" and \"n3_mean_bout_length_min\".\n5. \"n3_bout_count\" must be an INTEGER or null.\n6. \"n3_mean_bout_length_min\" must be a FLOAT or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"n3_bout_count\": , \"n3_mean_bout_length_min\": }", + "metrics": [ + { + "metric_id": "n3_bout_count_accuracy", + "type": "numeric_check", + "target_key": "n3_bout_count", + "weight": 50, + "params": { + "gt_value": 4, + "tolerance": 0.2 + } + }, + { + "metric_id": "n3_mean_bout_length_accuracy", + "type": "numeric_check", + "target_key": "n3_mean_bout_length_min", + "weight": 50, + "params": { + "gt_value": 2.625, + "tolerance": 0.13125 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case10/case10_23.json b/sleep_assessment/cases/case10/case10_23.json new file mode 100644 index 0000000000000000000000000000000000000000..161b62893921c70b2b47a0119d3527050338bd12 --- /dev/null +++ b/sleep_assessment/cases/case10/case10_23.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "Physionet2018_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_03.npy", + "instruction": "Please use the provided sleep-stage label file to calculate two N3-related sleep metrics for the whole night: the number of N3 bouts, and the mean N3 bout length in minutes. Here, a bout is defined as a contiguous N3 segment with length greater than or equal to 2 epochs. Please clearly report both final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two values from the agent's report:\n1) the number of N3 bouts across the night\n2) the mean N3 bout length in minutes\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"n3_bout_count\" and \"n3_mean_bout_length_min\".\n5. \"n3_bout_count\" must be an INTEGER or null.\n6. \"n3_mean_bout_length_min\" must be a FLOAT or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"n3_bout_count\": , \"n3_mean_bout_length_min\": }", + "metrics": [ + { + "metric_id": "n3_bout_count_accuracy", + "type": "numeric_check", + "target_key": "n3_bout_count", + "weight": 50, + "params": { + "gt_value": 4, + "tolerance": 0.2 + } + }, + { + "metric_id": "n3_mean_bout_length_accuracy", + "type": "numeric_check", + "target_key": "n3_mean_bout_length_min", + "weight": 50, + "params": { + "gt_value": 6.5, + "tolerance": 0.325 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case10/case10_24.json b/sleep_assessment/cases/case10/case10_24.json new file mode 100644 index 0000000000000000000000000000000000000000..a2a92ad05b806b541be6bf1d29c9ceccadcacc80 --- /dev/null +++ b/sleep_assessment/cases/case10/case10_24.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_04.npy", + "instruction": "Please use the provided sleep-stage label file to calculate two N3-related sleep metrics for the whole night: the number of N3 bouts, and the mean N3 bout length in minutes. Here, a bout is defined as a contiguous N3 segment with length greater than or equal to 2 epochs. Please clearly report both final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two values from the agent's report:\n1) the number of N3 bouts across the night\n2) the mean N3 bout length in minutes\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"n3_bout_count\" and \"n3_mean_bout_length_min\".\n5. \"n3_bout_count\" must be an INTEGER or null.\n6. \"n3_mean_bout_length_min\" must be a FLOAT or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"n3_bout_count\": , \"n3_mean_bout_length_min\": }", + "metrics": [ + { + "metric_id": "n3_bout_count_accuracy", + "type": "numeric_check", + "target_key": "n3_bout_count", + "weight": 50, + "params": { + "gt_value": 19, + "tolerance": 0.9500000000000001 + } + }, + { + "metric_id": "n3_mean_bout_length_accuracy", + "type": "numeric_check", + "target_key": "n3_mean_bout_length_min", + "weight": 50, + "params": { + "gt_value": 4.342105263157895, + "tolerance": 0.21710526315789475 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case10/case10_25.json b/sleep_assessment/cases/case10/case10_25.json new file mode 100644 index 0000000000000000000000000000000000000000..5fd90a6a9c9cd79cae01aa791d8b3fd63a94e01b --- /dev/null +++ b/sleep_assessment/cases/case10/case10_25.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_05.npy", + "instruction": "Please use the provided sleep-stage label file to calculate two N3-related sleep metrics for the whole night: the number of N3 bouts, and the mean N3 bout length in minutes. Here, a bout is defined as a contiguous N3 segment with length greater than or equal to 2 epochs. Please clearly report both final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two values from the agent's report:\n1) the number of N3 bouts across the night\n2) the mean N3 bout length in minutes\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"n3_bout_count\" and \"n3_mean_bout_length_min\".\n5. \"n3_bout_count\" must be an INTEGER or null.\n6. \"n3_mean_bout_length_min\" must be a FLOAT or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"n3_bout_count\": , \"n3_mean_bout_length_min\": }", + "metrics": [ + { + "metric_id": "n3_bout_count_accuracy", + "type": "numeric_check", + "target_key": "n3_bout_count", + "weight": 50, + "params": { + "gt_value": 4, + "tolerance": 0.2 + } + }, + { + "metric_id": "n3_mean_bout_length_accuracy", + "type": "numeric_check", + "target_key": "n3_mean_bout_length_min", + "weight": 50, + "params": { + "gt_value": 12.0, + "tolerance": 0.6000000000000001 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case11/case11_01.json b/sleep_assessment/cases/case11/case11_01.json new file mode 100644 index 0000000000000000000000000000000000000000..a5777f3bcb8d17f2c891fb8568849108af253d0e --- /dev/null +++ b/sleep_assessment/cases/case11/case11_01.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Please use the provided sleep-stage label file to locate the longest contiguous Wake (W) segment during the nocturnal sleep period. Define the nocturnal sleep period as the interval from the first non-Wake sleep epoch to the last non-Wake sleep epoch, excluding Wake epochs before sleep onset and after final awakening. Report the start time and end time of this longest Wake segment in minutes, using the first non-Wake sleep epoch as 0 minutes. Then classify the impact of this single longest awakening interruption on sleep continuity as low, medium, or high: low for <5 minutes, medium for 5 to <30 minutes, and high for >=30 minutes. Please clearly report all three final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three values from the agent's report:\n1) the start time in minutes of the longest contiguous Wake (W) segment during the sleep period\n2) the end time in minutes of that longest contiguous Wake (W) segment\n3) the sleep-continuity impact level of that Wake interruption\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"longest_w_start_min\", \"longest_w_end_min\", and \"wake_interruption_impact\".\n5. \"longest_w_start_min\" and \"longest_w_end_min\" must be FLOATS in minutes or null.\n6. \"wake_interruption_impact\" must be exactly one of \"low\", \"medium\", \"high\", or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_w_start_min\": , \"longest_w_end_min\": , \"wake_interruption_impact\": <\"low\"|\"medium\"|\"high\"|null>}", + "metrics": [ + { + "metric_id": "longest_w_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_start_min", + "weight": 30, + "params": { + "gt_value": 353.5, + "tolerance": 2.0 + } + }, + { + "metric_id": "longest_w_end_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_end_min", + "weight": 30, + "params": { + "gt_value": 391.5, + "tolerance": 2.0 + } + }, + { + "metric_id": "wake_interruption_impact_accuracy", + "type": "categorical_check", + "target_key": "wake_interruption_impact", + "weight": 40, + "params": { + "gt_value": "high" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case11/case11_02.json b/sleep_assessment/cases/case11/case11_02.json new file mode 100644 index 0000000000000000000000000000000000000000..9363519f89642b8b25a69bc588cfcbc97262eb40 --- /dev/null +++ b/sleep_assessment/cases/case11/case11_02.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_02.npy", + "instruction": "Please use the provided sleep-stage label file to locate the longest contiguous Wake (W) segment during the nocturnal sleep period. Define the nocturnal sleep period as the interval from the first non-Wake sleep epoch to the last non-Wake sleep epoch, excluding Wake epochs before sleep onset and after final awakening. Report the start time and end time of this longest Wake segment in minutes, using the first non-Wake sleep epoch as 0 minutes. Then classify the impact of this single longest awakening interruption on sleep continuity as low, medium, or high: low for <5 minutes, medium for 5 to <30 minutes, and high for >=30 minutes. Please clearly report all three final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three values from the agent's report:\n1) the start time in minutes of the longest contiguous Wake (W) segment during the sleep period\n2) the end time in minutes of that longest contiguous Wake (W) segment\n3) the sleep-continuity impact level of that Wake interruption\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"longest_w_start_min\", \"longest_w_end_min\", and \"wake_interruption_impact\".\n5. \"longest_w_start_min\" and \"longest_w_end_min\" must be FLOATS in minutes or null.\n6. \"wake_interruption_impact\" must be exactly one of \"low\", \"medium\", \"high\", or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_w_start_min\": , \"longest_w_end_min\": , \"wake_interruption_impact\": <\"low\"|\"medium\"|\"high\"|null>}", + "metrics": [ + { + "metric_id": "longest_w_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_start_min", + "weight": 30, + "params": { + "gt_value": 4.5, + "tolerance": 2.0 + } + }, + { + "metric_id": "longest_w_end_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_end_min", + "weight": 30, + "params": { + "gt_value": 32.0, + "tolerance": 2.0 + } + }, + { + "metric_id": "wake_interruption_impact_accuracy", + "type": "categorical_check", + "target_key": "wake_interruption_impact", + "weight": 40, + "params": { + "gt_value": "medium" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case11/case11_03.json b/sleep_assessment/cases/case11/case11_03.json new file mode 100644 index 0000000000000000000000000000000000000000..d5640803adb601e1e7ec3d1a95bef8637688d108 --- /dev/null +++ b/sleep_assessment/cases/case11/case11_03.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_03.npy", + "instruction": "Please use the provided sleep-stage label file to locate the longest contiguous Wake (W) segment during the nocturnal sleep period. Define the nocturnal sleep period as the interval from the first non-Wake sleep epoch to the last non-Wake sleep epoch, excluding Wake epochs before sleep onset and after final awakening. Report the start time and end time of this longest Wake segment in minutes, using the first non-Wake sleep epoch as 0 minutes. Then classify the impact of this single longest awakening interruption on sleep continuity as low, medium, or high: low for <5 minutes, medium for 5 to <30 minutes, and high for >=30 minutes. Please clearly report all three final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three values from the agent's report:\n1) the start time in minutes of the longest contiguous Wake (W) segment during the sleep period\n2) the end time in minutes of that longest contiguous Wake (W) segment\n3) the sleep-continuity impact level of that Wake interruption\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"longest_w_start_min\", \"longest_w_end_min\", and \"wake_interruption_impact\".\n5. \"longest_w_start_min\" and \"longest_w_end_min\" must be FLOATS in minutes or null.\n6. \"wake_interruption_impact\" must be exactly one of \"low\", \"medium\", \"high\", or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_w_start_min\": , \"longest_w_end_min\": , \"wake_interruption_impact\": <\"low\"|\"medium\"|\"high\"|null>}", + "metrics": [ + { + "metric_id": "longest_w_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_start_min", + "weight": 30, + "params": { + "gt_value": 241.5, + "tolerance": 2.0 + } + }, + { + "metric_id": "longest_w_end_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_end_min", + "weight": 30, + "params": { + "gt_value": 256.0, + "tolerance": 2.0 + } + }, + { + "metric_id": "wake_interruption_impact_accuracy", + "type": "categorical_check", + "target_key": "wake_interruption_impact", + "weight": 40, + "params": { + "gt_value": "medium" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case11/case11_04.json b/sleep_assessment/cases/case11/case11_04.json new file mode 100644 index 0000000000000000000000000000000000000000..5c857d315ac94ba5c1a029f73578ff426dbcf658 --- /dev/null +++ b/sleep_assessment/cases/case11/case11_04.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "ISRUC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_04.npy", + "instruction": "Please use the provided sleep-stage label file to locate the longest contiguous Wake (W) segment during the nocturnal sleep period. Define the nocturnal sleep period as the interval from the first non-Wake sleep epoch to the last non-Wake sleep epoch, excluding Wake epochs before sleep onset and after final awakening. Report the start time and end time of this longest Wake segment in minutes, using the first non-Wake sleep epoch as 0 minutes. Then classify the impact of this single longest awakening interruption on sleep continuity as low, medium, or high: low for <5 minutes, medium for 5 to <30 minutes, and high for >=30 minutes. Please clearly report all three final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three values from the agent's report:\n1) the start time in minutes of the longest contiguous Wake (W) segment during the sleep period\n2) the end time in minutes of that longest contiguous Wake (W) segment\n3) the sleep-continuity impact level of that Wake interruption\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"longest_w_start_min\", \"longest_w_end_min\", and \"wake_interruption_impact\".\n5. \"longest_w_start_min\" and \"longest_w_end_min\" must be FLOATS in minutes or null.\n6. \"wake_interruption_impact\" must be exactly one of \"low\", \"medium\", \"high\", or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_w_start_min\": , \"longest_w_end_min\": , \"wake_interruption_impact\": <\"low\"|\"medium\"|\"high\"|null>}", + "metrics": [ + { + "metric_id": "longest_w_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_start_min", + "weight": 30, + "params": { + "gt_value": 256.5, + "tolerance": 2.0 + } + }, + { + "metric_id": "longest_w_end_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_end_min", + "weight": 30, + "params": { + "gt_value": 258.0, + "tolerance": 2.0 + } + }, + { + "metric_id": "wake_interruption_impact_accuracy", + "type": "categorical_check", + "target_key": "wake_interruption_impact", + "weight": 40, + "params": { + "gt_value": "low" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case11/case11_05.json b/sleep_assessment/cases/case11/case11_05.json new file mode 100644 index 0000000000000000000000000000000000000000..d5c0440d3e1852bd0f8911f37ee4e44c333ba4e0 --- /dev/null +++ b/sleep_assessment/cases/case11/case11_05.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_05.npy", + "instruction": "Please use the provided sleep-stage label file to locate the longest contiguous Wake (W) segment during the nocturnal sleep period. Define the nocturnal sleep period as the interval from the first non-Wake sleep epoch to the last non-Wake sleep epoch, excluding Wake epochs before sleep onset and after final awakening. Report the start time and end time of this longest Wake segment in minutes, using the first non-Wake sleep epoch as 0 minutes. Then classify the impact of this single longest awakening interruption on sleep continuity as low, medium, or high: low for <5 minutes, medium for 5 to <30 minutes, and high for >=30 minutes. Please clearly report all three final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three values from the agent's report:\n1) the start time in minutes of the longest contiguous Wake (W) segment during the sleep period\n2) the end time in minutes of that longest contiguous Wake (W) segment\n3) the sleep-continuity impact level of that Wake interruption\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"longest_w_start_min\", \"longest_w_end_min\", and \"wake_interruption_impact\".\n5. \"longest_w_start_min\" and \"longest_w_end_min\" must be FLOATS in minutes or null.\n6. \"wake_interruption_impact\" must be exactly one of \"low\", \"medium\", \"high\", or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_w_start_min\": , \"longest_w_end_min\": , \"wake_interruption_impact\": <\"low\"|\"medium\"|\"high\"|null>}", + "metrics": [ + { + "metric_id": "longest_w_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_start_min", + "weight": 30, + "params": { + "gt_value": 21.5, + "tolerance": 2.0 + } + }, + { + "metric_id": "longest_w_end_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_end_min", + "weight": 30, + "params": { + "gt_value": 33.0, + "tolerance": 2.0 + } + }, + { + "metric_id": "wake_interruption_impact_accuracy", + "type": "categorical_check", + "target_key": "wake_interruption_impact", + "weight": 40, + "params": { + "gt_value": "medium" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case11/case11_06.json b/sleep_assessment/cases/case11/case11_06.json new file mode 100644 index 0000000000000000000000000000000000000000..225eb75b559fbafc25b685d583a8b61c3ae9907b --- /dev/null +++ b/sleep_assessment/cases/case11/case11_06.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "HMC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_01.npy", + "instruction": "Please use the provided sleep-stage label file to locate the longest contiguous Wake (W) segment during the nocturnal sleep period. Define the nocturnal sleep period as the interval from the first non-Wake sleep epoch to the last non-Wake sleep epoch, excluding Wake epochs before sleep onset and after final awakening. Report the start time and end time of this longest Wake segment in minutes, using the first non-Wake sleep epoch as 0 minutes. Then classify the impact of this single longest awakening interruption on sleep continuity as low, medium, or high: low for <5 minutes, medium for 5 to <30 minutes, and high for >=30 minutes. Please clearly report all three final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three values from the agent's report:\n1) the start time in minutes of the longest contiguous Wake (W) segment during the sleep period\n2) the end time in minutes of that longest contiguous Wake (W) segment\n3) the sleep-continuity impact level of that Wake interruption\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"longest_w_start_min\", \"longest_w_end_min\", and \"wake_interruption_impact\".\n5. \"longest_w_start_min\" and \"longest_w_end_min\" must be FLOATS in minutes or null.\n6. \"wake_interruption_impact\" must be exactly one of \"low\", \"medium\", \"high\", or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_w_start_min\": , \"longest_w_end_min\": , \"wake_interruption_impact\": <\"low\"|\"medium\"|\"high\"|null>}", + "metrics": [ + { + "metric_id": "longest_w_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_start_min", + "weight": 30, + "params": { + "gt_value": 154.5, + "tolerance": 2.0 + } + }, + { + "metric_id": "longest_w_end_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_end_min", + "weight": 30, + "params": { + "gt_value": 196.5, + "tolerance": 2.0 + } + }, + { + "metric_id": "wake_interruption_impact_accuracy", + "type": "categorical_check", + "target_key": "wake_interruption_impact", + "weight": 40, + "params": { + "gt_value": "high" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case11/case11_07.json b/sleep_assessment/cases/case11/case11_07.json new file mode 100644 index 0000000000000000000000000000000000000000..89f7897a1b8c075d5751d58742e72ba210cef1a9 --- /dev/null +++ b/sleep_assessment/cases/case11/case11_07.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "HMC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_02.npy", + "instruction": "Please use the provided sleep-stage label file to locate the longest contiguous Wake (W) segment during the nocturnal sleep period. Define the nocturnal sleep period as the interval from the first non-Wake sleep epoch to the last non-Wake sleep epoch, excluding Wake epochs before sleep onset and after final awakening. Report the start time and end time of this longest Wake segment in minutes, using the first non-Wake sleep epoch as 0 minutes. Then classify the impact of this single longest awakening interruption on sleep continuity as low, medium, or high: low for <5 minutes, medium for 5 to <30 minutes, and high for >=30 minutes. Please clearly report all three final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three values from the agent's report:\n1) the start time in minutes of the longest contiguous Wake (W) segment during the sleep period\n2) the end time in minutes of that longest contiguous Wake (W) segment\n3) the sleep-continuity impact level of that Wake interruption\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"longest_w_start_min\", \"longest_w_end_min\", and \"wake_interruption_impact\".\n5. \"longest_w_start_min\" and \"longest_w_end_min\" must be FLOATS in minutes or null.\n6. \"wake_interruption_impact\" must be exactly one of \"low\", \"medium\", \"high\", or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_w_start_min\": , \"longest_w_end_min\": , \"wake_interruption_impact\": <\"low\"|\"medium\"|\"high\"|null>}", + "metrics": [ + { + "metric_id": "longest_w_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_start_min", + "weight": 30, + "params": { + "gt_value": 397.0, + "tolerance": 2.0 + } + }, + { + "metric_id": "longest_w_end_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_end_min", + "weight": 30, + "params": { + "gt_value": 405.0, + "tolerance": 2.0 + } + }, + { + "metric_id": "wake_interruption_impact_accuracy", + "type": "categorical_check", + "target_key": "wake_interruption_impact", + "weight": 40, + "params": { + "gt_value": "medium" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case11/case11_08.json b/sleep_assessment/cases/case11/case11_08.json new file mode 100644 index 0000000000000000000000000000000000000000..633498c1b341201221b6fda6ca427450a28a57d4 --- /dev/null +++ b/sleep_assessment/cases/case11/case11_08.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "HMC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_03.npy", + "instruction": "Please use the provided sleep-stage label file to locate the longest contiguous Wake (W) segment during the nocturnal sleep period. Define the nocturnal sleep period as the interval from the first non-Wake sleep epoch to the last non-Wake sleep epoch, excluding Wake epochs before sleep onset and after final awakening. Report the start time and end time of this longest Wake segment in minutes, using the first non-Wake sleep epoch as 0 minutes. Then classify the impact of this single longest awakening interruption on sleep continuity as low, medium, or high: low for <5 minutes, medium for 5 to <30 minutes, and high for >=30 minutes. Please clearly report all three final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three values from the agent's report:\n1) the start time in minutes of the longest contiguous Wake (W) segment during the sleep period\n2) the end time in minutes of that longest contiguous Wake (W) segment\n3) the sleep-continuity impact level of that Wake interruption\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"longest_w_start_min\", \"longest_w_end_min\", and \"wake_interruption_impact\".\n5. \"longest_w_start_min\" and \"longest_w_end_min\" must be FLOATS in minutes or null.\n6. \"wake_interruption_impact\" must be exactly one of \"low\", \"medium\", \"high\", or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_w_start_min\": , \"longest_w_end_min\": , \"wake_interruption_impact\": <\"low\"|\"medium\"|\"high\"|null>}", + "metrics": [ + { + "metric_id": "longest_w_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_start_min", + "weight": 30, + "params": { + "gt_value": 148.5, + "tolerance": 2.0 + } + }, + { + "metric_id": "longest_w_end_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_end_min", + "weight": 30, + "params": { + "gt_value": 195.0, + "tolerance": 2.0 + } + }, + { + "metric_id": "wake_interruption_impact_accuracy", + "type": "categorical_check", + "target_key": "wake_interruption_impact", + "weight": 40, + "params": { + "gt_value": "high" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case11/case11_09.json b/sleep_assessment/cases/case11/case11_09.json new file mode 100644 index 0000000000000000000000000000000000000000..736d0865ce9de740f93011e371f3d288babb8301 --- /dev/null +++ b/sleep_assessment/cases/case11/case11_09.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "HMC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_04.npy", + "instruction": "Please use the provided sleep-stage label file to locate the longest contiguous Wake (W) segment during the nocturnal sleep period. Define the nocturnal sleep period as the interval from the first non-Wake sleep epoch to the last non-Wake sleep epoch, excluding Wake epochs before sleep onset and after final awakening. Report the start time and end time of this longest Wake segment in minutes, using the first non-Wake sleep epoch as 0 minutes. Then classify the impact of this single longest awakening interruption on sleep continuity as low, medium, or high: low for <5 minutes, medium for 5 to <30 minutes, and high for >=30 minutes. Please clearly report all three final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three values from the agent's report:\n1) the start time in minutes of the longest contiguous Wake (W) segment during the sleep period\n2) the end time in minutes of that longest contiguous Wake (W) segment\n3) the sleep-continuity impact level of that Wake interruption\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"longest_w_start_min\", \"longest_w_end_min\", and \"wake_interruption_impact\".\n5. \"longest_w_start_min\" and \"longest_w_end_min\" must be FLOATS in minutes or null.\n6. \"wake_interruption_impact\" must be exactly one of \"low\", \"medium\", \"high\", or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_w_start_min\": , \"longest_w_end_min\": , \"wake_interruption_impact\": <\"low\"|\"medium\"|\"high\"|null>}", + "metrics": [ + { + "metric_id": "longest_w_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_start_min", + "weight": 30, + "params": { + "gt_value": 493.5, + "tolerance": 2.0 + } + }, + { + "metric_id": "longest_w_end_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_end_min", + "weight": 30, + "params": { + "gt_value": 500.5, + "tolerance": 2.0 + } + }, + { + "metric_id": "wake_interruption_impact_accuracy", + "type": "categorical_check", + "target_key": "wake_interruption_impact", + "weight": 40, + "params": { + "gt_value": "medium" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case11/case11_10.json b/sleep_assessment/cases/case11/case11_10.json new file mode 100644 index 0000000000000000000000000000000000000000..6a9f93e547d1a7820dd1be53cd537cad6735e6df --- /dev/null +++ b/sleep_assessment/cases/case11/case11_10.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "HMC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_05.npy", + "instruction": "Please use the provided sleep-stage label file to locate the longest contiguous Wake (W) segment during the nocturnal sleep period. Define the nocturnal sleep period as the interval from the first non-Wake sleep epoch to the last non-Wake sleep epoch, excluding Wake epochs before sleep onset and after final awakening. Report the start time and end time of this longest Wake segment in minutes, using the first non-Wake sleep epoch as 0 minutes. Then classify the impact of this single longest awakening interruption on sleep continuity as low, medium, or high: low for <5 minutes, medium for 5 to <30 minutes, and high for >=30 minutes. Please clearly report all three final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three values from the agent's report:\n1) the start time in minutes of the longest contiguous Wake (W) segment during the sleep period\n2) the end time in minutes of that longest contiguous Wake (W) segment\n3) the sleep-continuity impact level of that Wake interruption\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"longest_w_start_min\", \"longest_w_end_min\", and \"wake_interruption_impact\".\n5. \"longest_w_start_min\" and \"longest_w_end_min\" must be FLOATS in minutes or null.\n6. \"wake_interruption_impact\" must be exactly one of \"low\", \"medium\", \"high\", or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_w_start_min\": , \"longest_w_end_min\": , \"wake_interruption_impact\": <\"low\"|\"medium\"|\"high\"|null>}", + "metrics": [ + { + "metric_id": "longest_w_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_start_min", + "weight": 30, + "params": { + "gt_value": 112.5, + "tolerance": 2.0 + } + }, + { + "metric_id": "longest_w_end_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_end_min", + "weight": 30, + "params": { + "gt_value": 122.0, + "tolerance": 2.0 + } + }, + { + "metric_id": "wake_interruption_impact_accuracy", + "type": "categorical_check", + "target_key": "wake_interruption_impact", + "weight": 40, + "params": { + "gt_value": "medium" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case11/case11_11.json b/sleep_assessment/cases/case11/case11_11.json new file mode 100644 index 0000000000000000000000000000000000000000..6bd51f1cc7a31b1dbb23f6031a1087d2c494b60a --- /dev/null +++ b/sleep_assessment/cases/case11/case11_11.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_01.npy", + "instruction": "Please use the provided sleep-stage label file to locate the longest contiguous Wake (W) segment during the nocturnal sleep period. Define the nocturnal sleep period as the interval from the first non-Wake sleep epoch to the last non-Wake sleep epoch, excluding Wake epochs before sleep onset and after final awakening. Report the start time and end time of this longest Wake segment in minutes, using the first non-Wake sleep epoch as 0 minutes. Then classify the impact of this single longest awakening interruption on sleep continuity as low, medium, or high: low for <5 minutes, medium for 5 to <30 minutes, and high for >=30 minutes. Please clearly report all three final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three values from the agent's report:\n1) the start time in minutes of the longest contiguous Wake (W) segment during the sleep period\n2) the end time in minutes of that longest contiguous Wake (W) segment\n3) the sleep-continuity impact level of that Wake interruption\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"longest_w_start_min\", \"longest_w_end_min\", and \"wake_interruption_impact\".\n5. \"longest_w_start_min\" and \"longest_w_end_min\" must be FLOATS in minutes or null.\n6. \"wake_interruption_impact\" must be exactly one of \"low\", \"medium\", \"high\", or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_w_start_min\": , \"longest_w_end_min\": , \"wake_interruption_impact\": <\"low\"|\"medium\"|\"high\"|null>}", + "metrics": [ + { + "metric_id": "longest_w_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_start_min", + "weight": 30, + "params": { + "gt_value": 47.5, + "tolerance": 2.0 + } + }, + { + "metric_id": "longest_w_end_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_end_min", + "weight": 30, + "params": { + "gt_value": 56.0, + "tolerance": 2.0 + } + }, + { + "metric_id": "wake_interruption_impact_accuracy", + "type": "categorical_check", + "target_key": "wake_interruption_impact", + "weight": 40, + "params": { + "gt_value": "medium" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case11/case11_12.json b/sleep_assessment/cases/case11/case11_12.json new file mode 100644 index 0000000000000000000000000000000000000000..ecc65e8e161e8f17f6d07c1a034e643c59594f72 --- /dev/null +++ b/sleep_assessment/cases/case11/case11_12.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_02.npy", + "instruction": "Please use the provided sleep-stage label file to locate the longest contiguous Wake (W) segment during the nocturnal sleep period. Define the nocturnal sleep period as the interval from the first non-Wake sleep epoch to the last non-Wake sleep epoch, excluding Wake epochs before sleep onset and after final awakening. Report the start time and end time of this longest Wake segment in minutes, using the first non-Wake sleep epoch as 0 minutes. Then classify the impact of this single longest awakening interruption on sleep continuity as low, medium, or high: low for <5 minutes, medium for 5 to <30 minutes, and high for >=30 minutes. Please clearly report all three final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three values from the agent's report:\n1) the start time in minutes of the longest contiguous Wake (W) segment during the sleep period\n2) the end time in minutes of that longest contiguous Wake (W) segment\n3) the sleep-continuity impact level of that Wake interruption\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"longest_w_start_min\", \"longest_w_end_min\", and \"wake_interruption_impact\".\n5. \"longest_w_start_min\" and \"longest_w_end_min\" must be FLOATS in minutes or null.\n6. \"wake_interruption_impact\" must be exactly one of \"low\", \"medium\", \"high\", or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_w_start_min\": , \"longest_w_end_min\": , \"wake_interruption_impact\": <\"low\"|\"medium\"|\"high\"|null>}", + "metrics": [ + { + "metric_id": "longest_w_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_start_min", + "weight": 30, + "params": { + "gt_value": 56.5, + "tolerance": 2.0 + } + }, + { + "metric_id": "longest_w_end_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_end_min", + "weight": 30, + "params": { + "gt_value": 66.5, + "tolerance": 2.0 + } + }, + { + "metric_id": "wake_interruption_impact_accuracy", + "type": "categorical_check", + "target_key": "wake_interruption_impact", + "weight": 40, + "params": { + "gt_value": "medium" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case11/case11_13.json b/sleep_assessment/cases/case11/case11_13.json new file mode 100644 index 0000000000000000000000000000000000000000..30192a129dd58532e7aa1b03b770adb833cf51b4 --- /dev/null +++ b/sleep_assessment/cases/case11/case11_13.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_03.npy", + "instruction": "Please use the provided sleep-stage label file to locate the longest contiguous Wake (W) segment during the nocturnal sleep period. Define the nocturnal sleep period as the interval from the first non-Wake sleep epoch to the last non-Wake sleep epoch, excluding Wake epochs before sleep onset and after final awakening. Report the start time and end time of this longest Wake segment in minutes, using the first non-Wake sleep epoch as 0 minutes. Then classify the impact of this single longest awakening interruption on sleep continuity as low, medium, or high: low for <5 minutes, medium for 5 to <30 minutes, and high for >=30 minutes. Please clearly report all three final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three values from the agent's report:\n1) the start time in minutes of the longest contiguous Wake (W) segment during the sleep period\n2) the end time in minutes of that longest contiguous Wake (W) segment\n3) the sleep-continuity impact level of that Wake interruption\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"longest_w_start_min\", \"longest_w_end_min\", and \"wake_interruption_impact\".\n5. \"longest_w_start_min\" and \"longest_w_end_min\" must be FLOATS in minutes or null.\n6. \"wake_interruption_impact\" must be exactly one of \"low\", \"medium\", \"high\", or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_w_start_min\": , \"longest_w_end_min\": , \"wake_interruption_impact\": <\"low\"|\"medium\"|\"high\"|null>}", + "metrics": [ + { + "metric_id": "longest_w_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_start_min", + "weight": 30, + "params": { + "gt_value": 282.0, + "tolerance": 2.0 + } + }, + { + "metric_id": "longest_w_end_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_end_min", + "weight": 30, + "params": { + "gt_value": 301.0, + "tolerance": 2.0 + } + }, + { + "metric_id": "wake_interruption_impact_accuracy", + "type": "categorical_check", + "target_key": "wake_interruption_impact", + "weight": 40, + "params": { + "gt_value": "medium" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case11/case11_14.json b/sleep_assessment/cases/case11/case11_14.json new file mode 100644 index 0000000000000000000000000000000000000000..0b0012884b15a42e123372a63700ea081ca16ab1 --- /dev/null +++ b/sleep_assessment/cases/case11/case11_14.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_04.npy", + "instruction": "Please use the provided sleep-stage label file to locate the longest contiguous Wake (W) segment during the nocturnal sleep period. Define the nocturnal sleep period as the interval from the first non-Wake sleep epoch to the last non-Wake sleep epoch, excluding Wake epochs before sleep onset and after final awakening. Report the start time and end time of this longest Wake segment in minutes, using the first non-Wake sleep epoch as 0 minutes. Then classify the impact of this single longest awakening interruption on sleep continuity as low, medium, or high: low for <5 minutes, medium for 5 to <30 minutes, and high for >=30 minutes. Please clearly report all three final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three values from the agent's report:\n1) the start time in minutes of the longest contiguous Wake (W) segment during the sleep period\n2) the end time in minutes of that longest contiguous Wake (W) segment\n3) the sleep-continuity impact level of that Wake interruption\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"longest_w_start_min\", \"longest_w_end_min\", and \"wake_interruption_impact\".\n5. \"longest_w_start_min\" and \"longest_w_end_min\" must be FLOATS in minutes or null.\n6. \"wake_interruption_impact\" must be exactly one of \"low\", \"medium\", \"high\", or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_w_start_min\": , \"longest_w_end_min\": , \"wake_interruption_impact\": <\"low\"|\"medium\"|\"high\"|null>}", + "metrics": [ + { + "metric_id": "longest_w_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_start_min", + "weight": 30, + "params": { + "gt_value": 50.5, + "tolerance": 2.0 + } + }, + { + "metric_id": "longest_w_end_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_end_min", + "weight": 30, + "params": { + "gt_value": 56.0, + "tolerance": 2.0 + } + }, + { + "metric_id": "wake_interruption_impact_accuracy", + "type": "categorical_check", + "target_key": "wake_interruption_impact", + "weight": 40, + "params": { + "gt_value": "medium" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case11/case11_15.json b/sleep_assessment/cases/case11/case11_15.json new file mode 100644 index 0000000000000000000000000000000000000000..4b1cd15a3a8a8c306ea41d7acd5a4ed9652f8294 --- /dev/null +++ b/sleep_assessment/cases/case11/case11_15.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_05.npy", + "instruction": "Please use the provided sleep-stage label file to locate the longest contiguous Wake (W) segment during the nocturnal sleep period. Define the nocturnal sleep period as the interval from the first non-Wake sleep epoch to the last non-Wake sleep epoch, excluding Wake epochs before sleep onset and after final awakening. Report the start time and end time of this longest Wake segment in minutes, using the first non-Wake sleep epoch as 0 minutes. Then classify the impact of this single longest awakening interruption on sleep continuity as low, medium, or high: low for <5 minutes, medium for 5 to <30 minutes, and high for >=30 minutes. Please clearly report all three final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three values from the agent's report:\n1) the start time in minutes of the longest contiguous Wake (W) segment during the sleep period\n2) the end time in minutes of that longest contiguous Wake (W) segment\n3) the sleep-continuity impact level of that Wake interruption\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"longest_w_start_min\", \"longest_w_end_min\", and \"wake_interruption_impact\".\n5. \"longest_w_start_min\" and \"longest_w_end_min\" must be FLOATS in minutes or null.\n6. \"wake_interruption_impact\" must be exactly one of \"low\", \"medium\", \"high\", or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_w_start_min\": , \"longest_w_end_min\": , \"wake_interruption_impact\": <\"low\"|\"medium\"|\"high\"|null>}", + "metrics": [ + { + "metric_id": "longest_w_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_start_min", + "weight": 30, + "params": { + "gt_value": 327.5, + "tolerance": 2.0 + } + }, + { + "metric_id": "longest_w_end_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_end_min", + "weight": 30, + "params": { + "gt_value": 346.5, + "tolerance": 2.0 + } + }, + { + "metric_id": "wake_interruption_impact_accuracy", + "type": "categorical_check", + "target_key": "wake_interruption_impact", + "weight": 40, + "params": { + "gt_value": "medium" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case11/case11_16.json b/sleep_assessment/cases/case11/case11_16.json new file mode 100644 index 0000000000000000000000000000000000000000..a00cd932e8df970150d487788b0a291fd44dfab2 --- /dev/null +++ b/sleep_assessment/cases/case11/case11_16.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "MASSSS3_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_01.npy", + "instruction": "Please use the provided sleep-stage label file to locate the longest contiguous Wake (W) segment during the nocturnal sleep period. Define the nocturnal sleep period as the interval from the first non-Wake sleep epoch to the last non-Wake sleep epoch, excluding Wake epochs before sleep onset and after final awakening. Report the start time and end time of this longest Wake segment in minutes, using the first non-Wake sleep epoch as 0 minutes. Then classify the impact of this single longest awakening interruption on sleep continuity as low, medium, or high: low for <5 minutes, medium for 5 to <30 minutes, and high for >=30 minutes. Please clearly report all three final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three values from the agent's report:\n1) the start time in minutes of the longest contiguous Wake (W) segment during the sleep period\n2) the end time in minutes of that longest contiguous Wake (W) segment\n3) the sleep-continuity impact level of that Wake interruption\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"longest_w_start_min\", \"longest_w_end_min\", and \"wake_interruption_impact\".\n5. \"longest_w_start_min\" and \"longest_w_end_min\" must be FLOATS in minutes or null.\n6. \"wake_interruption_impact\" must be exactly one of \"low\", \"medium\", \"high\", or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_w_start_min\": , \"longest_w_end_min\": , \"wake_interruption_impact\": <\"low\"|\"medium\"|\"high\"|null>}", + "metrics": [ + { + "metric_id": "longest_w_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_start_min", + "weight": 30, + "params": { + "gt_value": 410.0, + "tolerance": 2.0 + } + }, + { + "metric_id": "longest_w_end_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_end_min", + "weight": 30, + "params": { + "gt_value": 422.0, + "tolerance": 2.0 + } + }, + { + "metric_id": "wake_interruption_impact_accuracy", + "type": "categorical_check", + "target_key": "wake_interruption_impact", + "weight": 40, + "params": { + "gt_value": "medium" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case11/case11_17.json b/sleep_assessment/cases/case11/case11_17.json new file mode 100644 index 0000000000000000000000000000000000000000..1edf643a846743a1f65e21c882402bc455a96a9d --- /dev/null +++ b/sleep_assessment/cases/case11/case11_17.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "MASSSS3_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_02.npy", + "instruction": "Please use the provided sleep-stage label file to locate the longest contiguous Wake (W) segment during the nocturnal sleep period. Define the nocturnal sleep period as the interval from the first non-Wake sleep epoch to the last non-Wake sleep epoch, excluding Wake epochs before sleep onset and after final awakening. Report the start time and end time of this longest Wake segment in minutes, using the first non-Wake sleep epoch as 0 minutes. Then classify the impact of this single longest awakening interruption on sleep continuity as low, medium, or high: low for <5 minutes, medium for 5 to <30 minutes, and high for >=30 minutes. Please clearly report all three final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three values from the agent's report:\n1) the start time in minutes of the longest contiguous Wake (W) segment during the sleep period\n2) the end time in minutes of that longest contiguous Wake (W) segment\n3) the sleep-continuity impact level of that Wake interruption\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"longest_w_start_min\", \"longest_w_end_min\", and \"wake_interruption_impact\".\n5. \"longest_w_start_min\" and \"longest_w_end_min\" must be FLOATS in minutes or null.\n6. \"wake_interruption_impact\" must be exactly one of \"low\", \"medium\", \"high\", or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_w_start_min\": , \"longest_w_end_min\": , \"wake_interruption_impact\": <\"low\"|\"medium\"|\"high\"|null>}", + "metrics": [ + { + "metric_id": "longest_w_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_start_min", + "weight": 30, + "params": { + "gt_value": 486.0, + "tolerance": 2.0 + } + }, + { + "metric_id": "longest_w_end_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_end_min", + "weight": 30, + "params": { + "gt_value": 494.5, + "tolerance": 2.0 + } + }, + { + "metric_id": "wake_interruption_impact_accuracy", + "type": "categorical_check", + "target_key": "wake_interruption_impact", + "weight": 40, + "params": { + "gt_value": "medium" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case11/case11_18.json b/sleep_assessment/cases/case11/case11_18.json new file mode 100644 index 0000000000000000000000000000000000000000..7e3e319af21ae40a99ba3708c514067f37ff6571 --- /dev/null +++ b/sleep_assessment/cases/case11/case11_18.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "MASSSS3_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_03.npy", + "instruction": "Please use the provided sleep-stage label file to locate the longest contiguous Wake (W) segment during the nocturnal sleep period. Define the nocturnal sleep period as the interval from the first non-Wake sleep epoch to the last non-Wake sleep epoch, excluding Wake epochs before sleep onset and after final awakening. Report the start time and end time of this longest Wake segment in minutes, using the first non-Wake sleep epoch as 0 minutes. Then classify the impact of this single longest awakening interruption on sleep continuity as low, medium, or high: low for <5 minutes, medium for 5 to <30 minutes, and high for >=30 minutes. Please clearly report all three final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three values from the agent's report:\n1) the start time in minutes of the longest contiguous Wake (W) segment during the sleep period\n2) the end time in minutes of that longest contiguous Wake (W) segment\n3) the sleep-continuity impact level of that Wake interruption\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"longest_w_start_min\", \"longest_w_end_min\", and \"wake_interruption_impact\".\n5. \"longest_w_start_min\" and \"longest_w_end_min\" must be FLOATS in minutes or null.\n6. \"wake_interruption_impact\" must be exactly one of \"low\", \"medium\", \"high\", or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_w_start_min\": , \"longest_w_end_min\": , \"wake_interruption_impact\": <\"low\"|\"medium\"|\"high\"|null>}", + "metrics": [ + { + "metric_id": "longest_w_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_start_min", + "weight": 30, + "params": { + "gt_value": 400.5, + "tolerance": 2.0 + } + }, + { + "metric_id": "longest_w_end_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_end_min", + "weight": 30, + "params": { + "gt_value": 405.0, + "tolerance": 2.0 + } + }, + { + "metric_id": "wake_interruption_impact_accuracy", + "type": "categorical_check", + "target_key": "wake_interruption_impact", + "weight": 40, + "params": { + "gt_value": "low" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case11/case11_19.json b/sleep_assessment/cases/case11/case11_19.json new file mode 100644 index 0000000000000000000000000000000000000000..b2811637d9f2a248fe82e05f35d4fd7eb6079840 --- /dev/null +++ b/sleep_assessment/cases/case11/case11_19.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "MASSSS3_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_04.npy", + "instruction": "Please use the provided sleep-stage label file to locate the longest contiguous Wake (W) segment during the nocturnal sleep period. Define the nocturnal sleep period as the interval from the first non-Wake sleep epoch to the last non-Wake sleep epoch, excluding Wake epochs before sleep onset and after final awakening. Report the start time and end time of this longest Wake segment in minutes, using the first non-Wake sleep epoch as 0 minutes. Then classify the impact of this single longest awakening interruption on sleep continuity as low, medium, or high: low for <5 minutes, medium for 5 to <30 minutes, and high for >=30 minutes. Please clearly report all three final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three values from the agent's report:\n1) the start time in minutes of the longest contiguous Wake (W) segment during the sleep period\n2) the end time in minutes of that longest contiguous Wake (W) segment\n3) the sleep-continuity impact level of that Wake interruption\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"longest_w_start_min\", \"longest_w_end_min\", and \"wake_interruption_impact\".\n5. \"longest_w_start_min\" and \"longest_w_end_min\" must be FLOATS in minutes or null.\n6. \"wake_interruption_impact\" must be exactly one of \"low\", \"medium\", \"high\", or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_w_start_min\": , \"longest_w_end_min\": , \"wake_interruption_impact\": <\"low\"|\"medium\"|\"high\"|null>}", + "metrics": [ + { + "metric_id": "longest_w_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_start_min", + "weight": 30, + "params": { + "gt_value": 174.0, + "tolerance": 2.0 + } + }, + { + "metric_id": "longest_w_end_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_end_min", + "weight": 30, + "params": { + "gt_value": 174.5, + "tolerance": 2.0 + } + }, + { + "metric_id": "wake_interruption_impact_accuracy", + "type": "categorical_check", + "target_key": "wake_interruption_impact", + "weight": 40, + "params": { + "gt_value": "low" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case11/case11_20.json b/sleep_assessment/cases/case11/case11_20.json new file mode 100644 index 0000000000000000000000000000000000000000..f43b324e9f2b28aa6bd15fe29ddfcd6ceb46b7e9 --- /dev/null +++ b/sleep_assessment/cases/case11/case11_20.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "MASSSS3_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_05.npy", + "instruction": "Please use the provided sleep-stage label file to locate the longest contiguous Wake (W) segment during the nocturnal sleep period. Define the nocturnal sleep period as the interval from the first non-Wake sleep epoch to the last non-Wake sleep epoch, excluding Wake epochs before sleep onset and after final awakening. Report the start time and end time of this longest Wake segment in minutes, using the first non-Wake sleep epoch as 0 minutes. Then classify the impact of this single longest awakening interruption on sleep continuity as low, medium, or high: low for <5 minutes, medium for 5 to <30 minutes, and high for >=30 minutes. Please clearly report all three final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three values from the agent's report:\n1) the start time in minutes of the longest contiguous Wake (W) segment during the sleep period\n2) the end time in minutes of that longest contiguous Wake (W) segment\n3) the sleep-continuity impact level of that Wake interruption\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"longest_w_start_min\", \"longest_w_end_min\", and \"wake_interruption_impact\".\n5. \"longest_w_start_min\" and \"longest_w_end_min\" must be FLOATS in minutes or null.\n6. \"wake_interruption_impact\" must be exactly one of \"low\", \"medium\", \"high\", or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_w_start_min\": , \"longest_w_end_min\": , \"wake_interruption_impact\": <\"low\"|\"medium\"|\"high\"|null>}", + "metrics": [ + { + "metric_id": "longest_w_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_start_min", + "weight": 30, + "params": { + "gt_value": 385.0, + "tolerance": 2.0 + } + }, + { + "metric_id": "longest_w_end_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_end_min", + "weight": 30, + "params": { + "gt_value": 390.5, + "tolerance": 2.0 + } + }, + { + "metric_id": "wake_interruption_impact_accuracy", + "type": "categorical_check", + "target_key": "wake_interruption_impact", + "weight": 40, + "params": { + "gt_value": "medium" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case11/case11_21.json b/sleep_assessment/cases/case11/case11_21.json new file mode 100644 index 0000000000000000000000000000000000000000..67c0225b3292232d85280b6ad6ee797461228fcc --- /dev/null +++ b/sleep_assessment/cases/case11/case11_21.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_01.npy", + "instruction": "Please use the provided sleep-stage label file to locate the longest contiguous Wake (W) segment during the nocturnal sleep period. Define the nocturnal sleep period as the interval from the first non-Wake sleep epoch to the last non-Wake sleep epoch, excluding Wake epochs before sleep onset and after final awakening. Report the start time and end time of this longest Wake segment in minutes, using the first non-Wake sleep epoch as 0 minutes. Then classify the impact of this single longest awakening interruption on sleep continuity as low, medium, or high: low for <5 minutes, medium for 5 to <30 minutes, and high for >=30 minutes. Please clearly report all three final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three values from the agent's report:\n1) the start time in minutes of the longest contiguous Wake (W) segment during the sleep period\n2) the end time in minutes of that longest contiguous Wake (W) segment\n3) the sleep-continuity impact level of that Wake interruption\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"longest_w_start_min\", \"longest_w_end_min\", and \"wake_interruption_impact\".\n5. \"longest_w_start_min\" and \"longest_w_end_min\" must be FLOATS in minutes or null.\n6. \"wake_interruption_impact\" must be exactly one of \"low\", \"medium\", \"high\", or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_w_start_min\": , \"longest_w_end_min\": , \"wake_interruption_impact\": <\"low\"|\"medium\"|\"high\"|null>}", + "metrics": [ + { + "metric_id": "longest_w_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_start_min", + "weight": 30, + "params": { + "gt_value": 270.0, + "tolerance": 2.0 + } + }, + { + "metric_id": "longest_w_end_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_end_min", + "weight": 30, + "params": { + "gt_value": 353.0, + "tolerance": 2.0 + } + }, + { + "metric_id": "wake_interruption_impact_accuracy", + "type": "categorical_check", + "target_key": "wake_interruption_impact", + "weight": 40, + "params": { + "gt_value": "high" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case11/case11_22.json b/sleep_assessment/cases/case11/case11_22.json new file mode 100644 index 0000000000000000000000000000000000000000..8f651cf62398a26265d275a900e93ecbdf18eeea --- /dev/null +++ b/sleep_assessment/cases/case11/case11_22.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_02.npy", + "instruction": "Please use the provided sleep-stage label file to locate the longest contiguous Wake (W) segment during the nocturnal sleep period. Define the nocturnal sleep period as the interval from the first non-Wake sleep epoch to the last non-Wake sleep epoch, excluding Wake epochs before sleep onset and after final awakening. Report the start time and end time of this longest Wake segment in minutes, using the first non-Wake sleep epoch as 0 minutes. Then classify the impact of this single longest awakening interruption on sleep continuity as low, medium, or high: low for <5 minutes, medium for 5 to <30 minutes, and high for >=30 minutes. Please clearly report all three final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three values from the agent's report:\n1) the start time in minutes of the longest contiguous Wake (W) segment during the sleep period\n2) the end time in minutes of that longest contiguous Wake (W) segment\n3) the sleep-continuity impact level of that Wake interruption\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"longest_w_start_min\", \"longest_w_end_min\", and \"wake_interruption_impact\".\n5. \"longest_w_start_min\" and \"longest_w_end_min\" must be FLOATS in minutes or null.\n6. \"wake_interruption_impact\" must be exactly one of \"low\", \"medium\", \"high\", or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_w_start_min\": , \"longest_w_end_min\": , \"wake_interruption_impact\": <\"low\"|\"medium\"|\"high\"|null>}", + "metrics": [ + { + "metric_id": "longest_w_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_start_min", + "weight": 30, + "params": { + "gt_value": 81.5, + "tolerance": 2.0 + } + }, + { + "metric_id": "longest_w_end_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_end_min", + "weight": 30, + "params": { + "gt_value": 107.5, + "tolerance": 2.0 + } + }, + { + "metric_id": "wake_interruption_impact_accuracy", + "type": "categorical_check", + "target_key": "wake_interruption_impact", + "weight": 40, + "params": { + "gt_value": "medium" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case11/case11_23.json b/sleep_assessment/cases/case11/case11_23.json new file mode 100644 index 0000000000000000000000000000000000000000..85d5c721c8f89ca8229da53693cf1dd921b841ed --- /dev/null +++ b/sleep_assessment/cases/case11/case11_23.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "Physionet2018_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_03.npy", + "instruction": "Please use the provided sleep-stage label file to locate the longest contiguous Wake (W) segment during the nocturnal sleep period. Define the nocturnal sleep period as the interval from the first non-Wake sleep epoch to the last non-Wake sleep epoch, excluding Wake epochs before sleep onset and after final awakening. Report the start time and end time of this longest Wake segment in minutes, using the first non-Wake sleep epoch as 0 minutes. Then classify the impact of this single longest awakening interruption on sleep continuity as low, medium, or high: low for <5 minutes, medium for 5 to <30 minutes, and high for >=30 minutes. Please clearly report all three final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three values from the agent's report:\n1) the start time in minutes of the longest contiguous Wake (W) segment during the sleep period\n2) the end time in minutes of that longest contiguous Wake (W) segment\n3) the sleep-continuity impact level of that Wake interruption\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"longest_w_start_min\", \"longest_w_end_min\", and \"wake_interruption_impact\".\n5. \"longest_w_start_min\" and \"longest_w_end_min\" must be FLOATS in minutes or null.\n6. \"wake_interruption_impact\" must be exactly one of \"low\", \"medium\", \"high\", or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_w_start_min\": , \"longest_w_end_min\": , \"wake_interruption_impact\": <\"low\"|\"medium\"|\"high\"|null>}", + "metrics": [ + { + "metric_id": "longest_w_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_start_min", + "weight": 30, + "params": { + "gt_value": 131.0, + "tolerance": 2.0 + } + }, + { + "metric_id": "longest_w_end_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_end_min", + "weight": 30, + "params": { + "gt_value": 254.0, + "tolerance": 2.0 + } + }, + { + "metric_id": "wake_interruption_impact_accuracy", + "type": "categorical_check", + "target_key": "wake_interruption_impact", + "weight": 40, + "params": { + "gt_value": "high" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case11/case11_24.json b/sleep_assessment/cases/case11/case11_24.json new file mode 100644 index 0000000000000000000000000000000000000000..6417e0d99f6846ebde677e9fdf30eb908e50e4e9 --- /dev/null +++ b/sleep_assessment/cases/case11/case11_24.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_04.npy", + "instruction": "Please use the provided sleep-stage label file to locate the longest contiguous Wake (W) segment during the nocturnal sleep period. Define the nocturnal sleep period as the interval from the first non-Wake sleep epoch to the last non-Wake sleep epoch, excluding Wake epochs before sleep onset and after final awakening. Report the start time and end time of this longest Wake segment in minutes, using the first non-Wake sleep epoch as 0 minutes. Then classify the impact of this single longest awakening interruption on sleep continuity as low, medium, or high: low for <5 minutes, medium for 5 to <30 minutes, and high for >=30 minutes. Please clearly report all three final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three values from the agent's report:\n1) the start time in minutes of the longest contiguous Wake (W) segment during the sleep period\n2) the end time in minutes of that longest contiguous Wake (W) segment\n3) the sleep-continuity impact level of that Wake interruption\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"longest_w_start_min\", \"longest_w_end_min\", and \"wake_interruption_impact\".\n5. \"longest_w_start_min\" and \"longest_w_end_min\" must be FLOATS in minutes or null.\n6. \"wake_interruption_impact\" must be exactly one of \"low\", \"medium\", \"high\", or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_w_start_min\": , \"longest_w_end_min\": , \"wake_interruption_impact\": <\"low\"|\"medium\"|\"high\"|null>}", + "metrics": [ + { + "metric_id": "longest_w_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_start_min", + "weight": 30, + "params": { + "gt_value": 265.5, + "tolerance": 2.0 + } + }, + { + "metric_id": "longest_w_end_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_end_min", + "weight": 30, + "params": { + "gt_value": 281.0, + "tolerance": 2.0 + } + }, + { + "metric_id": "wake_interruption_impact_accuracy", + "type": "categorical_check", + "target_key": "wake_interruption_impact", + "weight": 40, + "params": { + "gt_value": "medium" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case11/case11_25.json b/sleep_assessment/cases/case11/case11_25.json new file mode 100644 index 0000000000000000000000000000000000000000..d73fd0aa844fd22336fa55d14f6b021b8ea131f4 --- /dev/null +++ b/sleep_assessment/cases/case11/case11_25.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_05.npy", + "instruction": "Please use the provided sleep-stage label file to locate the longest contiguous Wake (W) segment during the nocturnal sleep period. Define the nocturnal sleep period as the interval from the first non-Wake sleep epoch to the last non-Wake sleep epoch, excluding Wake epochs before sleep onset and after final awakening. Report the start time and end time of this longest Wake segment in minutes, using the first non-Wake sleep epoch as 0 minutes. Then classify the impact of this single longest awakening interruption on sleep continuity as low, medium, or high: low for <5 minutes, medium for 5 to <30 minutes, and high for >=30 minutes. Please clearly report all three final values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three values from the agent's report:\n1) the start time in minutes of the longest contiguous Wake (W) segment during the sleep period\n2) the end time in minutes of that longest contiguous Wake (W) segment\n3) the sleep-continuity impact level of that Wake interruption\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"longest_w_start_min\", \"longest_w_end_min\", and \"wake_interruption_impact\".\n5. \"longest_w_start_min\" and \"longest_w_end_min\" must be FLOATS in minutes or null.\n6. \"wake_interruption_impact\" must be exactly one of \"low\", \"medium\", \"high\", or null.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_w_start_min\": , \"longest_w_end_min\": , \"wake_interruption_impact\": <\"low\"|\"medium\"|\"high\"|null>}", + "metrics": [ + { + "metric_id": "longest_w_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_start_min", + "weight": 30, + "params": { + "gt_value": 204.0, + "tolerance": 2.0 + } + }, + { + "metric_id": "longest_w_end_time_accuracy", + "type": "numeric_check", + "target_key": "longest_w_end_min", + "weight": 30, + "params": { + "gt_value": 231.0, + "tolerance": 2.0 + } + }, + { + "metric_id": "wake_interruption_impact_accuracy", + "type": "categorical_check", + "target_key": "wake_interruption_impact", + "weight": 40, + "params": { + "gt_value": "medium" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case12/case12_01.json b/sleep_assessment/cases/case12/case12_01.json new file mode 100644 index 0000000000000000000000000000000000000000..e974afb4ed7d7a0efed4a8f4cd8bc85c63788f19 --- /dev/null +++ b/sleep_assessment/cases/case12/case12_01.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Please use the provided sleep-stage label file to analyze half-night N3 and REM patterns. Split the entire label file into a first half and a second half using the midpoint of the label sequence. Calculate the percentage of total N3 sleep that occurs in the first half of the label file, and the percentage of total REM sleep that occurs in the second half of the label file. Then determine whether deep sleep is concentrated in the first half of the night, defined as first-half N3 being at least 70% of total N3, and whether REM is enhanced in the second half of the night, defined as second-half REM being at least 60% of total REM. If total N3 duration is less than 30 minutes, explicitly state that N3 is limited or short before interpreting whether the limited N3 is mainly concentrated in the first half. Please clearly report both percentage values and the two final interpretations." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two percentage values from the agent's report:\n1) the percentage of total N3 sleep that occurs in the first half of the night\n2) the percentage of total REM sleep that occurs in the second half of the night\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_n3_ratio_percent\" and \"second_half_rem_ratio_percent\".\n5. Both values must be FLOATS or null.\n6. Values are percentages; if the report says 72.5%, output 72.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_n3_ratio_percent\": , \"second_half_rem_ratio_percent\": }", + "metrics": [ + { + "metric_id": "first_half_n3_ratio_accuracy", + "type": "numeric_check", + "target_key": "first_half_n3_ratio_percent", + "weight": 30, + "params": { + "gt_value": 70.99567099567099, + "tolerance": 3.54978354978355 + } + }, + { + "metric_id": "second_half_rem_ratio_accuracy", + "type": "numeric_check", + "target_key": "second_half_rem_ratio_percent", + "weight": 30, + "params": { + "gt_value": 45.76271186440678, + "tolerance": 2.288135593220339 + } + }, + { + "metric_id": "half_night_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_01.edf. The first-half N3 ratio is 70.995671% of total N3, the second-half REM ratio is 45.762712% of total REM, total N3 duration is 115.500000 minutes, and total REM duration is 59.000000 minutes. The benchmark rules are: first-half deep-sleep concentration is true only when first-half N3 is >=70% of total N3; second-half REM enhancement is true only when second-half REM is >=60% of total REM; total N3 <30 minutes is a short-N3 edge case. The correct response must say that deep sleep is concentrated in the first half of the night. The correct response must say that REM is not enhanced or increased in the second half of the night. Mark the response as correct only if it gets both interpretations right and also handles any short-N3, absent-N3, or absent-REM edge case described above. If the response gives the opposite conclusion, ignores a required short-stage caveat, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case12/case12_02.json b/sleep_assessment/cases/case12/case12_02.json new file mode 100644 index 0000000000000000000000000000000000000000..677c407e317e619e1fa8d56d1533e9eb15eacd1f --- /dev/null +++ b/sleep_assessment/cases/case12/case12_02.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_02.npy", + "instruction": "Please use the provided sleep-stage label file to analyze half-night N3 and REM patterns. Split the entire label file into a first half and a second half using the midpoint of the label sequence. Calculate the percentage of total N3 sleep that occurs in the first half of the label file, and the percentage of total REM sleep that occurs in the second half of the label file. Then determine whether deep sleep is concentrated in the first half of the night, defined as first-half N3 being at least 70% of total N3, and whether REM is enhanced in the second half of the night, defined as second-half REM being at least 60% of total REM. If total N3 duration is less than 30 minutes, explicitly state that N3 is limited or short before interpreting whether the limited N3 is mainly concentrated in the first half. Please clearly report both percentage values and the two final interpretations." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two percentage values from the agent's report:\n1) the percentage of total N3 sleep that occurs in the first half of the night\n2) the percentage of total REM sleep that occurs in the second half of the night\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_n3_ratio_percent\" and \"second_half_rem_ratio_percent\".\n5. Both values must be FLOATS or null.\n6. Values are percentages; if the report says 72.5%, output 72.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_n3_ratio_percent\": , \"second_half_rem_ratio_percent\": }", + "metrics": [ + { + "metric_id": "first_half_n3_ratio_accuracy", + "type": "numeric_check", + "target_key": "first_half_n3_ratio_percent", + "weight": 30, + "params": { + "gt_value": 52.22929936305732, + "tolerance": 2.611464968152866 + } + }, + { + "metric_id": "second_half_rem_ratio_accuracy", + "type": "numeric_check", + "target_key": "second_half_rem_ratio_percent", + "weight": 30, + "params": { + "gt_value": 77.45098039215686, + "tolerance": 3.8725490196078436 + } + }, + { + "metric_id": "half_night_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_02.edf. The first-half N3 ratio is 52.229299% of total N3, the second-half REM ratio is 77.450980% of total REM, total N3 duration is 78.500000 minutes, and total REM duration is 51.000000 minutes. The benchmark rules are: first-half deep-sleep concentration is true only when first-half N3 is >=70% of total N3; second-half REM enhancement is true only when second-half REM is >=60% of total REM; total N3 <30 minutes is a short-N3 edge case. The correct response must say that deep sleep is not concentrated in the first half of the night. The correct response must say that REM is enhanced or increased in the second half of the night. Mark the response as correct only if it gets both interpretations right and also handles any short-N3, absent-N3, or absent-REM edge case described above. If the response gives the opposite conclusion, ignores a required short-stage caveat, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case12/case12_03.json b/sleep_assessment/cases/case12/case12_03.json new file mode 100644 index 0000000000000000000000000000000000000000..9bd757e1d5dcb968b62c4086247c06238c33dfd2 --- /dev/null +++ b/sleep_assessment/cases/case12/case12_03.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_03.npy", + "instruction": "Please use the provided sleep-stage label file to analyze half-night N3 and REM patterns. Split the entire label file into a first half and a second half using the midpoint of the label sequence. Calculate the percentage of total N3 sleep that occurs in the first half of the label file, and the percentage of total REM sleep that occurs in the second half of the label file. Then determine whether deep sleep is concentrated in the first half of the night, defined as first-half N3 being at least 70% of total N3, and whether REM is enhanced in the second half of the night, defined as second-half REM being at least 60% of total REM. If total N3 duration is less than 30 minutes, explicitly state that N3 is limited or short before interpreting whether the limited N3 is mainly concentrated in the first half. Please clearly report both percentage values and the two final interpretations." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two percentage values from the agent's report:\n1) the percentage of total N3 sleep that occurs in the first half of the night\n2) the percentage of total REM sleep that occurs in the second half of the night\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_n3_ratio_percent\" and \"second_half_rem_ratio_percent\".\n5. Both values must be FLOATS or null.\n6. Values are percentages; if the report says 72.5%, output 72.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_n3_ratio_percent\": , \"second_half_rem_ratio_percent\": }", + "metrics": [ + { + "metric_id": "first_half_n3_ratio_accuracy", + "type": "numeric_check", + "target_key": "first_half_n3_ratio_percent", + "weight": 30, + "params": { + "gt_value": 72.83236994219654, + "tolerance": 3.641618497109827 + } + }, + { + "metric_id": "second_half_rem_ratio_accuracy", + "type": "numeric_check", + "target_key": "second_half_rem_ratio_percent", + "weight": 30, + "params": { + "gt_value": 84.14096916299559, + "tolerance": 4.20704845814978 + } + }, + { + "metric_id": "half_night_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_03.edf. The first-half N3 ratio is 72.832370% of total N3, the second-half REM ratio is 84.140969% of total REM, total N3 duration is 86.500000 minutes, and total REM duration is 113.500000 minutes. The benchmark rules are: first-half deep-sleep concentration is true only when first-half N3 is >=70% of total N3; second-half REM enhancement is true only when second-half REM is >=60% of total REM; total N3 <30 minutes is a short-N3 edge case. The correct response must say that deep sleep is concentrated in the first half of the night. The correct response must say that REM is enhanced or increased in the second half of the night. Mark the response as correct only if it gets both interpretations right and also handles any short-N3, absent-N3, or absent-REM edge case described above. If the response gives the opposite conclusion, ignores a required short-stage caveat, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case12/case12_04.json b/sleep_assessment/cases/case12/case12_04.json new file mode 100644 index 0000000000000000000000000000000000000000..e88b50a7dcaa76b727bda0fb6aa6b7074cc6a016 --- /dev/null +++ b/sleep_assessment/cases/case12/case12_04.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "ISRUC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_04.npy", + "instruction": "Please use the provided sleep-stage label file to analyze half-night N3 and REM patterns. Split the entire label file into a first half and a second half using the midpoint of the label sequence. Calculate the percentage of total N3 sleep that occurs in the first half of the label file, and the percentage of total REM sleep that occurs in the second half of the label file. Then determine whether deep sleep is concentrated in the first half of the night, defined as first-half N3 being at least 70% of total N3, and whether REM is enhanced in the second half of the night, defined as second-half REM being at least 60% of total REM. If total N3 duration is less than 30 minutes, explicitly state that N3 is limited or short before interpreting whether the limited N3 is mainly concentrated in the first half. Please clearly report both percentage values and the two final interpretations." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two percentage values from the agent's report:\n1) the percentage of total N3 sleep that occurs in the first half of the night\n2) the percentage of total REM sleep that occurs in the second half of the night\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_n3_ratio_percent\" and \"second_half_rem_ratio_percent\".\n5. Both values must be FLOATS or null.\n6. Values are percentages; if the report says 72.5%, output 72.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_n3_ratio_percent\": , \"second_half_rem_ratio_percent\": }", + "metrics": [ + { + "metric_id": "first_half_n3_ratio_accuracy", + "type": "numeric_check", + "target_key": "first_half_n3_ratio_percent", + "weight": 30, + "params": { + "gt_value": 85.04672897196262, + "tolerance": 4.252336448598132 + } + }, + { + "metric_id": "second_half_rem_ratio_accuracy", + "type": "numeric_check", + "target_key": "second_half_rem_ratio_percent", + "weight": 30, + "params": { + "gt_value": 60.43478260869565, + "tolerance": 3.0217391304347827 + } + }, + { + "metric_id": "half_night_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_04.edf. The first-half N3 ratio is 85.046729% of total N3, the second-half REM ratio is 60.434783% of total REM, total N3 duration is 107.000000 minutes, and total REM duration is 115.000000 minutes. The benchmark rules are: first-half deep-sleep concentration is true only when first-half N3 is >=70% of total N3; second-half REM enhancement is true only when second-half REM is >=60% of total REM; total N3 <30 minutes is a short-N3 edge case. The correct response must say that deep sleep is concentrated in the first half of the night. The correct response must say that REM is enhanced or increased in the second half of the night. Mark the response as correct only if it gets both interpretations right and also handles any short-N3, absent-N3, or absent-REM edge case described above. If the response gives the opposite conclusion, ignores a required short-stage caveat, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case12/case12_05.json b/sleep_assessment/cases/case12/case12_05.json new file mode 100644 index 0000000000000000000000000000000000000000..bfc7d8f9ca32588193b2d070c69bd973b72135e0 --- /dev/null +++ b/sleep_assessment/cases/case12/case12_05.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_05.npy", + "instruction": "Please use the provided sleep-stage label file to analyze half-night N3 and REM patterns. Split the entire label file into a first half and a second half using the midpoint of the label sequence. Calculate the percentage of total N3 sleep that occurs in the first half of the label file, and the percentage of total REM sleep that occurs in the second half of the label file. Then determine whether deep sleep is concentrated in the first half of the night, defined as first-half N3 being at least 70% of total N3, and whether REM is enhanced in the second half of the night, defined as second-half REM being at least 60% of total REM. If total N3 duration is less than 30 minutes, explicitly state that N3 is limited or short before interpreting whether the limited N3 is mainly concentrated in the first half. Please clearly report both percentage values and the two final interpretations." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two percentage values from the agent's report:\n1) the percentage of total N3 sleep that occurs in the first half of the night\n2) the percentage of total REM sleep that occurs in the second half of the night\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_n3_ratio_percent\" and \"second_half_rem_ratio_percent\".\n5. Both values must be FLOATS or null.\n6. Values are percentages; if the report says 72.5%, output 72.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_n3_ratio_percent\": , \"second_half_rem_ratio_percent\": }", + "metrics": [ + { + "metric_id": "first_half_n3_ratio_accuracy", + "type": "numeric_check", + "target_key": "first_half_n3_ratio_percent", + "weight": 30, + "params": { + "gt_value": 60.36585365853659, + "tolerance": 3.0182926829268295 + } + }, + { + "metric_id": "second_half_rem_ratio_accuracy", + "type": "numeric_check", + "target_key": "second_half_rem_ratio_percent", + "weight": 30, + "params": { + "gt_value": 95.23809523809524, + "tolerance": 4.761904761904762 + } + }, + { + "metric_id": "half_night_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_05.edf. The first-half N3 ratio is 60.365854% of total N3, the second-half REM ratio is 95.238095% of total REM, total N3 duration is 82.000000 minutes, and total REM duration is 21.000000 minutes. The benchmark rules are: first-half deep-sleep concentration is true only when first-half N3 is >=70% of total N3; second-half REM enhancement is true only when second-half REM is >=60% of total REM; total N3 <30 minutes is a short-N3 edge case. The correct response must say that deep sleep is not concentrated in the first half of the night. The correct response must say that REM is enhanced or increased in the second half of the night. Mark the response as correct only if it gets both interpretations right and also handles any short-N3, absent-N3, or absent-REM edge case described above. If the response gives the opposite conclusion, ignores a required short-stage caveat, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case12/case12_06.json b/sleep_assessment/cases/case12/case12_06.json new file mode 100644 index 0000000000000000000000000000000000000000..3b477acf25d8797cf1ba1b39ffe27df5c38db0a9 --- /dev/null +++ b/sleep_assessment/cases/case12/case12_06.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "HMC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_01.npy", + "instruction": "Please use the provided sleep-stage label file to analyze half-night N3 and REM patterns. Split the entire label file into a first half and a second half using the midpoint of the label sequence. Calculate the percentage of total N3 sleep that occurs in the first half of the label file, and the percentage of total REM sleep that occurs in the second half of the label file. Then determine whether deep sleep is concentrated in the first half of the night, defined as first-half N3 being at least 70% of total N3, and whether REM is enhanced in the second half of the night, defined as second-half REM being at least 60% of total REM. If total N3 duration is less than 30 minutes, explicitly state that N3 is limited or short before interpreting whether the limited N3 is mainly concentrated in the first half. Please clearly report both percentage values and the two final interpretations." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two percentage values from the agent's report:\n1) the percentage of total N3 sleep that occurs in the first half of the night\n2) the percentage of total REM sleep that occurs in the second half of the night\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_n3_ratio_percent\" and \"second_half_rem_ratio_percent\".\n5. Both values must be FLOATS or null.\n6. Values are percentages; if the report says 72.5%, output 72.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_n3_ratio_percent\": , \"second_half_rem_ratio_percent\": }", + "metrics": [ + { + "metric_id": "first_half_n3_ratio_accuracy", + "type": "numeric_check", + "target_key": "first_half_n3_ratio_percent", + "weight": 30, + "params": { + "gt_value": 56.52173913043478, + "tolerance": 2.8260869565217392 + } + }, + { + "metric_id": "second_half_rem_ratio_accuracy", + "type": "numeric_check", + "target_key": "second_half_rem_ratio_percent", + "weight": 30, + "params": { + "gt_value": 81.56028368794327, + "tolerance": 4.078014184397164 + } + }, + { + "metric_id": "half_night_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is HMC_01.edf. The first-half N3 ratio is 56.521739% of total N3, the second-half REM ratio is 81.560284% of total REM, total N3 duration is 11.500000 minutes, and total REM duration is 70.500000 minutes. The benchmark rules are: first-half deep-sleep concentration is true only when first-half N3 is >=70% of total N3; second-half REM enhancement is true only when second-half REM is >=60% of total REM; total N3 <30 minutes is a short-N3 edge case. The correct response must say that total N3 is limited or short, and that deep sleep is not concentrated in the first half of the night. The correct response must say that REM is enhanced or increased in the second half of the night. Mark the response as correct only if it gets both interpretations right and also handles any short-N3, absent-N3, or absent-REM edge case described above. If the response gives the opposite conclusion, ignores a required short-stage caveat, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case12/case12_07.json b/sleep_assessment/cases/case12/case12_07.json new file mode 100644 index 0000000000000000000000000000000000000000..80b9a6d6aaa2e9e2d22f9933439135392ef44ba7 --- /dev/null +++ b/sleep_assessment/cases/case12/case12_07.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "HMC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_02.npy", + "instruction": "Please use the provided sleep-stage label file to analyze half-night N3 and REM patterns. Split the entire label file into a first half and a second half using the midpoint of the label sequence. Calculate the percentage of total N3 sleep that occurs in the first half of the label file, and the percentage of total REM sleep that occurs in the second half of the label file. Then determine whether deep sleep is concentrated in the first half of the night, defined as first-half N3 being at least 70% of total N3, and whether REM is enhanced in the second half of the night, defined as second-half REM being at least 60% of total REM. If total N3 duration is less than 30 minutes, explicitly state that N3 is limited or short before interpreting whether the limited N3 is mainly concentrated in the first half. Please clearly report both percentage values and the two final interpretations." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two percentage values from the agent's report:\n1) the percentage of total N3 sleep that occurs in the first half of the night\n2) the percentage of total REM sleep that occurs in the second half of the night\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_n3_ratio_percent\" and \"second_half_rem_ratio_percent\".\n5. Both values must be FLOATS or null.\n6. Values are percentages; if the report says 72.5%, output 72.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_n3_ratio_percent\": , \"second_half_rem_ratio_percent\": }", + "metrics": [ + { + "metric_id": "first_half_n3_ratio_accuracy", + "type": "numeric_check", + "target_key": "first_half_n3_ratio_percent", + "weight": 30, + "params": { + "gt_value": 68.30357142857143, + "tolerance": 3.4151785714285716 + } + }, + { + "metric_id": "second_half_rem_ratio_accuracy", + "type": "numeric_check", + "target_key": "second_half_rem_ratio_percent", + "weight": 30, + "params": { + "gt_value": 58.04597701149425, + "tolerance": 2.902298850574713 + } + }, + { + "metric_id": "half_night_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is HMC_02.edf. The first-half N3 ratio is 68.303571% of total N3, the second-half REM ratio is 58.045977% of total REM, total N3 duration is 112.000000 minutes, and total REM duration is 87.000000 minutes. The benchmark rules are: first-half deep-sleep concentration is true only when first-half N3 is >=70% of total N3; second-half REM enhancement is true only when second-half REM is >=60% of total REM; total N3 <30 minutes is a short-N3 edge case. The correct response must say that deep sleep is not concentrated in the first half of the night. The correct response must say that REM is not enhanced or increased in the second half of the night. Mark the response as correct only if it gets both interpretations right and also handles any short-N3, absent-N3, or absent-REM edge case described above. If the response gives the opposite conclusion, ignores a required short-stage caveat, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case12/case12_08.json b/sleep_assessment/cases/case12/case12_08.json new file mode 100644 index 0000000000000000000000000000000000000000..103d42d3d9b6026189dcc43a1b0bdb20e5451ca6 --- /dev/null +++ b/sleep_assessment/cases/case12/case12_08.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "HMC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_03.npy", + "instruction": "Please use the provided sleep-stage label file to analyze half-night N3 and REM patterns. Split the entire label file into a first half and a second half using the midpoint of the label sequence. Calculate the percentage of total N3 sleep that occurs in the first half of the label file, and the percentage of total REM sleep that occurs in the second half of the label file. Then determine whether deep sleep is concentrated in the first half of the night, defined as first-half N3 being at least 70% of total N3, and whether REM is enhanced in the second half of the night, defined as second-half REM being at least 60% of total REM. If total N3 duration is less than 30 minutes, explicitly state that N3 is limited or short before interpreting whether the limited N3 is mainly concentrated in the first half. Please clearly report both percentage values and the two final interpretations." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two percentage values from the agent's report:\n1) the percentage of total N3 sleep that occurs in the first half of the night\n2) the percentage of total REM sleep that occurs in the second half of the night\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_n3_ratio_percent\" and \"second_half_rem_ratio_percent\".\n5. Both values must be FLOATS or null.\n6. Values are percentages; if the report says 72.5%, output 72.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_n3_ratio_percent\": , \"second_half_rem_ratio_percent\": }", + "metrics": [ + { + "metric_id": "first_half_n3_ratio_accuracy", + "type": "numeric_check", + "target_key": "first_half_n3_ratio_percent", + "weight": 30, + "params": { + "gt_value": 66.0, + "tolerance": 3.3000000000000003 + } + }, + { + "metric_id": "second_half_rem_ratio_accuracy", + "type": "numeric_check", + "target_key": "second_half_rem_ratio_percent", + "weight": 30, + "params": { + "gt_value": 100.0, + "tolerance": 5.0 + } + }, + { + "metric_id": "half_night_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is HMC_03.edf. The first-half N3 ratio is 66.000000% of total N3, the second-half REM ratio is 100.000000% of total REM, total N3 duration is 75.000000 minutes, and total REM duration is 76.000000 minutes. The benchmark rules are: first-half deep-sleep concentration is true only when first-half N3 is >=70% of total N3; second-half REM enhancement is true only when second-half REM is >=60% of total REM; total N3 <30 minutes is a short-N3 edge case. The correct response must say that deep sleep is not concentrated in the first half of the night. The correct response must say that REM is enhanced or increased in the second half of the night. Mark the response as correct only if it gets both interpretations right and also handles any short-N3, absent-N3, or absent-REM edge case described above. If the response gives the opposite conclusion, ignores a required short-stage caveat, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case12/case12_09.json b/sleep_assessment/cases/case12/case12_09.json new file mode 100644 index 0000000000000000000000000000000000000000..b4e2113186b22f1e20b16ac5c6b534bbc504e15c --- /dev/null +++ b/sleep_assessment/cases/case12/case12_09.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "HMC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_04.npy", + "instruction": "Please use the provided sleep-stage label file to analyze half-night N3 and REM patterns. Split the entire label file into a first half and a second half using the midpoint of the label sequence. Calculate the percentage of total N3 sleep that occurs in the first half of the label file, and the percentage of total REM sleep that occurs in the second half of the label file. Then determine whether deep sleep is concentrated in the first half of the night, defined as first-half N3 being at least 70% of total N3, and whether REM is enhanced in the second half of the night, defined as second-half REM being at least 60% of total REM. If total N3 duration is less than 30 minutes, explicitly state that N3 is limited or short before interpreting whether the limited N3 is mainly concentrated in the first half. Please clearly report both percentage values and the two final interpretations." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two percentage values from the agent's report:\n1) the percentage of total N3 sleep that occurs in the first half of the night\n2) the percentage of total REM sleep that occurs in the second half of the night\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_n3_ratio_percent\" and \"second_half_rem_ratio_percent\".\n5. Both values must be FLOATS or null.\n6. Values are percentages; if the report says 72.5%, output 72.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_n3_ratio_percent\": , \"second_half_rem_ratio_percent\": }", + "metrics": [ + { + "metric_id": "first_half_n3_ratio_accuracy", + "type": "numeric_check", + "target_key": "first_half_n3_ratio_percent", + "weight": 30, + "params": { + "gt_value": 100.0, + "tolerance": 5.0 + } + }, + { + "metric_id": "second_half_rem_ratio_accuracy", + "type": "numeric_check", + "target_key": "second_half_rem_ratio_percent", + "weight": 30, + "params": { + "gt_value": 52.22672064777328, + "tolerance": 2.611336032388664 + } + }, + { + "metric_id": "half_night_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is HMC_04.edf. The first-half N3 ratio is 100.000000% of total N3, the second-half REM ratio is 52.226721% of total REM, total N3 duration is 32.000000 minutes, and total REM duration is 123.500000 minutes. The benchmark rules are: first-half deep-sleep concentration is true only when first-half N3 is >=70% of total N3; second-half REM enhancement is true only when second-half REM is >=60% of total REM; total N3 <30 minutes is a short-N3 edge case. The correct response must say that deep sleep is concentrated in the first half of the night. The correct response must say that REM is not enhanced or increased in the second half of the night. Mark the response as correct only if it gets both interpretations right and also handles any short-N3, absent-N3, or absent-REM edge case described above. If the response gives the opposite conclusion, ignores a required short-stage caveat, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case12/case12_10.json b/sleep_assessment/cases/case12/case12_10.json new file mode 100644 index 0000000000000000000000000000000000000000..8f9dba7f5a07b1e1747616f9b14f24f73a4698cc --- /dev/null +++ b/sleep_assessment/cases/case12/case12_10.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "HMC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_05.npy", + "instruction": "Please use the provided sleep-stage label file to analyze half-night N3 and REM patterns. Split the entire label file into a first half and a second half using the midpoint of the label sequence. Calculate the percentage of total N3 sleep that occurs in the first half of the label file, and the percentage of total REM sleep that occurs in the second half of the label file. Then determine whether deep sleep is concentrated in the first half of the night, defined as first-half N3 being at least 70% of total N3, and whether REM is enhanced in the second half of the night, defined as second-half REM being at least 60% of total REM. If total N3 duration is less than 30 minutes, explicitly state that N3 is limited or short before interpreting whether the limited N3 is mainly concentrated in the first half. Please clearly report both percentage values and the two final interpretations." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two percentage values from the agent's report:\n1) the percentage of total N3 sleep that occurs in the first half of the night\n2) the percentage of total REM sleep that occurs in the second half of the night\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_n3_ratio_percent\" and \"second_half_rem_ratio_percent\".\n5. Both values must be FLOATS or null.\n6. Values are percentages; if the report says 72.5%, output 72.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_n3_ratio_percent\": , \"second_half_rem_ratio_percent\": }", + "metrics": [ + { + "metric_id": "first_half_n3_ratio_accuracy", + "type": "numeric_check", + "target_key": "first_half_n3_ratio_percent", + "weight": 30, + "params": { + "gt_value": 72.48322147651007, + "tolerance": 3.6241610738255035 + } + }, + { + "metric_id": "second_half_rem_ratio_accuracy", + "type": "numeric_check", + "target_key": "second_half_rem_ratio_percent", + "weight": 30, + "params": { + "gt_value": 57.48031496062992, + "tolerance": 2.8740157480314963 + } + }, + { + "metric_id": "half_night_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is HMC_05.edf. The first-half N3 ratio is 72.483221% of total N3, the second-half REM ratio is 57.480315% of total REM, total N3 duration is 74.500000 minutes, and total REM duration is 63.500000 minutes. The benchmark rules are: first-half deep-sleep concentration is true only when first-half N3 is >=70% of total N3; second-half REM enhancement is true only when second-half REM is >=60% of total REM; total N3 <30 minutes is a short-N3 edge case. The correct response must say that deep sleep is concentrated in the first half of the night. The correct response must say that REM is not enhanced or increased in the second half of the night. Mark the response as correct only if it gets both interpretations right and also handles any short-N3, absent-N3, or absent-REM edge case described above. If the response gives the opposite conclusion, ignores a required short-stage caveat, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case12/case12_11.json b/sleep_assessment/cases/case12/case12_11.json new file mode 100644 index 0000000000000000000000000000000000000000..091b7689ea9d2b6c680b6f8971f1ba4ba7332373 --- /dev/null +++ b/sleep_assessment/cases/case12/case12_11.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_01.npy", + "instruction": "Please use the provided sleep-stage label file to analyze half-night N3 and REM patterns. Split the entire label file into a first half and a second half using the midpoint of the label sequence. Calculate the percentage of total N3 sleep that occurs in the first half of the label file, and the percentage of total REM sleep that occurs in the second half of the label file. Then determine whether deep sleep is concentrated in the first half of the night, defined as first-half N3 being at least 70% of total N3, and whether REM is enhanced in the second half of the night, defined as second-half REM being at least 60% of total REM. If total N3 duration is less than 30 minutes, explicitly state that N3 is limited or short before interpreting whether the limited N3 is mainly concentrated in the first half. Please clearly report both percentage values and the two final interpretations." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two percentage values from the agent's report:\n1) the percentage of total N3 sleep that occurs in the first half of the night\n2) the percentage of total REM sleep that occurs in the second half of the night\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_n3_ratio_percent\" and \"second_half_rem_ratio_percent\".\n5. Both values must be FLOATS or null.\n6. Values are percentages; if the report says 72.5%, output 72.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_n3_ratio_percent\": , \"second_half_rem_ratio_percent\": }", + "metrics": [ + { + "metric_id": "first_half_n3_ratio_accuracy", + "type": "numeric_check", + "target_key": "first_half_n3_ratio_percent", + "weight": 30, + "params": { + "gt_value": 97.93103448275862, + "tolerance": 4.8965517241379315 + } + }, + { + "metric_id": "second_half_rem_ratio_accuracy", + "type": "numeric_check", + "target_key": "second_half_rem_ratio_percent", + "weight": 30, + "params": { + "gt_value": 51.96078431372549, + "tolerance": 2.5980392156862746 + } + }, + { + "metric_id": "half_night_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_01.edf. The first-half N3 ratio is 97.931034% of total N3, the second-half REM ratio is 51.960784% of total REM, total N3 duration is 72.500000 minutes, and total REM duration is 51.000000 minutes. The benchmark rules are: first-half deep-sleep concentration is true only when first-half N3 is >=70% of total N3; second-half REM enhancement is true only when second-half REM is >=60% of total REM; total N3 <30 minutes is a short-N3 edge case. The correct response must say that deep sleep is concentrated in the first half of the night. The correct response must say that REM is not enhanced or increased in the second half of the night. Mark the response as correct only if it gets both interpretations right and also handles any short-N3, absent-N3, or absent-REM edge case described above. If the response gives the opposite conclusion, ignores a required short-stage caveat, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case12/case12_12.json b/sleep_assessment/cases/case12/case12_12.json new file mode 100644 index 0000000000000000000000000000000000000000..69350db2ed621f076ce118e9b97aeea2695fad23 --- /dev/null +++ b/sleep_assessment/cases/case12/case12_12.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_02.npy", + "instruction": "Please use the provided sleep-stage label file to analyze half-night N3 and REM patterns. Split the entire label file into a first half and a second half using the midpoint of the label sequence. Calculate the percentage of total N3 sleep that occurs in the first half of the label file, and the percentage of total REM sleep that occurs in the second half of the label file. Then determine whether deep sleep is concentrated in the first half of the night, defined as first-half N3 being at least 70% of total N3, and whether REM is enhanced in the second half of the night, defined as second-half REM being at least 60% of total REM. If total N3 duration is less than 30 minutes, explicitly state that N3 is limited or short before interpreting whether the limited N3 is mainly concentrated in the first half. Please clearly report both percentage values and the two final interpretations." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two percentage values from the agent's report:\n1) the percentage of total N3 sleep that occurs in the first half of the night\n2) the percentage of total REM sleep that occurs in the second half of the night\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_n3_ratio_percent\" and \"second_half_rem_ratio_percent\".\n5. Both values must be FLOATS or null.\n6. Values are percentages; if the report says 72.5%, output 72.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_n3_ratio_percent\": , \"second_half_rem_ratio_percent\": }", + "metrics": [ + { + "metric_id": "first_half_n3_ratio_accuracy", + "type": "numeric_check", + "target_key": "first_half_n3_ratio_percent", + "weight": 30, + "params": { + "gt_value": 100.0, + "tolerance": 5.0 + } + }, + { + "metric_id": "second_half_rem_ratio_accuracy", + "type": "numeric_check", + "target_key": "second_half_rem_ratio_percent", + "weight": 30, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + }, + { + "metric_id": "half_night_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_02.edf. The first-half N3 ratio is 100.000000% of total N3, the second-half REM ratio is 0.000000% of total REM, total N3 duration is 30.500000 minutes, and total REM duration is 30.500000 minutes. The benchmark rules are: first-half deep-sleep concentration is true only when first-half N3 is >=70% of total N3; second-half REM enhancement is true only when second-half REM is >=60% of total REM; total N3 <30 minutes is a short-N3 edge case. The correct response must say that deep sleep is concentrated in the first half of the night. The correct response must say that REM is not enhanced or increased in the second half of the night. Mark the response as correct only if it gets both interpretations right and also handles any short-N3, absent-N3, or absent-REM edge case described above. If the response gives the opposite conclusion, ignores a required short-stage caveat, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case12/case12_13.json b/sleep_assessment/cases/case12/case12_13.json new file mode 100644 index 0000000000000000000000000000000000000000..88db810a488be830f272f553440b2bde3f9eaff6 --- /dev/null +++ b/sleep_assessment/cases/case12/case12_13.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_03.npy", + "instruction": "Please use the provided sleep-stage label file to analyze half-night N3 and REM patterns. Split the entire label file into a first half and a second half using the midpoint of the label sequence. Calculate the percentage of total N3 sleep that occurs in the first half of the label file, and the percentage of total REM sleep that occurs in the second half of the label file. Then determine whether deep sleep is concentrated in the first half of the night, defined as first-half N3 being at least 70% of total N3, and whether REM is enhanced in the second half of the night, defined as second-half REM being at least 60% of total REM. If total N3 duration is less than 30 minutes, explicitly state that N3 is limited or short before interpreting whether the limited N3 is mainly concentrated in the first half. Please clearly report both percentage values and the two final interpretations." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two percentage values from the agent's report:\n1) the percentage of total N3 sleep that occurs in the first half of the night\n2) the percentage of total REM sleep that occurs in the second half of the night\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_n3_ratio_percent\" and \"second_half_rem_ratio_percent\".\n5. Both values must be FLOATS or null.\n6. Values are percentages; if the report says 72.5%, output 72.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_n3_ratio_percent\": , \"second_half_rem_ratio_percent\": }", + "metrics": [ + { + "metric_id": "first_half_n3_ratio_accuracy", + "type": "numeric_check", + "target_key": "first_half_n3_ratio_percent", + "weight": 30, + "params": { + "gt_value": 57.00325732899023, + "tolerance": 2.850162866449512 + } + }, + { + "metric_id": "second_half_rem_ratio_accuracy", + "type": "numeric_check", + "target_key": "second_half_rem_ratio_percent", + "weight": 30, + "params": { + "gt_value": 81.3953488372093, + "tolerance": 4.069767441860465 + } + }, + { + "metric_id": "half_night_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_03.edf. The first-half N3 ratio is 57.003257% of total N3, the second-half REM ratio is 81.395349% of total REM, total N3 duration is 153.500000 minutes, and total REM duration is 43.000000 minutes. The benchmark rules are: first-half deep-sleep concentration is true only when first-half N3 is >=70% of total N3; second-half REM enhancement is true only when second-half REM is >=60% of total REM; total N3 <30 minutes is a short-N3 edge case. The correct response must say that deep sleep is not concentrated in the first half of the night. The correct response must say that REM is enhanced or increased in the second half of the night. Mark the response as correct only if it gets both interpretations right and also handles any short-N3, absent-N3, or absent-REM edge case described above. If the response gives the opposite conclusion, ignores a required short-stage caveat, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case12/case12_14.json b/sleep_assessment/cases/case12/case12_14.json new file mode 100644 index 0000000000000000000000000000000000000000..d9694d76b8f39d9185207a8cd06df99f3a895bee --- /dev/null +++ b/sleep_assessment/cases/case12/case12_14.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_04.npy", + "instruction": "Please use the provided sleep-stage label file to analyze half-night N3 and REM patterns. Split the entire label file into a first half and a second half using the midpoint of the label sequence. Calculate the percentage of total N3 sleep that occurs in the first half of the label file, and the percentage of total REM sleep that occurs in the second half of the label file. Then determine whether deep sleep is concentrated in the first half of the night, defined as first-half N3 being at least 70% of total N3, and whether REM is enhanced in the second half of the night, defined as second-half REM being at least 60% of total REM. If total N3 duration is less than 30 minutes, explicitly state that N3 is limited or short before interpreting whether the limited N3 is mainly concentrated in the first half. Please clearly report both percentage values and the two final interpretations." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two percentage values from the agent's report:\n1) the percentage of total N3 sleep that occurs in the first half of the night\n2) the percentage of total REM sleep that occurs in the second half of the night\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_n3_ratio_percent\" and \"second_half_rem_ratio_percent\".\n5. Both values must be FLOATS or null.\n6. Values are percentages; if the report says 72.5%, output 72.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_n3_ratio_percent\": , \"second_half_rem_ratio_percent\": }", + "metrics": [ + { + "metric_id": "first_half_n3_ratio_accuracy", + "type": "numeric_check", + "target_key": "first_half_n3_ratio_percent", + "weight": 30, + "params": { + "gt_value": 65.39682539682539, + "tolerance": 3.2698412698412698 + } + }, + { + "metric_id": "second_half_rem_ratio_accuracy", + "type": "numeric_check", + "target_key": "second_half_rem_ratio_percent", + "weight": 30, + "params": { + "gt_value": 56.52173913043478, + "tolerance": 2.8260869565217392 + } + }, + { + "metric_id": "half_night_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_04.edf. The first-half N3 ratio is 65.396825% of total N3, the second-half REM ratio is 56.521739% of total REM, total N3 duration is 157.500000 minutes, and total REM duration is 46.000000 minutes. The benchmark rules are: first-half deep-sleep concentration is true only when first-half N3 is >=70% of total N3; second-half REM enhancement is true only when second-half REM is >=60% of total REM; total N3 <30 minutes is a short-N3 edge case. The correct response must say that deep sleep is not concentrated in the first half of the night. The correct response must say that REM is not enhanced or increased in the second half of the night. Mark the response as correct only if it gets both interpretations right and also handles any short-N3, absent-N3, or absent-REM edge case described above. If the response gives the opposite conclusion, ignores a required short-stage caveat, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case12/case12_15.json b/sleep_assessment/cases/case12/case12_15.json new file mode 100644 index 0000000000000000000000000000000000000000..0fbd49990dfd535b285d8dbf7a35c989130cedc4 --- /dev/null +++ b/sleep_assessment/cases/case12/case12_15.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_05.npy", + "instruction": "Please use the provided sleep-stage label file to analyze half-night N3 and REM patterns. Split the entire label file into a first half and a second half using the midpoint of the label sequence. Calculate the percentage of total N3 sleep that occurs in the first half of the label file, and the percentage of total REM sleep that occurs in the second half of the label file. Then determine whether deep sleep is concentrated in the first half of the night, defined as first-half N3 being at least 70% of total N3, and whether REM is enhanced in the second half of the night, defined as second-half REM being at least 60% of total REM. If total N3 duration is less than 30 minutes, explicitly state that N3 is limited or short before interpreting whether the limited N3 is mainly concentrated in the first half. Please clearly report both percentage values and the two final interpretations." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two percentage values from the agent's report:\n1) the percentage of total N3 sleep that occurs in the first half of the night\n2) the percentage of total REM sleep that occurs in the second half of the night\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_n3_ratio_percent\" and \"second_half_rem_ratio_percent\".\n5. Both values must be FLOATS or null.\n6. Values are percentages; if the report says 72.5%, output 72.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_n3_ratio_percent\": , \"second_half_rem_ratio_percent\": }", + "metrics": [ + { + "metric_id": "first_half_n3_ratio_accuracy", + "type": "numeric_check", + "target_key": "first_half_n3_ratio_percent", + "weight": 30, + "params": { + "gt_value": 100.0, + "tolerance": 5.0 + } + }, + { + "metric_id": "second_half_rem_ratio_accuracy", + "type": "numeric_check", + "target_key": "second_half_rem_ratio_percent", + "weight": 30, + "params": { + "gt_value": 65.93406593406593, + "tolerance": 3.2967032967032965 + } + }, + { + "metric_id": "half_night_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_05.edf. The first-half N3 ratio is 100.000000% of total N3, the second-half REM ratio is 65.934066% of total REM, total N3 duration is 49.500000 minutes, and total REM duration is 45.500000 minutes. The benchmark rules are: first-half deep-sleep concentration is true only when first-half N3 is >=70% of total N3; second-half REM enhancement is true only when second-half REM is >=60% of total REM; total N3 <30 minutes is a short-N3 edge case. The correct response must say that deep sleep is concentrated in the first half of the night. The correct response must say that REM is enhanced or increased in the second half of the night. Mark the response as correct only if it gets both interpretations right and also handles any short-N3, absent-N3, or absent-REM edge case described above. If the response gives the opposite conclusion, ignores a required short-stage caveat, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case12/case12_16.json b/sleep_assessment/cases/case12/case12_16.json new file mode 100644 index 0000000000000000000000000000000000000000..21ac88832cc34902d25843866348169396ddead5 --- /dev/null +++ b/sleep_assessment/cases/case12/case12_16.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "MASSSS3_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_01.npy", + "instruction": "Please use the provided sleep-stage label file to analyze half-night N3 and REM patterns. Split the entire label file into a first half and a second half using the midpoint of the label sequence. Calculate the percentage of total N3 sleep that occurs in the first half of the label file, and the percentage of total REM sleep that occurs in the second half of the label file. Then determine whether deep sleep is concentrated in the first half of the night, defined as first-half N3 being at least 70% of total N3, and whether REM is enhanced in the second half of the night, defined as second-half REM being at least 60% of total REM. If total N3 duration is less than 30 minutes, explicitly state that N3 is limited or short before interpreting whether the limited N3 is mainly concentrated in the first half. Please clearly report both percentage values and the two final interpretations." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two percentage values from the agent's report:\n1) the percentage of total N3 sleep that occurs in the first half of the night\n2) the percentage of total REM sleep that occurs in the second half of the night\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_n3_ratio_percent\" and \"second_half_rem_ratio_percent\".\n5. Both values must be FLOATS or null.\n6. Values are percentages; if the report says 72.5%, output 72.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_n3_ratio_percent\": , \"second_half_rem_ratio_percent\": }", + "metrics": [ + { + "metric_id": "first_half_n3_ratio_accuracy", + "type": "numeric_check", + "target_key": "first_half_n3_ratio_percent", + "weight": 30, + "params": { + "gt_value": 85.61643835616438, + "tolerance": 4.280821917808219 + } + }, + { + "metric_id": "second_half_rem_ratio_accuracy", + "type": "numeric_check", + "target_key": "second_half_rem_ratio_percent", + "weight": 30, + "params": { + "gt_value": 64.64088397790056, + "tolerance": 3.232044198895028 + } + }, + { + "metric_id": "half_night_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is MASSSS3_01.edf. The first-half N3 ratio is 85.616438% of total N3, the second-half REM ratio is 64.640884% of total REM, total N3 duration is 73.000000 minutes, and total REM duration is 90.500000 minutes. The benchmark rules are: first-half deep-sleep concentration is true only when first-half N3 is >=70% of total N3; second-half REM enhancement is true only when second-half REM is >=60% of total REM; total N3 <30 minutes is a short-N3 edge case. The correct response must say that deep sleep is concentrated in the first half of the night. The correct response must say that REM is enhanced or increased in the second half of the night. Mark the response as correct only if it gets both interpretations right and also handles any short-N3, absent-N3, or absent-REM edge case described above. If the response gives the opposite conclusion, ignores a required short-stage caveat, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case12/case12_17.json b/sleep_assessment/cases/case12/case12_17.json new file mode 100644 index 0000000000000000000000000000000000000000..503a4b8a6d717e37a10b7aae99ad6e5a282b7bdd --- /dev/null +++ b/sleep_assessment/cases/case12/case12_17.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "MASSSS3_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_02.npy", + "instruction": "Please use the provided sleep-stage label file to analyze half-night N3 and REM patterns. Split the entire label file into a first half and a second half using the midpoint of the label sequence. Calculate the percentage of total N3 sleep that occurs in the first half of the label file, and the percentage of total REM sleep that occurs in the second half of the label file. Then determine whether deep sleep is concentrated in the first half of the night, defined as first-half N3 being at least 70% of total N3, and whether REM is enhanced in the second half of the night, defined as second-half REM being at least 60% of total REM. If total N3 duration is less than 30 minutes, explicitly state that N3 is limited or short before interpreting whether the limited N3 is mainly concentrated in the first half. Please clearly report both percentage values and the two final interpretations." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two percentage values from the agent's report:\n1) the percentage of total N3 sleep that occurs in the first half of the night\n2) the percentage of total REM sleep that occurs in the second half of the night\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_n3_ratio_percent\" and \"second_half_rem_ratio_percent\".\n5. Both values must be FLOATS or null.\n6. Values are percentages; if the report says 72.5%, output 72.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_n3_ratio_percent\": , \"second_half_rem_ratio_percent\": }", + "metrics": [ + { + "metric_id": "first_half_n3_ratio_accuracy", + "type": "numeric_check", + "target_key": "first_half_n3_ratio_percent", + "weight": 30, + "params": { + "gt_value": 89.65517241379311, + "tolerance": 4.482758620689656 + } + }, + { + "metric_id": "second_half_rem_ratio_accuracy", + "type": "numeric_check", + "target_key": "second_half_rem_ratio_percent", + "weight": 30, + "params": { + "gt_value": 54.54545454545455, + "tolerance": 2.7272727272727275 + } + }, + { + "metric_id": "half_night_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is MASSSS3_02.edf. The first-half N3 ratio is 89.655172% of total N3, the second-half REM ratio is 54.545455% of total REM, total N3 duration is 87.000000 minutes, and total REM duration is 121.000000 minutes. The benchmark rules are: first-half deep-sleep concentration is true only when first-half N3 is >=70% of total N3; second-half REM enhancement is true only when second-half REM is >=60% of total REM; total N3 <30 minutes is a short-N3 edge case. The correct response must say that deep sleep is concentrated in the first half of the night. The correct response must say that REM is not enhanced or increased in the second half of the night. Mark the response as correct only if it gets both interpretations right and also handles any short-N3, absent-N3, or absent-REM edge case described above. If the response gives the opposite conclusion, ignores a required short-stage caveat, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case12/case12_18.json b/sleep_assessment/cases/case12/case12_18.json new file mode 100644 index 0000000000000000000000000000000000000000..3ddd0374148d1fb4ff1ef7e89453084a135ce82e --- /dev/null +++ b/sleep_assessment/cases/case12/case12_18.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "MASSSS3_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_03.npy", + "instruction": "Please use the provided sleep-stage label file to analyze half-night N3 and REM patterns. Split the entire label file into a first half and a second half using the midpoint of the label sequence. Calculate the percentage of total N3 sleep that occurs in the first half of the label file, and the percentage of total REM sleep that occurs in the second half of the label file. Then determine whether deep sleep is concentrated in the first half of the night, defined as first-half N3 being at least 70% of total N3, and whether REM is enhanced in the second half of the night, defined as second-half REM being at least 60% of total REM. If total N3 duration is less than 30 minutes, explicitly state that N3 is limited or short before interpreting whether the limited N3 is mainly concentrated in the first half. Please clearly report both percentage values and the two final interpretations." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two percentage values from the agent's report:\n1) the percentage of total N3 sleep that occurs in the first half of the night\n2) the percentage of total REM sleep that occurs in the second half of the night\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_n3_ratio_percent\" and \"second_half_rem_ratio_percent\".\n5. Both values must be FLOATS or null.\n6. Values are percentages; if the report says 72.5%, output 72.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_n3_ratio_percent\": , \"second_half_rem_ratio_percent\": }", + "metrics": [ + { + "metric_id": "first_half_n3_ratio_accuracy", + "type": "numeric_check", + "target_key": "first_half_n3_ratio_percent", + "weight": 30, + "params": { + "gt_value": 84.6774193548387, + "tolerance": 4.233870967741935 + } + }, + { + "metric_id": "second_half_rem_ratio_accuracy", + "type": "numeric_check", + "target_key": "second_half_rem_ratio_percent", + "weight": 30, + "params": { + "gt_value": 66.04477611940298, + "tolerance": 3.3022388059701493 + } + }, + { + "metric_id": "half_night_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is MASSSS3_03.edf. The first-half N3 ratio is 84.677419% of total N3, the second-half REM ratio is 66.044776% of total REM, total N3 duration is 62.000000 minutes, and total REM duration is 134.000000 minutes. The benchmark rules are: first-half deep-sleep concentration is true only when first-half N3 is >=70% of total N3; second-half REM enhancement is true only when second-half REM is >=60% of total REM; total N3 <30 minutes is a short-N3 edge case. The correct response must say that deep sleep is concentrated in the first half of the night. The correct response must say that REM is enhanced or increased in the second half of the night. Mark the response as correct only if it gets both interpretations right and also handles any short-N3, absent-N3, or absent-REM edge case described above. If the response gives the opposite conclusion, ignores a required short-stage caveat, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case12/case12_19.json b/sleep_assessment/cases/case12/case12_19.json new file mode 100644 index 0000000000000000000000000000000000000000..a3ef726cab9570ab8286e0161565fd1baaa96bfb --- /dev/null +++ b/sleep_assessment/cases/case12/case12_19.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "MASSSS3_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_04.npy", + "instruction": "Please use the provided sleep-stage label file to analyze half-night N3 and REM patterns. Split the entire label file into a first half and a second half using the midpoint of the label sequence. Calculate the percentage of total N3 sleep that occurs in the first half of the label file, and the percentage of total REM sleep that occurs in the second half of the label file. Then determine whether deep sleep is concentrated in the first half of the night, defined as first-half N3 being at least 70% of total N3, and whether REM is enhanced in the second half of the night, defined as second-half REM being at least 60% of total REM. If total N3 duration is less than 30 minutes, explicitly state that N3 is limited or short before interpreting whether the limited N3 is mainly concentrated in the first half. Please clearly report both percentage values and the two final interpretations." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two percentage values from the agent's report:\n1) the percentage of total N3 sleep that occurs in the first half of the night\n2) the percentage of total REM sleep that occurs in the second half of the night\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_n3_ratio_percent\" and \"second_half_rem_ratio_percent\".\n5. Both values must be FLOATS or null.\n6. Values are percentages; if the report says 72.5%, output 72.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_n3_ratio_percent\": , \"second_half_rem_ratio_percent\": }", + "metrics": [ + { + "metric_id": "first_half_n3_ratio_accuracy", + "type": "numeric_check", + "target_key": "first_half_n3_ratio_percent", + "weight": 30, + "params": { + "gt_value": 100.0, + "tolerance": 5.0 + } + }, + { + "metric_id": "second_half_rem_ratio_accuracy", + "type": "numeric_check", + "target_key": "second_half_rem_ratio_percent", + "weight": 30, + "params": { + "gt_value": 70.61224489795919, + "tolerance": 3.5306122448979593 + } + }, + { + "metric_id": "half_night_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is MASSSS3_04.edf. The first-half N3 ratio is 100.000000% of total N3, the second-half REM ratio is 70.612245% of total REM, total N3 duration is 32.000000 minutes, and total REM duration is 122.500000 minutes. The benchmark rules are: first-half deep-sleep concentration is true only when first-half N3 is >=70% of total N3; second-half REM enhancement is true only when second-half REM is >=60% of total REM; total N3 <30 minutes is a short-N3 edge case. The correct response must say that deep sleep is concentrated in the first half of the night. The correct response must say that REM is enhanced or increased in the second half of the night. Mark the response as correct only if it gets both interpretations right and also handles any short-N3, absent-N3, or absent-REM edge case described above. If the response gives the opposite conclusion, ignores a required short-stage caveat, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case12/case12_20.json b/sleep_assessment/cases/case12/case12_20.json new file mode 100644 index 0000000000000000000000000000000000000000..8457d907f568dfe15391110f2c4f862d24160a9a --- /dev/null +++ b/sleep_assessment/cases/case12/case12_20.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "MASSSS3_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_05.npy", + "instruction": "Please use the provided sleep-stage label file to analyze half-night N3 and REM patterns. Split the entire label file into a first half and a second half using the midpoint of the label sequence. Calculate the percentage of total N3 sleep that occurs in the first half of the label file, and the percentage of total REM sleep that occurs in the second half of the label file. Then determine whether deep sleep is concentrated in the first half of the night, defined as first-half N3 being at least 70% of total N3, and whether REM is enhanced in the second half of the night, defined as second-half REM being at least 60% of total REM. If total N3 duration is less than 30 minutes, explicitly state that N3 is limited or short before interpreting whether the limited N3 is mainly concentrated in the first half. Please clearly report both percentage values and the two final interpretations." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two percentage values from the agent's report:\n1) the percentage of total N3 sleep that occurs in the first half of the night\n2) the percentage of total REM sleep that occurs in the second half of the night\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_n3_ratio_percent\" and \"second_half_rem_ratio_percent\".\n5. Both values must be FLOATS or null.\n6. Values are percentages; if the report says 72.5%, output 72.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_n3_ratio_percent\": , \"second_half_rem_ratio_percent\": }", + "metrics": [ + { + "metric_id": "first_half_n3_ratio_accuracy", + "type": "numeric_check", + "target_key": "first_half_n3_ratio_percent", + "weight": 30, + "params": { + "gt_value": 97.43589743589743, + "tolerance": 4.871794871794872 + } + }, + { + "metric_id": "second_half_rem_ratio_accuracy", + "type": "numeric_check", + "target_key": "second_half_rem_ratio_percent", + "weight": 30, + "params": { + "gt_value": 80.50541516245487, + "tolerance": 4.025270758122743 + } + }, + { + "metric_id": "half_night_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is MASSSS3_05.edf. The first-half N3 ratio is 97.435897% of total N3, the second-half REM ratio is 80.505415% of total REM, total N3 duration is 97.500000 minutes, and total REM duration is 138.500000 minutes. The benchmark rules are: first-half deep-sleep concentration is true only when first-half N3 is >=70% of total N3; second-half REM enhancement is true only when second-half REM is >=60% of total REM; total N3 <30 minutes is a short-N3 edge case. The correct response must say that deep sleep is concentrated in the first half of the night. The correct response must say that REM is enhanced or increased in the second half of the night. Mark the response as correct only if it gets both interpretations right and also handles any short-N3, absent-N3, or absent-REM edge case described above. If the response gives the opposite conclusion, ignores a required short-stage caveat, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case12/case12_21.json b/sleep_assessment/cases/case12/case12_21.json new file mode 100644 index 0000000000000000000000000000000000000000..1d619ee321dcf986a71ee47a187c9dfb47b95e51 --- /dev/null +++ b/sleep_assessment/cases/case12/case12_21.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_01.npy", + "instruction": "Please use the provided sleep-stage label file to analyze half-night N3 and REM patterns. Split the entire label file into a first half and a second half using the midpoint of the label sequence. Calculate the percentage of total N3 sleep that occurs in the first half of the label file, and the percentage of total REM sleep that occurs in the second half of the label file. Then determine whether deep sleep is concentrated in the first half of the night, defined as first-half N3 being at least 70% of total N3, and whether REM is enhanced in the second half of the night, defined as second-half REM being at least 60% of total REM. If total N3 duration is less than 30 minutes, explicitly state that N3 is limited or short before interpreting whether the limited N3 is mainly concentrated in the first half. Please clearly report both percentage values and the two final interpretations." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two percentage values from the agent's report:\n1) the percentage of total N3 sleep that occurs in the first half of the night\n2) the percentage of total REM sleep that occurs in the second half of the night\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_n3_ratio_percent\" and \"second_half_rem_ratio_percent\".\n5. Both values must be FLOATS or null.\n6. Values are percentages; if the report says 72.5%, output 72.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_n3_ratio_percent\": , \"second_half_rem_ratio_percent\": }", + "metrics": [ + { + "metric_id": "first_half_n3_ratio_accuracy", + "type": "numeric_check", + "target_key": "first_half_n3_ratio_percent", + "weight": 30, + "params": { + "gt_value": 100.0, + "tolerance": 5.0 + } + }, + { + "metric_id": "second_half_rem_ratio_accuracy", + "type": "numeric_check", + "target_key": "second_half_rem_ratio_percent", + "weight": 30, + "params": { + "gt_value": 46.91358024691358, + "tolerance": 2.345679012345679 + } + }, + { + "metric_id": "half_night_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is Physionet2018_01.edf. The first-half N3 ratio is 100.000000% of total N3, the second-half REM ratio is 46.913580% of total REM, total N3 duration is 36.000000 minutes, and total REM duration is 40.500000 minutes. The benchmark rules are: first-half deep-sleep concentration is true only when first-half N3 is >=70% of total N3; second-half REM enhancement is true only when second-half REM is >=60% of total REM; total N3 <30 minutes is a short-N3 edge case. The correct response must say that deep sleep is concentrated in the first half of the night. The correct response must say that REM is not enhanced or increased in the second half of the night. Mark the response as correct only if it gets both interpretations right and also handles any short-N3, absent-N3, or absent-REM edge case described above. If the response gives the opposite conclusion, ignores a required short-stage caveat, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case12/case12_22.json b/sleep_assessment/cases/case12/case12_22.json new file mode 100644 index 0000000000000000000000000000000000000000..6f0cf0397e4303983da0d8d54a3d1c2e6d1869b7 --- /dev/null +++ b/sleep_assessment/cases/case12/case12_22.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_02.npy", + "instruction": "Please use the provided sleep-stage label file to analyze half-night N3 and REM patterns. Split the entire label file into a first half and a second half using the midpoint of the label sequence. Calculate the percentage of total N3 sleep that occurs in the first half of the label file, and the percentage of total REM sleep that occurs in the second half of the label file. Then determine whether deep sleep is concentrated in the first half of the night, defined as first-half N3 being at least 70% of total N3, and whether REM is enhanced in the second half of the night, defined as second-half REM being at least 60% of total REM. If total N3 duration is less than 30 minutes, explicitly state that N3 is limited or short before interpreting whether the limited N3 is mainly concentrated in the first half. Please clearly report both percentage values and the two final interpretations." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two percentage values from the agent's report:\n1) the percentage of total N3 sleep that occurs in the first half of the night\n2) the percentage of total REM sleep that occurs in the second half of the night\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_n3_ratio_percent\" and \"second_half_rem_ratio_percent\".\n5. Both values must be FLOATS or null.\n6. Values are percentages; if the report says 72.5%, output 72.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_n3_ratio_percent\": , \"second_half_rem_ratio_percent\": }", + "metrics": [ + { + "metric_id": "first_half_n3_ratio_accuracy", + "type": "numeric_check", + "target_key": "first_half_n3_ratio_percent", + "weight": 30, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + }, + { + "metric_id": "second_half_rem_ratio_accuracy", + "type": "numeric_check", + "target_key": "second_half_rem_ratio_percent", + "weight": 30, + "params": { + "gt_value": 100.0, + "tolerance": 5.0 + } + }, + { + "metric_id": "half_night_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is Physionet2018_02.edf. The first-half N3 ratio is 0.000000% of total N3, the second-half REM ratio is 100.000000% of total REM, total N3 duration is 10.500000 minutes, and total REM duration is 0.500000 minutes. The benchmark rules are: first-half deep-sleep concentration is true only when first-half N3 is >=70% of total N3; second-half REM enhancement is true only when second-half REM is >=60% of total REM; total N3 <30 minutes is a short-N3 edge case. The correct response must say that total N3 is limited or short, and that deep sleep is not concentrated in the first half of the night. The correct response must say that REM is enhanced or increased in the second half of the night. Mark the response as correct only if it gets both interpretations right and also handles any short-N3, absent-N3, or absent-REM edge case described above. If the response gives the opposite conclusion, ignores a required short-stage caveat, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case12/case12_23.json b/sleep_assessment/cases/case12/case12_23.json new file mode 100644 index 0000000000000000000000000000000000000000..d79fb70c83f7157893cc541ee301c507905b6619 --- /dev/null +++ b/sleep_assessment/cases/case12/case12_23.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "Physionet2018_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_03.npy", + "instruction": "Please use the provided sleep-stage label file to analyze half-night N3 and REM patterns. Split the entire label file into a first half and a second half using the midpoint of the label sequence. Calculate the percentage of total N3 sleep that occurs in the first half of the label file, and the percentage of total REM sleep that occurs in the second half of the label file. Then determine whether deep sleep is concentrated in the first half of the night, defined as first-half N3 being at least 70% of total N3, and whether REM is enhanced in the second half of the night, defined as second-half REM being at least 60% of total REM. If total N3 duration is less than 30 minutes, explicitly state that N3 is limited or short before interpreting whether the limited N3 is mainly concentrated in the first half. Please clearly report both percentage values and the two final interpretations." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two percentage values from the agent's report:\n1) the percentage of total N3 sleep that occurs in the first half of the night\n2) the percentage of total REM sleep that occurs in the second half of the night\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_n3_ratio_percent\" and \"second_half_rem_ratio_percent\".\n5. Both values must be FLOATS or null.\n6. Values are percentages; if the report says 72.5%, output 72.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_n3_ratio_percent\": , \"second_half_rem_ratio_percent\": }", + "metrics": [ + { + "metric_id": "first_half_n3_ratio_accuracy", + "type": "numeric_check", + "target_key": "first_half_n3_ratio_percent", + "weight": 30, + "params": { + "gt_value": 90.3225806451613, + "tolerance": 4.516129032258065 + } + }, + { + "metric_id": "second_half_rem_ratio_accuracy", + "type": "numeric_check", + "target_key": "second_half_rem_ratio_percent", + "weight": 30, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + }, + { + "metric_id": "half_night_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is Physionet2018_03.edf. The first-half N3 ratio is 90.322581% of total N3, the second-half REM ratio is 0.000000% of total REM, total N3 duration is 31.000000 minutes, and total REM duration is 0.500000 minutes. The benchmark rules are: first-half deep-sleep concentration is true only when first-half N3 is >=70% of total N3; second-half REM enhancement is true only when second-half REM is >=60% of total REM; total N3 <30 minutes is a short-N3 edge case. The correct response must say that deep sleep is concentrated in the first half of the night. The correct response must say that REM is not enhanced or increased in the second half of the night. Mark the response as correct only if it gets both interpretations right and also handles any short-N3, absent-N3, or absent-REM edge case described above. If the response gives the opposite conclusion, ignores a required short-stage caveat, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case12/case12_24.json b/sleep_assessment/cases/case12/case12_24.json new file mode 100644 index 0000000000000000000000000000000000000000..23f47adb2b0067e699d1eac99314803b90d7a5b1 --- /dev/null +++ b/sleep_assessment/cases/case12/case12_24.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_04.npy", + "instruction": "Please use the provided sleep-stage label file to analyze half-night N3 and REM patterns. Split the entire label file into a first half and a second half using the midpoint of the label sequence. Calculate the percentage of total N3 sleep that occurs in the first half of the label file, and the percentage of total REM sleep that occurs in the second half of the label file. Then determine whether deep sleep is concentrated in the first half of the night, defined as first-half N3 being at least 70% of total N3, and whether REM is enhanced in the second half of the night, defined as second-half REM being at least 60% of total REM. If total N3 duration is less than 30 minutes, explicitly state that N3 is limited or short before interpreting whether the limited N3 is mainly concentrated in the first half. Please clearly report both percentage values and the two final interpretations." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two percentage values from the agent's report:\n1) the percentage of total N3 sleep that occurs in the first half of the night\n2) the percentage of total REM sleep that occurs in the second half of the night\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_n3_ratio_percent\" and \"second_half_rem_ratio_percent\".\n5. Both values must be FLOATS or null.\n6. Values are percentages; if the report says 72.5%, output 72.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_n3_ratio_percent\": , \"second_half_rem_ratio_percent\": }", + "metrics": [ + { + "metric_id": "first_half_n3_ratio_accuracy", + "type": "numeric_check", + "target_key": "first_half_n3_ratio_percent", + "weight": 30, + "params": { + "gt_value": 58.72093023255814, + "tolerance": 2.9360465116279073 + } + }, + { + "metric_id": "second_half_rem_ratio_accuracy", + "type": "numeric_check", + "target_key": "second_half_rem_ratio_percent", + "weight": 30, + "params": { + "gt_value": 46.666666666666664, + "tolerance": 2.3333333333333335 + } + }, + { + "metric_id": "half_night_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is Physionet2018_04.edf. The first-half N3 ratio is 58.720930% of total N3, the second-half REM ratio is 46.666667% of total REM, total N3 duration is 86.000000 minutes, and total REM duration is 45.000000 minutes. The benchmark rules are: first-half deep-sleep concentration is true only when first-half N3 is >=70% of total N3; second-half REM enhancement is true only when second-half REM is >=60% of total REM; total N3 <30 minutes is a short-N3 edge case. The correct response must say that deep sleep is not concentrated in the first half of the night. The correct response must say that REM is not enhanced or increased in the second half of the night. Mark the response as correct only if it gets both interpretations right and also handles any short-N3, absent-N3, or absent-REM edge case described above. If the response gives the opposite conclusion, ignores a required short-stage caveat, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case12/case12_25.json b/sleep_assessment/cases/case12/case12_25.json new file mode 100644 index 0000000000000000000000000000000000000000..45a05ae2aa1fe77f3bbdcc292910689baa17913c --- /dev/null +++ b/sleep_assessment/cases/case12/case12_25.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_05.npy", + "instruction": "Please use the provided sleep-stage label file to analyze half-night N3 and REM patterns. Split the entire label file into a first half and a second half using the midpoint of the label sequence. Calculate the percentage of total N3 sleep that occurs in the first half of the label file, and the percentage of total REM sleep that occurs in the second half of the label file. Then determine whether deep sleep is concentrated in the first half of the night, defined as first-half N3 being at least 70% of total N3, and whether REM is enhanced in the second half of the night, defined as second-half REM being at least 60% of total REM. If total N3 duration is less than 30 minutes, explicitly state that N3 is limited or short before interpreting whether the limited N3 is mainly concentrated in the first half. Please clearly report both percentage values and the two final interpretations." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two percentage values from the agent's report:\n1) the percentage of total N3 sleep that occurs in the first half of the night\n2) the percentage of total REM sleep that occurs in the second half of the night\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_n3_ratio_percent\" and \"second_half_rem_ratio_percent\".\n5. Both values must be FLOATS or null.\n6. Values are percentages; if the report says 72.5%, output 72.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_n3_ratio_percent\": , \"second_half_rem_ratio_percent\": }", + "metrics": [ + { + "metric_id": "first_half_n3_ratio_accuracy", + "type": "numeric_check", + "target_key": "first_half_n3_ratio_percent", + "weight": 30, + "params": { + "gt_value": 25.252525252525253, + "tolerance": 1.2626262626262628 + } + }, + { + "metric_id": "second_half_rem_ratio_accuracy", + "type": "numeric_check", + "target_key": "second_half_rem_ratio_percent", + "weight": 30, + "params": { + "gt_value": 87.25490196078431, + "tolerance": 4.362745098039215 + } + }, + { + "metric_id": "half_night_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is Physionet2018_05.edf. The first-half N3 ratio is 25.252525% of total N3, the second-half REM ratio is 87.254902% of total REM, total N3 duration is 49.500000 minutes, and total REM duration is 51.000000 minutes. The benchmark rules are: first-half deep-sleep concentration is true only when first-half N3 is >=70% of total N3; second-half REM enhancement is true only when second-half REM is >=60% of total REM; total N3 <30 minutes is a short-N3 edge case. The correct response must say that deep sleep is not concentrated in the first half of the night. The correct response must say that REM is enhanced or increased in the second half of the night. Mark the response as correct only if it gets both interpretations right and also handles any short-N3, absent-N3, or absent-REM edge case described above. If the response gives the opposite conclusion, ignores a required short-stage caveat, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case13/case13_01.json b/sleep_assessment/cases/case13/case13_01.json new file mode 100644 index 0000000000000000000000000000000000000000..13b2fd1e08c62d10e97d534beffa7637c6871a11 --- /dev/null +++ b/sleep_assessment/cases/case13/case13_01.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the Sleep Fragmentation Index (SFI). Define the sleep period as the interval from the first non-Wake epoch to the last non-Wake epoch. Within this sleep period, count the number of awakenings as the number of contiguous Wake (W) bouts, and count sleep-stage shifts as the number of adjacent epoch stage changes where both adjacent epochs are non-Wake sleep stages (N1, N2, N3, or R/REM). Calculate total sleep time as the number of non-Wake epochs multiplied by 30 seconds. SFI is defined as (number of awakenings + sleep-stage shifts) divided by total sleep time in hours, reported as events per hour of total sleep time. Please clearly report the final SFI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the Sleep Fragmentation Index (SFI) from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"sleep_fragmentation_index\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"sleep_fragmentation_index\": null}.\n\n### OUTPUT TEMPLATE\n{\"sleep_fragmentation_index\": }", + "metrics": [ + { + "metric_id": "sleep_fragmentation_index_accuracy", + "type": "numeric_check", + "target_key": "sleep_fragmentation_index", + "weight": 100, + "params": { + "gt_value": 20.454545454545453, + "tolerance": 1.0227272727272727 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case13/case13_02.json b/sleep_assessment/cases/case13/case13_02.json new file mode 100644 index 0000000000000000000000000000000000000000..313cb53408bf21072a93f2efe97d0f4122fda9ed --- /dev/null +++ b/sleep_assessment/cases/case13/case13_02.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_02.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the Sleep Fragmentation Index (SFI). Define the sleep period as the interval from the first non-Wake epoch to the last non-Wake epoch. Within this sleep period, count the number of awakenings as the number of contiguous Wake (W) bouts, and count sleep-stage shifts as the number of adjacent epoch stage changes where both adjacent epochs are non-Wake sleep stages (N1, N2, N3, or R/REM). Calculate total sleep time as the number of non-Wake epochs multiplied by 30 seconds. SFI is defined as (number of awakenings + sleep-stage shifts) divided by total sleep time in hours, reported as events per hour of total sleep time. Please clearly report the final SFI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the Sleep Fragmentation Index (SFI) from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"sleep_fragmentation_index\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"sleep_fragmentation_index\": null}.\n\n### OUTPUT TEMPLATE\n{\"sleep_fragmentation_index\": }", + "metrics": [ + { + "metric_id": "sleep_fragmentation_index_accuracy", + "type": "numeric_check", + "target_key": "sleep_fragmentation_index", + "weight": 100, + "params": { + "gt_value": 24.86787204450626, + "tolerance": 1.2433936022253131 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case13/case13_03.json b/sleep_assessment/cases/case13/case13_03.json new file mode 100644 index 0000000000000000000000000000000000000000..85e14a74203c3a7fce94c0cd42842870fc5159db --- /dev/null +++ b/sleep_assessment/cases/case13/case13_03.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_03.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the Sleep Fragmentation Index (SFI). Define the sleep period as the interval from the first non-Wake epoch to the last non-Wake epoch. Within this sleep period, count the number of awakenings as the number of contiguous Wake (W) bouts, and count sleep-stage shifts as the number of adjacent epoch stage changes where both adjacent epochs are non-Wake sleep stages (N1, N2, N3, or R/REM). Calculate total sleep time as the number of non-Wake epochs multiplied by 30 seconds. SFI is defined as (number of awakenings + sleep-stage shifts) divided by total sleep time in hours, reported as events per hour of total sleep time. Please clearly report the final SFI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the Sleep Fragmentation Index (SFI) from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"sleep_fragmentation_index\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"sleep_fragmentation_index\": null}.\n\n### OUTPUT TEMPLATE\n{\"sleep_fragmentation_index\": }", + "metrics": [ + { + "metric_id": "sleep_fragmentation_index_accuracy", + "type": "numeric_check", + "target_key": "sleep_fragmentation_index", + "weight": 100, + "params": { + "gt_value": 19.827373612823674, + "tolerance": 0.9913686806411838 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case13/case13_04.json b/sleep_assessment/cases/case13/case13_04.json new file mode 100644 index 0000000000000000000000000000000000000000..2f8c60392cbf1edbd3e1cdf2fa73dc74d856f8fa --- /dev/null +++ b/sleep_assessment/cases/case13/case13_04.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "ISRUC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_04.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the Sleep Fragmentation Index (SFI). Define the sleep period as the interval from the first non-Wake epoch to the last non-Wake epoch. Within this sleep period, count the number of awakenings as the number of contiguous Wake (W) bouts, and count sleep-stage shifts as the number of adjacent epoch stage changes where both adjacent epochs are non-Wake sleep stages (N1, N2, N3, or R/REM). Calculate total sleep time as the number of non-Wake epochs multiplied by 30 seconds. SFI is defined as (number of awakenings + sleep-stage shifts) divided by total sleep time in hours, reported as events per hour of total sleep time. Please clearly report the final SFI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the Sleep Fragmentation Index (SFI) from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"sleep_fragmentation_index\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"sleep_fragmentation_index\": null}.\n\n### OUTPUT TEMPLATE\n{\"sleep_fragmentation_index\": }", + "metrics": [ + { + "metric_id": "sleep_fragmentation_index_accuracy", + "type": "numeric_check", + "target_key": "sleep_fragmentation_index", + "weight": 100, + "params": { + "gt_value": 12.834224598930481, + "tolerance": 0.6417112299465241 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case13/case13_05.json b/sleep_assessment/cases/case13/case13_05.json new file mode 100644 index 0000000000000000000000000000000000000000..a915d35193d63a8aa0a3dd06273700db9477be78 --- /dev/null +++ b/sleep_assessment/cases/case13/case13_05.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_05.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the Sleep Fragmentation Index (SFI). Define the sleep period as the interval from the first non-Wake epoch to the last non-Wake epoch. Within this sleep period, count the number of awakenings as the number of contiguous Wake (W) bouts, and count sleep-stage shifts as the number of adjacent epoch stage changes where both adjacent epochs are non-Wake sleep stages (N1, N2, N3, or R/REM). Calculate total sleep time as the number of non-Wake epochs multiplied by 30 seconds. SFI is defined as (number of awakenings + sleep-stage shifts) divided by total sleep time in hours, reported as events per hour of total sleep time. Please clearly report the final SFI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the Sleep Fragmentation Index (SFI) from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"sleep_fragmentation_index\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"sleep_fragmentation_index\": null}.\n\n### OUTPUT TEMPLATE\n{\"sleep_fragmentation_index\": }", + "metrics": [ + { + "metric_id": "sleep_fragmentation_index_accuracy", + "type": "numeric_check", + "target_key": "sleep_fragmentation_index", + "weight": 100, + "params": { + "gt_value": 20.932642487046632, + "tolerance": 1.0466321243523315 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case13/case13_06.json b/sleep_assessment/cases/case13/case13_06.json new file mode 100644 index 0000000000000000000000000000000000000000..78f886bd3980311e97a24d9ef82e6262d7be73ee --- /dev/null +++ b/sleep_assessment/cases/case13/case13_06.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "HMC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_01.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the Sleep Fragmentation Index (SFI). Define the sleep period as the interval from the first non-Wake epoch to the last non-Wake epoch. Within this sleep period, count the number of awakenings as the number of contiguous Wake (W) bouts, and count sleep-stage shifts as the number of adjacent epoch stage changes where both adjacent epochs are non-Wake sleep stages (N1, N2, N3, or R/REM). Calculate total sleep time as the number of non-Wake epochs multiplied by 30 seconds. SFI is defined as (number of awakenings + sleep-stage shifts) divided by total sleep time in hours, reported as events per hour of total sleep time. Please clearly report the final SFI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the Sleep Fragmentation Index (SFI) from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"sleep_fragmentation_index\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"sleep_fragmentation_index\": null}.\n\n### OUTPUT TEMPLATE\n{\"sleep_fragmentation_index\": }", + "metrics": [ + { + "metric_id": "sleep_fragmentation_index_accuracy", + "type": "numeric_check", + "target_key": "sleep_fragmentation_index", + "weight": 100, + "params": { + "gt_value": 14.33854907539118, + "tolerance": 0.716927453769559 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case13/case13_07.json b/sleep_assessment/cases/case13/case13_07.json new file mode 100644 index 0000000000000000000000000000000000000000..acd632b866e5c697968767053af80ff0b7c5a658 --- /dev/null +++ b/sleep_assessment/cases/case13/case13_07.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "HMC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_02.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the Sleep Fragmentation Index (SFI). Define the sleep period as the interval from the first non-Wake epoch to the last non-Wake epoch. Within this sleep period, count the number of awakenings as the number of contiguous Wake (W) bouts, and count sleep-stage shifts as the number of adjacent epoch stage changes where both adjacent epochs are non-Wake sleep stages (N1, N2, N3, or R/REM). Calculate total sleep time as the number of non-Wake epochs multiplied by 30 seconds. SFI is defined as (number of awakenings + sleep-stage shifts) divided by total sleep time in hours, reported as events per hour of total sleep time. Please clearly report the final SFI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the Sleep Fragmentation Index (SFI) from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"sleep_fragmentation_index\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"sleep_fragmentation_index\": null}.\n\n### OUTPUT TEMPLATE\n{\"sleep_fragmentation_index\": }", + "metrics": [ + { + "metric_id": "sleep_fragmentation_index_accuracy", + "type": "numeric_check", + "target_key": "sleep_fragmentation_index", + "weight": 100, + "params": { + "gt_value": 13.7984496124031, + "tolerance": 0.689922480620155 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case13/case13_08.json b/sleep_assessment/cases/case13/case13_08.json new file mode 100644 index 0000000000000000000000000000000000000000..63de99e2129a1080a529c9ac717e5e18da48d80a --- /dev/null +++ b/sleep_assessment/cases/case13/case13_08.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "HMC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_03.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the Sleep Fragmentation Index (SFI). Define the sleep period as the interval from the first non-Wake epoch to the last non-Wake epoch. Within this sleep period, count the number of awakenings as the number of contiguous Wake (W) bouts, and count sleep-stage shifts as the number of adjacent epoch stage changes where both adjacent epochs are non-Wake sleep stages (N1, N2, N3, or R/REM). Calculate total sleep time as the number of non-Wake epochs multiplied by 30 seconds. SFI is defined as (number of awakenings + sleep-stage shifts) divided by total sleep time in hours, reported as events per hour of total sleep time. Please clearly report the final SFI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the Sleep Fragmentation Index (SFI) from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"sleep_fragmentation_index\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"sleep_fragmentation_index\": null}.\n\n### OUTPUT TEMPLATE\n{\"sleep_fragmentation_index\": }", + "metrics": [ + { + "metric_id": "sleep_fragmentation_index_accuracy", + "type": "numeric_check", + "target_key": "sleep_fragmentation_index", + "weight": 100, + "params": { + "gt_value": 25.263157894736842, + "tolerance": 1.2631578947368423 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case13/case13_09.json b/sleep_assessment/cases/case13/case13_09.json new file mode 100644 index 0000000000000000000000000000000000000000..8029c6edcf5ee01b2a0c3dc13cf78457603aaacc --- /dev/null +++ b/sleep_assessment/cases/case13/case13_09.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "HMC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_04.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the Sleep Fragmentation Index (SFI). Define the sleep period as the interval from the first non-Wake epoch to the last non-Wake epoch. Within this sleep period, count the number of awakenings as the number of contiguous Wake (W) bouts, and count sleep-stage shifts as the number of adjacent epoch stage changes where both adjacent epochs are non-Wake sleep stages (N1, N2, N3, or R/REM). Calculate total sleep time as the number of non-Wake epochs multiplied by 30 seconds. SFI is defined as (number of awakenings + sleep-stage shifts) divided by total sleep time in hours, reported as events per hour of total sleep time. Please clearly report the final SFI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the Sleep Fragmentation Index (SFI) from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"sleep_fragmentation_index\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"sleep_fragmentation_index\": null}.\n\n### OUTPUT TEMPLATE\n{\"sleep_fragmentation_index\": }", + "metrics": [ + { + "metric_id": "sleep_fragmentation_index_accuracy", + "type": "numeric_check", + "target_key": "sleep_fragmentation_index", + "weight": 100, + "params": { + "gt_value": 13.461538461538462, + "tolerance": 0.6730769230769231 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case13/case13_10.json b/sleep_assessment/cases/case13/case13_10.json new file mode 100644 index 0000000000000000000000000000000000000000..112242fd0b6182ef8c037a2385a451197a178782 --- /dev/null +++ b/sleep_assessment/cases/case13/case13_10.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "HMC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_05.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the Sleep Fragmentation Index (SFI). Define the sleep period as the interval from the first non-Wake epoch to the last non-Wake epoch. Within this sleep period, count the number of awakenings as the number of contiguous Wake (W) bouts, and count sleep-stage shifts as the number of adjacent epoch stage changes where both adjacent epochs are non-Wake sleep stages (N1, N2, N3, or R/REM). Calculate total sleep time as the number of non-Wake epochs multiplied by 30 seconds. SFI is defined as (number of awakenings + sleep-stage shifts) divided by total sleep time in hours, reported as events per hour of total sleep time. Please clearly report the final SFI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the Sleep Fragmentation Index (SFI) from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"sleep_fragmentation_index\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"sleep_fragmentation_index\": null}.\n\n### OUTPUT TEMPLATE\n{\"sleep_fragmentation_index\": }", + "metrics": [ + { + "metric_id": "sleep_fragmentation_index_accuracy", + "type": "numeric_check", + "target_key": "sleep_fragmentation_index", + "weight": 100, + "params": { + "gt_value": 32.13720316622691, + "tolerance": 1.6068601583113455 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case13/case13_11.json b/sleep_assessment/cases/case13/case13_11.json new file mode 100644 index 0000000000000000000000000000000000000000..c06d7ff056cc1c6949bc6bf69b65a5af63d9c8ee --- /dev/null +++ b/sleep_assessment/cases/case13/case13_11.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_01.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the Sleep Fragmentation Index (SFI). Define the sleep period as the interval from the first non-Wake epoch to the last non-Wake epoch. Within this sleep period, count the number of awakenings as the number of contiguous Wake (W) bouts, and count sleep-stage shifts as the number of adjacent epoch stage changes where both adjacent epochs are non-Wake sleep stages (N1, N2, N3, or R/REM). Calculate total sleep time as the number of non-Wake epochs multiplied by 30 seconds. SFI is defined as (number of awakenings + sleep-stage shifts) divided by total sleep time in hours, reported as events per hour of total sleep time. Please clearly report the final SFI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the Sleep Fragmentation Index (SFI) from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"sleep_fragmentation_index\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"sleep_fragmentation_index\": null}.\n\n### OUTPUT TEMPLATE\n{\"sleep_fragmentation_index\": }", + "metrics": [ + { + "metric_id": "sleep_fragmentation_index_accuracy", + "type": "numeric_check", + "target_key": "sleep_fragmentation_index", + "weight": 100, + "params": { + "gt_value": 21.091877496671103, + "tolerance": 1.0545938748335553 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case13/case13_12.json b/sleep_assessment/cases/case13/case13_12.json new file mode 100644 index 0000000000000000000000000000000000000000..bdcb56ef85b086a83964527343422fd189b10245 --- /dev/null +++ b/sleep_assessment/cases/case13/case13_12.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_02.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the Sleep Fragmentation Index (SFI). Define the sleep period as the interval from the first non-Wake epoch to the last non-Wake epoch. Within this sleep period, count the number of awakenings as the number of contiguous Wake (W) bouts, and count sleep-stage shifts as the number of adjacent epoch stage changes where both adjacent epochs are non-Wake sleep stages (N1, N2, N3, or R/REM). Calculate total sleep time as the number of non-Wake epochs multiplied by 30 seconds. SFI is defined as (number of awakenings + sleep-stage shifts) divided by total sleep time in hours, reported as events per hour of total sleep time. Please clearly report the final SFI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the Sleep Fragmentation Index (SFI) from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"sleep_fragmentation_index\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"sleep_fragmentation_index\": null}.\n\n### OUTPUT TEMPLATE\n{\"sleep_fragmentation_index\": }", + "metrics": [ + { + "metric_id": "sleep_fragmentation_index_accuracy", + "type": "numeric_check", + "target_key": "sleep_fragmentation_index", + "weight": 100, + "params": { + "gt_value": 26.373626373626376, + "tolerance": 1.3186813186813189 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case13/case13_13.json b/sleep_assessment/cases/case13/case13_13.json new file mode 100644 index 0000000000000000000000000000000000000000..9aca30b37ff7733188de3cf67d69c775cf1a3ac7 --- /dev/null +++ b/sleep_assessment/cases/case13/case13_13.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_03.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the Sleep Fragmentation Index (SFI). Define the sleep period as the interval from the first non-Wake epoch to the last non-Wake epoch. Within this sleep period, count the number of awakenings as the number of contiguous Wake (W) bouts, and count sleep-stage shifts as the number of adjacent epoch stage changes where both adjacent epochs are non-Wake sleep stages (N1, N2, N3, or R/REM). Calculate total sleep time as the number of non-Wake epochs multiplied by 30 seconds. SFI is defined as (number of awakenings + sleep-stage shifts) divided by total sleep time in hours, reported as events per hour of total sleep time. Please clearly report the final SFI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the Sleep Fragmentation Index (SFI) from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"sleep_fragmentation_index\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"sleep_fragmentation_index\": null}.\n\n### OUTPUT TEMPLATE\n{\"sleep_fragmentation_index\": }", + "metrics": [ + { + "metric_id": "sleep_fragmentation_index_accuracy", + "type": "numeric_check", + "target_key": "sleep_fragmentation_index", + "weight": 100, + "params": { + "gt_value": 18.410041841004187, + "tolerance": 0.9205020920502094 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case13/case13_14.json b/sleep_assessment/cases/case13/case13_14.json new file mode 100644 index 0000000000000000000000000000000000000000..c08f4230721c4ae6fa710838db008c9802c61105 --- /dev/null +++ b/sleep_assessment/cases/case13/case13_14.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_04.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the Sleep Fragmentation Index (SFI). Define the sleep period as the interval from the first non-Wake epoch to the last non-Wake epoch. Within this sleep period, count the number of awakenings as the number of contiguous Wake (W) bouts, and count sleep-stage shifts as the number of adjacent epoch stage changes where both adjacent epochs are non-Wake sleep stages (N1, N2, N3, or R/REM). Calculate total sleep time as the number of non-Wake epochs multiplied by 30 seconds. SFI is defined as (number of awakenings + sleep-stage shifts) divided by total sleep time in hours, reported as events per hour of total sleep time. Please clearly report the final SFI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the Sleep Fragmentation Index (SFI) from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"sleep_fragmentation_index\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"sleep_fragmentation_index\": null}.\n\n### OUTPUT TEMPLATE\n{\"sleep_fragmentation_index\": }", + "metrics": [ + { + "metric_id": "sleep_fragmentation_index_accuracy", + "type": "numeric_check", + "target_key": "sleep_fragmentation_index", + "weight": 100, + "params": { + "gt_value": 22.524916943521596, + "tolerance": 1.12624584717608 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case13/case13_15.json b/sleep_assessment/cases/case13/case13_15.json new file mode 100644 index 0000000000000000000000000000000000000000..1537dcd668312fb6428b2a0d04a10ba35866b95c --- /dev/null +++ b/sleep_assessment/cases/case13/case13_15.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_05.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the Sleep Fragmentation Index (SFI). Define the sleep period as the interval from the first non-Wake epoch to the last non-Wake epoch. Within this sleep period, count the number of awakenings as the number of contiguous Wake (W) bouts, and count sleep-stage shifts as the number of adjacent epoch stage changes where both adjacent epochs are non-Wake sleep stages (N1, N2, N3, or R/REM). Calculate total sleep time as the number of non-Wake epochs multiplied by 30 seconds. SFI is defined as (number of awakenings + sleep-stage shifts) divided by total sleep time in hours, reported as events per hour of total sleep time. Please clearly report the final SFI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the Sleep Fragmentation Index (SFI) from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"sleep_fragmentation_index\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"sleep_fragmentation_index\": null}.\n\n### OUTPUT TEMPLATE\n{\"sleep_fragmentation_index\": }", + "metrics": [ + { + "metric_id": "sleep_fragmentation_index_accuracy", + "type": "numeric_check", + "target_key": "sleep_fragmentation_index", + "weight": 100, + "params": { + "gt_value": 15.567567567567567, + "tolerance": 0.7783783783783784 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case13/case13_16.json b/sleep_assessment/cases/case13/case13_16.json new file mode 100644 index 0000000000000000000000000000000000000000..9ea0e11d6d3c8cd59deee1b7ee4fec47b5405fde --- /dev/null +++ b/sleep_assessment/cases/case13/case13_16.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "MASSSS3_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_01.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the Sleep Fragmentation Index (SFI). Define the sleep period as the interval from the first non-Wake epoch to the last non-Wake epoch. Within this sleep period, count the number of awakenings as the number of contiguous Wake (W) bouts, and count sleep-stage shifts as the number of adjacent epoch stage changes where both adjacent epochs are non-Wake sleep stages (N1, N2, N3, or R/REM). Calculate total sleep time as the number of non-Wake epochs multiplied by 30 seconds. SFI is defined as (number of awakenings + sleep-stage shifts) divided by total sleep time in hours, reported as events per hour of total sleep time. Please clearly report the final SFI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the Sleep Fragmentation Index (SFI) from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"sleep_fragmentation_index\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"sleep_fragmentation_index\": null}.\n\n### OUTPUT TEMPLATE\n{\"sleep_fragmentation_index\": }", + "metrics": [ + { + "metric_id": "sleep_fragmentation_index_accuracy", + "type": "numeric_check", + "target_key": "sleep_fragmentation_index", + "weight": 100, + "params": { + "gt_value": 19.808306709265175, + "tolerance": 0.9904153354632588 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case13/case13_17.json b/sleep_assessment/cases/case13/case13_17.json new file mode 100644 index 0000000000000000000000000000000000000000..7175cb4678fcb3ff150b11ce599df3d803de0de0 --- /dev/null +++ b/sleep_assessment/cases/case13/case13_17.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "MASSSS3_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_02.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the Sleep Fragmentation Index (SFI). Define the sleep period as the interval from the first non-Wake epoch to the last non-Wake epoch. Within this sleep period, count the number of awakenings as the number of contiguous Wake (W) bouts, and count sleep-stage shifts as the number of adjacent epoch stage changes where both adjacent epochs are non-Wake sleep stages (N1, N2, N3, or R/REM). Calculate total sleep time as the number of non-Wake epochs multiplied by 30 seconds. SFI is defined as (number of awakenings + sleep-stage shifts) divided by total sleep time in hours, reported as events per hour of total sleep time. Please clearly report the final SFI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the Sleep Fragmentation Index (SFI) from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"sleep_fragmentation_index\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"sleep_fragmentation_index\": null}.\n\n### OUTPUT TEMPLATE\n{\"sleep_fragmentation_index\": }", + "metrics": [ + { + "metric_id": "sleep_fragmentation_index_accuracy", + "type": "numeric_check", + "target_key": "sleep_fragmentation_index", + "weight": 100, + "params": { + "gt_value": 20.48167539267016, + "tolerance": 1.024083769633508 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case13/case13_18.json b/sleep_assessment/cases/case13/case13_18.json new file mode 100644 index 0000000000000000000000000000000000000000..2fc1c6816335739b49be6f20886a20a7401903d3 --- /dev/null +++ b/sleep_assessment/cases/case13/case13_18.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "MASSSS3_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_03.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the Sleep Fragmentation Index (SFI). Define the sleep period as the interval from the first non-Wake epoch to the last non-Wake epoch. Within this sleep period, count the number of awakenings as the number of contiguous Wake (W) bouts, and count sleep-stage shifts as the number of adjacent epoch stage changes where both adjacent epochs are non-Wake sleep stages (N1, N2, N3, or R/REM). Calculate total sleep time as the number of non-Wake epochs multiplied by 30 seconds. SFI is defined as (number of awakenings + sleep-stage shifts) divided by total sleep time in hours, reported as events per hour of total sleep time. Please clearly report the final SFI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the Sleep Fragmentation Index (SFI) from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"sleep_fragmentation_index\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"sleep_fragmentation_index\": null}.\n\n### OUTPUT TEMPLATE\n{\"sleep_fragmentation_index\": }", + "metrics": [ + { + "metric_id": "sleep_fragmentation_index_accuracy", + "type": "numeric_check", + "target_key": "sleep_fragmentation_index", + "weight": 100, + "params": { + "gt_value": 16.478405315614616, + "tolerance": 0.8239202657807309 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case13/case13_19.json b/sleep_assessment/cases/case13/case13_19.json new file mode 100644 index 0000000000000000000000000000000000000000..c92b5cafaa03abbeff9dbbeb524feffc51344e50 --- /dev/null +++ b/sleep_assessment/cases/case13/case13_19.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "MASSSS3_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_04.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the Sleep Fragmentation Index (SFI). Define the sleep period as the interval from the first non-Wake epoch to the last non-Wake epoch. Within this sleep period, count the number of awakenings as the number of contiguous Wake (W) bouts, and count sleep-stage shifts as the number of adjacent epoch stage changes where both adjacent epochs are non-Wake sleep stages (N1, N2, N3, or R/REM). Calculate total sleep time as the number of non-Wake epochs multiplied by 30 seconds. SFI is defined as (number of awakenings + sleep-stage shifts) divided by total sleep time in hours, reported as events per hour of total sleep time. Please clearly report the final SFI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the Sleep Fragmentation Index (SFI) from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"sleep_fragmentation_index\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"sleep_fragmentation_index\": null}.\n\n### OUTPUT TEMPLATE\n{\"sleep_fragmentation_index\": }", + "metrics": [ + { + "metric_id": "sleep_fragmentation_index_accuracy", + "type": "numeric_check", + "target_key": "sleep_fragmentation_index", + "weight": 100, + "params": { + "gt_value": 12.25305216426193, + "tolerance": 0.6126526082130965 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case13/case13_20.json b/sleep_assessment/cases/case13/case13_20.json new file mode 100644 index 0000000000000000000000000000000000000000..be4cd64a8f2c154abac9baf80c0c2415ce8ee1e7 --- /dev/null +++ b/sleep_assessment/cases/case13/case13_20.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "MASSSS3_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_05.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the Sleep Fragmentation Index (SFI). Define the sleep period as the interval from the first non-Wake epoch to the last non-Wake epoch. Within this sleep period, count the number of awakenings as the number of contiguous Wake (W) bouts, and count sleep-stage shifts as the number of adjacent epoch stage changes where both adjacent epochs are non-Wake sleep stages (N1, N2, N3, or R/REM). Calculate total sleep time as the number of non-Wake epochs multiplied by 30 seconds. SFI is defined as (number of awakenings + sleep-stage shifts) divided by total sleep time in hours, reported as events per hour of total sleep time. Please clearly report the final SFI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the Sleep Fragmentation Index (SFI) from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"sleep_fragmentation_index\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"sleep_fragmentation_index\": null}.\n\n### OUTPUT TEMPLATE\n{\"sleep_fragmentation_index\": }", + "metrics": [ + { + "metric_id": "sleep_fragmentation_index_accuracy", + "type": "numeric_check", + "target_key": "sleep_fragmentation_index", + "weight": 100, + "params": { + "gt_value": 10.984251968503937, + "tolerance": 0.5492125984251969 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case13/case13_21.json b/sleep_assessment/cases/case13/case13_21.json new file mode 100644 index 0000000000000000000000000000000000000000..5a0646d2564a9a398f8b321377eb0e6eed018c9b --- /dev/null +++ b/sleep_assessment/cases/case13/case13_21.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_01.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the Sleep Fragmentation Index (SFI). Define the sleep period as the interval from the first non-Wake epoch to the last non-Wake epoch. Within this sleep period, count the number of awakenings as the number of contiguous Wake (W) bouts, and count sleep-stage shifts as the number of adjacent epoch stage changes where both adjacent epochs are non-Wake sleep stages (N1, N2, N3, or R/REM). Calculate total sleep time as the number of non-Wake epochs multiplied by 30 seconds. SFI is defined as (number of awakenings + sleep-stage shifts) divided by total sleep time in hours, reported as events per hour of total sleep time. Please clearly report the final SFI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the Sleep Fragmentation Index (SFI) from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"sleep_fragmentation_index\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"sleep_fragmentation_index\": null}.\n\n### OUTPUT TEMPLATE\n{\"sleep_fragmentation_index\": }", + "metrics": [ + { + "metric_id": "sleep_fragmentation_index_accuracy", + "type": "numeric_check", + "target_key": "sleep_fragmentation_index", + "weight": 100, + "params": { + "gt_value": 17.573221757322177, + "tolerance": 0.8786610878661089 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case13/case13_22.json b/sleep_assessment/cases/case13/case13_22.json new file mode 100644 index 0000000000000000000000000000000000000000..44cb8f384d03068ae295f24cfbf919e1105782bd --- /dev/null +++ b/sleep_assessment/cases/case13/case13_22.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_02.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the Sleep Fragmentation Index (SFI). Define the sleep period as the interval from the first non-Wake epoch to the last non-Wake epoch. Within this sleep period, count the number of awakenings as the number of contiguous Wake (W) bouts, and count sleep-stage shifts as the number of adjacent epoch stage changes where both adjacent epochs are non-Wake sleep stages (N1, N2, N3, or R/REM). Calculate total sleep time as the number of non-Wake epochs multiplied by 30 seconds. SFI is defined as (number of awakenings + sleep-stage shifts) divided by total sleep time in hours, reported as events per hour of total sleep time. Please clearly report the final SFI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the Sleep Fragmentation Index (SFI) from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"sleep_fragmentation_index\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"sleep_fragmentation_index\": null}.\n\n### OUTPUT TEMPLATE\n{\"sleep_fragmentation_index\": }", + "metrics": [ + { + "metric_id": "sleep_fragmentation_index_accuracy", + "type": "numeric_check", + "target_key": "sleep_fragmentation_index", + "weight": 100, + "params": { + "gt_value": 18.0, + "tolerance": 0.9 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case13/case13_23.json b/sleep_assessment/cases/case13/case13_23.json new file mode 100644 index 0000000000000000000000000000000000000000..60bc6b090b459ed50c159611b168f5ef722b17db --- /dev/null +++ b/sleep_assessment/cases/case13/case13_23.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "Physionet2018_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_03.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the Sleep Fragmentation Index (SFI). Define the sleep period as the interval from the first non-Wake epoch to the last non-Wake epoch. Within this sleep period, count the number of awakenings as the number of contiguous Wake (W) bouts, and count sleep-stage shifts as the number of adjacent epoch stage changes where both adjacent epochs are non-Wake sleep stages (N1, N2, N3, or R/REM). Calculate total sleep time as the number of non-Wake epochs multiplied by 30 seconds. SFI is defined as (number of awakenings + sleep-stage shifts) divided by total sleep time in hours, reported as events per hour of total sleep time. Please clearly report the final SFI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the Sleep Fragmentation Index (SFI) from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"sleep_fragmentation_index\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"sleep_fragmentation_index\": null}.\n\n### OUTPUT TEMPLATE\n{\"sleep_fragmentation_index\": }", + "metrics": [ + { + "metric_id": "sleep_fragmentation_index_accuracy", + "type": "numeric_check", + "target_key": "sleep_fragmentation_index", + "weight": 100, + "params": { + "gt_value": 18.849557522123895, + "tolerance": 0.9424778761061948 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case13/case13_24.json b/sleep_assessment/cases/case13/case13_24.json new file mode 100644 index 0000000000000000000000000000000000000000..1a76d957fc5f5093d151ebd45654a141d30495c7 --- /dev/null +++ b/sleep_assessment/cases/case13/case13_24.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_04.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the Sleep Fragmentation Index (SFI). Define the sleep period as the interval from the first non-Wake epoch to the last non-Wake epoch. Within this sleep period, count the number of awakenings as the number of contiguous Wake (W) bouts, and count sleep-stage shifts as the number of adjacent epoch stage changes where both adjacent epochs are non-Wake sleep stages (N1, N2, N3, or R/REM). Calculate total sleep time as the number of non-Wake epochs multiplied by 30 seconds. SFI is defined as (number of awakenings + sleep-stage shifts) divided by total sleep time in hours, reported as events per hour of total sleep time. Please clearly report the final SFI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the Sleep Fragmentation Index (SFI) from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"sleep_fragmentation_index\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"sleep_fragmentation_index\": null}.\n\n### OUTPUT TEMPLATE\n{\"sleep_fragmentation_index\": }", + "metrics": [ + { + "metric_id": "sleep_fragmentation_index_accuracy", + "type": "numeric_check", + "target_key": "sleep_fragmentation_index", + "weight": 100, + "params": { + "gt_value": 20.91743119266055, + "tolerance": 1.0458715596330277 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case13/case13_25.json b/sleep_assessment/cases/case13/case13_25.json new file mode 100644 index 0000000000000000000000000000000000000000..7111aefc72ddcb73a63f41e561a778374e9797d7 --- /dev/null +++ b/sleep_assessment/cases/case13/case13_25.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_05.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the Sleep Fragmentation Index (SFI). Define the sleep period as the interval from the first non-Wake epoch to the last non-Wake epoch. Within this sleep period, count the number of awakenings as the number of contiguous Wake (W) bouts, and count sleep-stage shifts as the number of adjacent epoch stage changes where both adjacent epochs are non-Wake sleep stages (N1, N2, N3, or R/REM). Calculate total sleep time as the number of non-Wake epochs multiplied by 30 seconds. SFI is defined as (number of awakenings + sleep-stage shifts) divided by total sleep time in hours, reported as events per hour of total sleep time. Please clearly report the final SFI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the Sleep Fragmentation Index (SFI) from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"sleep_fragmentation_index\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"sleep_fragmentation_index\": null}.\n\n### OUTPUT TEMPLATE\n{\"sleep_fragmentation_index\": }", + "metrics": [ + { + "metric_id": "sleep_fragmentation_index_accuracy", + "type": "numeric_check", + "target_key": "sleep_fragmentation_index", + "weight": 100, + "params": { + "gt_value": 10.4, + "tolerance": 0.52 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case14/case14_01.json b/sleep_assessment/cases/case14/case14_01.json new file mode 100644 index 0000000000000000000000000000000000000000..f44328a0873dc7b8bd1cfbbdd60811c758af715b --- /dev/null +++ b/sleep_assessment/cases/case14/case14_01.json @@ -0,0 +1,67 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the percentage distribution of the five sleep stages W, N1, N2, N3, and R. Report the percentage of the total analyzed sleep-stage sequence occupied by each stage. Please clearly report all five final percentage values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract five sleep-stage percentage values from the agent's report:\n1) W stage percentage\n2) N1 stage percentage\n3) N2 stage percentage\n4) N3 stage percentage\n5) R/REM stage percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"w_stage_ratio_percent\", \"n1_stage_ratio_percent\", \"n2_stage_ratio_percent\", \"n3_stage_ratio_percent\", and \"r_stage_ratio_percent\".\n5. All values must be FLOATS or null.\n6. Values are percentages; if the report says 25.5%, output 25.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"w_stage_ratio_percent\": , \"n1_stage_ratio_percent\": , \"n2_stage_ratio_percent\": , \"n3_stage_ratio_percent\": , \"r_stage_ratio_percent\": }", + "metrics": [ + { + "metric_id": "w_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "w_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 30.0, + "tolerance": 1.5 + } + }, + { + "metric_id": "n1_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n1_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 8.295454545454545, + "tolerance": 0.4147727272727273 + } + }, + { + "metric_id": "n2_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n2_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 22.045454545454547, + "tolerance": 1.1022727272727273 + } + }, + { + "metric_id": "n3_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n3_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 26.25, + "tolerance": 1.3125 + } + }, + { + "metric_id": "r_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "r_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 13.409090909090908, + "tolerance": 0.6704545454545454 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case14/case14_02.json b/sleep_assessment/cases/case14/case14_02.json new file mode 100644 index 0000000000000000000000000000000000000000..373c83459029f60748a8fc89377b374cd4572b64 --- /dev/null +++ b/sleep_assessment/cases/case14/case14_02.json @@ -0,0 +1,67 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_02.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the percentage distribution of the five sleep stages W, N1, N2, N3, and R. Report the percentage of the total analyzed sleep-stage sequence occupied by each stage. Please clearly report all five final percentage values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract five sleep-stage percentage values from the agent's report:\n1) W stage percentage\n2) N1 stage percentage\n3) N2 stage percentage\n4) N3 stage percentage\n5) R/REM stage percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"w_stage_ratio_percent\", \"n1_stage_ratio_percent\", \"n2_stage_ratio_percent\", \"n3_stage_ratio_percent\", and \"r_stage_ratio_percent\".\n5. All values must be FLOATS or null.\n6. Values are percentages; if the report says 25.5%, output 25.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"w_stage_ratio_percent\": , \"n1_stage_ratio_percent\": , \"n2_stage_ratio_percent\": , \"n3_stage_ratio_percent\": , \"r_stage_ratio_percent\": }", + "metrics": [ + { + "metric_id": "w_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "w_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 25.414937759336098, + "tolerance": 1.270746887966805 + } + }, + { + "metric_id": "n1_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n1_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 11.929460580912863, + "tolerance": 0.5964730290456431 + } + }, + { + "metric_id": "n2_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n2_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 35.788381742738586, + "tolerance": 1.7894190871369293 + } + }, + { + "metric_id": "n3_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n3_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 16.286307053941908, + "tolerance": 0.8143153526970954 + } + }, + { + "metric_id": "r_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "r_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 10.58091286307054, + "tolerance": 0.529045643153527 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case14/case14_03.json b/sleep_assessment/cases/case14/case14_03.json new file mode 100644 index 0000000000000000000000000000000000000000..ab79d9c29ba4a0c45ec1701c1030381752c14225 --- /dev/null +++ b/sleep_assessment/cases/case14/case14_03.json @@ -0,0 +1,67 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_03.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the percentage distribution of the five sleep stages W, N1, N2, N3, and R. Report the percentage of the total analyzed sleep-stage sequence occupied by each stage. Please clearly report all five final percentage values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract five sleep-stage percentage values from the agent's report:\n1) W stage percentage\n2) N1 stage percentage\n3) N2 stage percentage\n4) N3 stage percentage\n5) R/REM stage percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"w_stage_ratio_percent\", \"n1_stage_ratio_percent\", \"n2_stage_ratio_percent\", \"n3_stage_ratio_percent\", and \"r_stage_ratio_percent\".\n5. All values must be FLOATS or null.\n6. Values are percentages; if the report says 25.5%, output 25.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"w_stage_ratio_percent\": , \"n1_stage_ratio_percent\": , \"n2_stage_ratio_percent\": , \"n3_stage_ratio_percent\": , \"r_stage_ratio_percent\": }", + "metrics": [ + { + "metric_id": "w_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "w_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 13.997879109225876, + "tolerance": 0.6998939554612938 + } + }, + { + "metric_id": "n1_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n1_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 17.497348886532343, + "tolerance": 0.8748674443266172 + } + }, + { + "metric_id": "n2_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n2_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 26.08695652173913, + "tolerance": 1.3043478260869565 + } + }, + { + "metric_id": "n3_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n3_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 18.345705196182397, + "tolerance": 0.9172852598091199 + } + }, + { + "metric_id": "r_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "r_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 24.072110286320253, + "tolerance": 1.2036055143160127 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case14/case14_04.json b/sleep_assessment/cases/case14/case14_04.json new file mode 100644 index 0000000000000000000000000000000000000000..734082a0812ac9d77a78b0aad7f00a475204cdd9 --- /dev/null +++ b/sleep_assessment/cases/case14/case14_04.json @@ -0,0 +1,67 @@ +{ + "meta_info": { + "case_id": "ISRUC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_04.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the percentage distribution of the five sleep stages W, N1, N2, N3, and R. Report the percentage of the total analyzed sleep-stage sequence occupied by each stage. Please clearly report all five final percentage values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract five sleep-stage percentage values from the agent's report:\n1) W stage percentage\n2) N1 stage percentage\n3) N2 stage percentage\n4) N3 stage percentage\n5) R/REM stage percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"w_stage_ratio_percent\", \"n1_stage_ratio_percent\", \"n2_stage_ratio_percent\", \"n3_stage_ratio_percent\", and \"r_stage_ratio_percent\".\n5. All values must be FLOATS or null.\n6. Values are percentages; if the report says 25.5%, output 25.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"w_stage_ratio_percent\": , \"n1_stage_ratio_percent\": , \"n2_stage_ratio_percent\": , \"n3_stage_ratio_percent\": , \"r_stage_ratio_percent\": }", + "metrics": [ + { + "metric_id": "w_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "w_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 2.907580477673936, + "tolerance": 0.1453790238836968 + } + }, + { + "metric_id": "n1_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n1_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 6.749740394600208, + "tolerance": 0.3374870197300104 + } + }, + { + "metric_id": "n2_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n2_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 44.23676012461059, + "tolerance": 2.2118380062305296 + } + }, + { + "metric_id": "n3_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n3_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 22.22222222222222, + "tolerance": 1.1111111111111112 + } + }, + { + "metric_id": "r_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "r_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 23.883696780893043, + "tolerance": 1.1941848390446521 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case14/case14_05.json b/sleep_assessment/cases/case14/case14_05.json new file mode 100644 index 0000000000000000000000000000000000000000..e98487e90ba9cb78a09275ebe53a48fc9c97570d --- /dev/null +++ b/sleep_assessment/cases/case14/case14_05.json @@ -0,0 +1,67 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_05.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the percentage distribution of the five sleep stages W, N1, N2, N3, and R. Report the percentage of the total analyzed sleep-stage sequence occupied by each stage. Please clearly report all five final percentage values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract five sleep-stage percentage values from the agent's report:\n1) W stage percentage\n2) N1 stage percentage\n3) N2 stage percentage\n4) N3 stage percentage\n5) R/REM stage percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"w_stage_ratio_percent\", \"n1_stage_ratio_percent\", \"n2_stage_ratio_percent\", \"n3_stage_ratio_percent\", and \"r_stage_ratio_percent\".\n5. All values must be FLOATS or null.\n6. Values are percentages; if the report says 25.5%, output 25.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"w_stage_ratio_percent\": , \"n1_stage_ratio_percent\": , \"n2_stage_ratio_percent\": , \"n3_stage_ratio_percent\": , \"r_stage_ratio_percent\": }", + "metrics": [ + { + "metric_id": "w_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "w_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 33.82857142857143, + "tolerance": 1.6914285714285715 + } + }, + { + "metric_id": "n1_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n1_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 12.342857142857143, + "tolerance": 0.6171428571428572 + } + }, + { + "metric_id": "n2_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n2_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 30.285714285714285, + "tolerance": 1.5142857142857142 + } + }, + { + "metric_id": "n3_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n3_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 18.742857142857144, + "tolerance": 0.9371428571428573 + } + }, + { + "metric_id": "r_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "r_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 4.8, + "tolerance": 0.24 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case14/case14_06.json b/sleep_assessment/cases/case14/case14_06.json new file mode 100644 index 0000000000000000000000000000000000000000..b4ea47329b79a82f2fb71ab96b5a6d0a3ce8a219 --- /dev/null +++ b/sleep_assessment/cases/case14/case14_06.json @@ -0,0 +1,67 @@ +{ + "meta_info": { + "case_id": "HMC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_01.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the percentage distribution of the five sleep stages W, N1, N2, N3, and R. Report the percentage of the total analyzed sleep-stage sequence occupied by each stage. Please clearly report all five final percentage values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract five sleep-stage percentage values from the agent's report:\n1) W stage percentage\n2) N1 stage percentage\n3) N2 stage percentage\n4) N3 stage percentage\n5) R/REM stage percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"w_stage_ratio_percent\", \"n1_stage_ratio_percent\", \"n2_stage_ratio_percent\", \"n3_stage_ratio_percent\", and \"r_stage_ratio_percent\".\n5. All values must be FLOATS or null.\n6. Values are percentages; if the report says 25.5%, output 25.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"w_stage_ratio_percent\": , \"n1_stage_ratio_percent\": , \"n2_stage_ratio_percent\": , \"n3_stage_ratio_percent\": , \"r_stage_ratio_percent\": }", + "metrics": [ + { + "metric_id": "w_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "w_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 17.681498829039814, + "tolerance": 0.8840749414519907 + } + }, + { + "metric_id": "n1_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n1_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 12.763466042154567, + "tolerance": 0.6381733021077284 + } + }, + { + "metric_id": "n2_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n2_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 50.35128805620609, + "tolerance": 2.5175644028103044 + } + }, + { + "metric_id": "n3_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n3_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 2.693208430913349, + "tolerance": 0.13466042154566746 + } + }, + { + "metric_id": "r_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "r_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 16.510538641686182, + "tolerance": 0.8255269320843092 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case14/case14_07.json b/sleep_assessment/cases/case14/case14_07.json new file mode 100644 index 0000000000000000000000000000000000000000..9c4520360ec3948e0b08ba048b39bbb3155554ff --- /dev/null +++ b/sleep_assessment/cases/case14/case14_07.json @@ -0,0 +1,67 @@ +{ + "meta_info": { + "case_id": "HMC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_02.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the percentage distribution of the five sleep stages W, N1, N2, N3, and R. Report the percentage of the total analyzed sleep-stage sequence occupied by each stage. Please clearly report all five final percentage values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract five sleep-stage percentage values from the agent's report:\n1) W stage percentage\n2) N1 stage percentage\n3) N2 stage percentage\n4) N3 stage percentage\n5) R/REM stage percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"w_stage_ratio_percent\", \"n1_stage_ratio_percent\", \"n2_stage_ratio_percent\", \"n3_stage_ratio_percent\", and \"r_stage_ratio_percent\".\n5. All values must be FLOATS or null.\n6. Values are percentages; if the report says 25.5%, output 25.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"w_stage_ratio_percent\": , \"n1_stage_ratio_percent\": , \"n2_stage_ratio_percent\": , \"n3_stage_ratio_percent\": , \"r_stage_ratio_percent\": }", + "metrics": [ + { + "metric_id": "w_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "w_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 9.57943925233645, + "tolerance": 0.4789719626168225 + } + }, + { + "metric_id": "n1_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n1_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 7.827102803738318, + "tolerance": 0.3913551401869159 + } + }, + { + "metric_id": "n2_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n2_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 36.098130841121495, + "tolerance": 1.8049065420560748 + } + }, + { + "metric_id": "n3_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n3_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 26.16822429906542, + "tolerance": 1.3084112149532712 + } + }, + { + "metric_id": "r_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "r_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 20.327102803738317, + "tolerance": 1.016355140186916 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case14/case14_08.json b/sleep_assessment/cases/case14/case14_08.json new file mode 100644 index 0000000000000000000000000000000000000000..9e34014d8182cde7ad4d573633b4c3c17fdf7d4a --- /dev/null +++ b/sleep_assessment/cases/case14/case14_08.json @@ -0,0 +1,67 @@ +{ + "meta_info": { + "case_id": "HMC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_03.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the percentage distribution of the five sleep stages W, N1, N2, N3, and R. Report the percentage of the total analyzed sleep-stage sequence occupied by each stage. Please clearly report all five final percentage values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract five sleep-stage percentage values from the agent's report:\n1) W stage percentage\n2) N1 stage percentage\n3) N2 stage percentage\n4) N3 stage percentage\n5) R/REM stage percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"w_stage_ratio_percent\", \"n1_stage_ratio_percent\", \"n2_stage_ratio_percent\", \"n3_stage_ratio_percent\", and \"r_stage_ratio_percent\".\n5. All values must be FLOATS or null.\n6. Values are percentages; if the report says 25.5%, output 25.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"w_stage_ratio_percent\": , \"n1_stage_ratio_percent\": , \"n2_stage_ratio_percent\": , \"n3_stage_ratio_percent\": , \"r_stage_ratio_percent\": }", + "metrics": [ + { + "metric_id": "w_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "w_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 42.24318658280922, + "tolerance": 2.1121593291404612 + } + }, + { + "metric_id": "n1_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n1_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 8.49056603773585, + "tolerance": 0.42452830188679247 + } + }, + { + "metric_id": "n2_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n2_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 17.61006289308176, + "tolerance": 0.8805031446540881 + } + }, + { + "metric_id": "n3_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n3_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 15.723270440251572, + "tolerance": 0.7861635220125787 + } + }, + { + "metric_id": "r_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "r_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 15.932914046121594, + "tolerance": 0.7966457023060798 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case14/case14_09.json b/sleep_assessment/cases/case14/case14_09.json new file mode 100644 index 0000000000000000000000000000000000000000..f27127f7c272d56d734e6d2eb116adb0eeccfe4d --- /dev/null +++ b/sleep_assessment/cases/case14/case14_09.json @@ -0,0 +1,67 @@ +{ + "meta_info": { + "case_id": "HMC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_04.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the percentage distribution of the five sleep stages W, N1, N2, N3, and R. Report the percentage of the total analyzed sleep-stage sequence occupied by each stage. Please clearly report all five final percentage values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract five sleep-stage percentage values from the agent's report:\n1) W stage percentage\n2) N1 stage percentage\n3) N2 stage percentage\n4) N3 stage percentage\n5) R/REM stage percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"w_stage_ratio_percent\", \"n1_stage_ratio_percent\", \"n2_stage_ratio_percent\", \"n3_stage_ratio_percent\", and \"r_stage_ratio_percent\".\n5. All values must be FLOATS or null.\n6. Values are percentages; if the report says 25.5%, output 25.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"w_stage_ratio_percent\": , \"n1_stage_ratio_percent\": , \"n2_stage_ratio_percent\": , \"n3_stage_ratio_percent\": , \"r_stage_ratio_percent\": }", + "metrics": [ + { + "metric_id": "w_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "w_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 7.874015748031496, + "tolerance": 0.3937007874015748 + } + }, + { + "metric_id": "n1_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n1_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 8.956692913385826, + "tolerance": 0.4478346456692913 + } + }, + { + "metric_id": "n2_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n2_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 52.55905511811024, + "tolerance": 2.627952755905512 + } + }, + { + "metric_id": "n3_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n3_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 6.299212598425197, + "tolerance": 0.3149606299212599 + } + }, + { + "metric_id": "r_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "r_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 24.311023622047244, + "tolerance": 1.2155511811023623 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case14/case14_10.json b/sleep_assessment/cases/case14/case14_10.json new file mode 100644 index 0000000000000000000000000000000000000000..756b1038d8ec4147c217be0a4857bd9150dba6f0 --- /dev/null +++ b/sleep_assessment/cases/case14/case14_10.json @@ -0,0 +1,67 @@ +{ + "meta_info": { + "case_id": "HMC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_05.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the percentage distribution of the five sleep stages W, N1, N2, N3, and R. Report the percentage of the total analyzed sleep-stage sequence occupied by each stage. Please clearly report all five final percentage values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract five sleep-stage percentage values from the agent's report:\n1) W stage percentage\n2) N1 stage percentage\n3) N2 stage percentage\n4) N3 stage percentage\n5) R/REM stage percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"w_stage_ratio_percent\", \"n1_stage_ratio_percent\", \"n2_stage_ratio_percent\", \"n3_stage_ratio_percent\", and \"r_stage_ratio_percent\".\n5. All values must be FLOATS or null.\n6. Values are percentages; if the report says 25.5%, output 25.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"w_stage_ratio_percent\": , \"n1_stage_ratio_percent\": , \"n2_stage_ratio_percent\": , \"n3_stage_ratio_percent\": , \"r_stage_ratio_percent\": }", + "metrics": [ + { + "metric_id": "w_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "w_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 20.959332638164756, + "tolerance": 1.047966631908238 + } + }, + { + "metric_id": "n1_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n1_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 14.70281543274244, + "tolerance": 0.7351407716371221 + } + }, + { + "metric_id": "n2_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n2_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 35.557872784150156, + "tolerance": 1.777893639207508 + } + }, + { + "metric_id": "n3_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n3_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 15.537017726798748, + "tolerance": 0.7768508863399375 + } + }, + { + "metric_id": "r_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "r_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 13.2429614181439, + "tolerance": 0.662148070907195 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case14/case14_11.json b/sleep_assessment/cases/case14/case14_11.json new file mode 100644 index 0000000000000000000000000000000000000000..77d4894d6bd6cb9d83b6b712e6464217e0257600 --- /dev/null +++ b/sleep_assessment/cases/case14/case14_11.json @@ -0,0 +1,67 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_01.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the percentage distribution of the five sleep stages W, N1, N2, N3, and R. Report the percentage of the total analyzed sleep-stage sequence occupied by each stage. Please clearly report all five final percentage values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract five sleep-stage percentage values from the agent's report:\n1) W stage percentage\n2) N1 stage percentage\n3) N2 stage percentage\n4) N3 stage percentage\n5) R/REM stage percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"w_stage_ratio_percent\", \"n1_stage_ratio_percent\", \"n2_stage_ratio_percent\", \"n3_stage_ratio_percent\", and \"r_stage_ratio_percent\".\n5. All values must be FLOATS or null.\n6. Values are percentages; if the report says 25.5%, output 25.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"w_stage_ratio_percent\": , \"n1_stage_ratio_percent\": , \"n2_stage_ratio_percent\": , \"n3_stage_ratio_percent\": , \"r_stage_ratio_percent\": }", + "metrics": [ + { + "metric_id": "w_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "w_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 30.719557195571955, + "tolerance": 1.5359778597785978 + } + }, + { + "metric_id": "n1_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n1_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 4.3357933579335795, + "tolerance": 0.21678966789667897 + } + }, + { + "metric_id": "n2_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n2_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 42.15867158671587, + "tolerance": 2.1079335793357936 + } + }, + { + "metric_id": "n3_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n3_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 13.376383763837639, + "tolerance": 0.668819188191882 + } + }, + { + "metric_id": "r_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "r_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 9.409594095940959, + "tolerance": 0.4704797047970479 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case14/case14_12.json b/sleep_assessment/cases/case14/case14_12.json new file mode 100644 index 0000000000000000000000000000000000000000..4eac2169761dd0ec25f73861ca76d035c53f9c26 --- /dev/null +++ b/sleep_assessment/cases/case14/case14_12.json @@ -0,0 +1,67 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_02.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the percentage distribution of the five sleep stages W, N1, N2, N3, and R. Report the percentage of the total analyzed sleep-stage sequence occupied by each stage. Please clearly report all five final percentage values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract five sleep-stage percentage values from the agent's report:\n1) W stage percentage\n2) N1 stage percentage\n3) N2 stage percentage\n4) N3 stage percentage\n5) R/REM stage percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"w_stage_ratio_percent\", \"n1_stage_ratio_percent\", \"n2_stage_ratio_percent\", \"n3_stage_ratio_percent\", and \"r_stage_ratio_percent\".\n5. All values must be FLOATS or null.\n6. Values are percentages; if the report says 25.5%, output 25.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"w_stage_ratio_percent\": , \"n1_stage_ratio_percent\": , \"n2_stage_ratio_percent\": , \"n3_stage_ratio_percent\": , \"r_stage_ratio_percent\": }", + "metrics": [ + { + "metric_id": "w_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "w_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 66.26506024096386, + "tolerance": 3.313253012048193 + } + }, + { + "metric_id": "n1_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n1_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 0.27803521779425394, + "tolerance": 0.013901760889712698 + } + }, + { + "metric_id": "n2_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n2_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 22.150139017608897, + "tolerance": 1.107506950880445 + } + }, + { + "metric_id": "n3_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n3_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 5.653382761816497, + "tolerance": 0.28266913809082483 + } + }, + { + "metric_id": "r_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "r_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 5.653382761816497, + "tolerance": 0.28266913809082483 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case14/case14_13.json b/sleep_assessment/cases/case14/case14_13.json new file mode 100644 index 0000000000000000000000000000000000000000..52da55d2de66cada8fc7ffb21888dd1bfa766ba6 --- /dev/null +++ b/sleep_assessment/cases/case14/case14_13.json @@ -0,0 +1,67 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_03.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the percentage distribution of the five sleep stages W, N1, N2, N3, and R. Report the percentage of the total analyzed sleep-stage sequence occupied by each stage. Please clearly report all five final percentage values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract five sleep-stage percentage values from the agent's report:\n1) W stage percentage\n2) N1 stage percentage\n3) N2 stage percentage\n4) N3 stage percentage\n5) R/REM stage percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"w_stage_ratio_percent\", \"n1_stage_ratio_percent\", \"n2_stage_ratio_percent\", \"n3_stage_ratio_percent\", and \"r_stage_ratio_percent\".\n5. All values must be FLOATS or null.\n6. Values are percentages; if the report says 25.5%, output 25.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"w_stage_ratio_percent\": , \"n1_stage_ratio_percent\": , \"n2_stage_ratio_percent\": , \"n3_stage_ratio_percent\": , \"r_stage_ratio_percent\": }", + "metrics": [ + { + "metric_id": "w_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "w_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 31.649189704480456, + "tolerance": 1.5824594852240228 + } + }, + { + "metric_id": "n1_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n1_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 3.336510962821735, + "tolerance": 0.16682554814108674 + } + }, + { + "metric_id": "n2_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n2_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 27.550047664442324, + "tolerance": 1.3775023832221163 + } + }, + { + "metric_id": "n3_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n3_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 29.26596758817922, + "tolerance": 1.4632983794089611 + } + }, + { + "metric_id": "r_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "r_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 8.198284080076263, + "tolerance": 0.4099142040038132 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case14/case14_14.json b/sleep_assessment/cases/case14/case14_14.json new file mode 100644 index 0000000000000000000000000000000000000000..238cbe0eac2a0086b9de817b017e8be120aa438b --- /dev/null +++ b/sleep_assessment/cases/case14/case14_14.json @@ -0,0 +1,67 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_04.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the percentage distribution of the five sleep stages W, N1, N2, N3, and R. Report the percentage of the total analyzed sleep-stage sequence occupied by each stage. Please clearly report all five final percentage values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract five sleep-stage percentage values from the agent's report:\n1) W stage percentage\n2) N1 stage percentage\n3) N2 stage percentage\n4) N3 stage percentage\n5) R/REM stage percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"w_stage_ratio_percent\", \"n1_stage_ratio_percent\", \"n2_stage_ratio_percent\", \"n3_stage_ratio_percent\", and \"r_stage_ratio_percent\".\n5. All values must be FLOATS or null.\n6. Values are percentages; if the report says 25.5%, output 25.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"w_stage_ratio_percent\": , \"n1_stage_ratio_percent\": , \"n2_stage_ratio_percent\": , \"n3_stage_ratio_percent\": , \"r_stage_ratio_percent\": }", + "metrics": [ + { + "metric_id": "w_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "w_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 31.2, + "tolerance": 1.56 + } + }, + { + "metric_id": "n1_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n1_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 2.057142857142857, + "tolerance": 0.10285714285714286 + } + }, + { + "metric_id": "n2_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n2_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 20.228571428571428, + "tolerance": 1.0114285714285713 + } + }, + { + "metric_id": "n3_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n3_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 36.0, + "tolerance": 1.8 + } + }, + { + "metric_id": "r_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "r_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 10.514285714285714, + "tolerance": 0.5257142857142857 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case14/case14_15.json b/sleep_assessment/cases/case14/case14_15.json new file mode 100644 index 0000000000000000000000000000000000000000..49b096095bf940e113b98ed220610660d3319794 --- /dev/null +++ b/sleep_assessment/cases/case14/case14_15.json @@ -0,0 +1,67 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_05.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the percentage distribution of the five sleep stages W, N1, N2, N3, and R. Report the percentage of the total analyzed sleep-stage sequence occupied by each stage. Please clearly report all five final percentage values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract five sleep-stage percentage values from the agent's report:\n1) W stage percentage\n2) N1 stage percentage\n3) N2 stage percentage\n4) N3 stage percentage\n5) R/REM stage percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"w_stage_ratio_percent\", \"n1_stage_ratio_percent\", \"n2_stage_ratio_percent\", \"n3_stage_ratio_percent\", and \"r_stage_ratio_percent\".\n5. All values must be FLOATS or null.\n6. Values are percentages; if the report says 25.5%, output 25.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"w_stage_ratio_percent\": , \"n1_stage_ratio_percent\": , \"n2_stage_ratio_percent\": , \"n3_stage_ratio_percent\": , \"r_stage_ratio_percent\": }", + "metrics": [ + { + "metric_id": "w_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "w_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 31.73431734317343, + "tolerance": 1.5867158671586716 + } + }, + { + "metric_id": "n1_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n1_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 3.874538745387454, + "tolerance": 0.1937269372693727 + } + }, + { + "metric_id": "n2_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n2_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 46.86346863468635, + "tolerance": 2.3431734317343174 + } + }, + { + "metric_id": "n3_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n3_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 9.132841328413285, + "tolerance": 0.4566420664206643 + } + }, + { + "metric_id": "r_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "r_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 8.394833948339484, + "tolerance": 0.4197416974169742 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case14/case14_16.json b/sleep_assessment/cases/case14/case14_16.json new file mode 100644 index 0000000000000000000000000000000000000000..d90bbf5e1898c28e74f402b8f7b19417f7f47f4f --- /dev/null +++ b/sleep_assessment/cases/case14/case14_16.json @@ -0,0 +1,67 @@ +{ + "meta_info": { + "case_id": "MASSSS3_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_01.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the percentage distribution of the five sleep stages W, N1, N2, N3, and R. Report the percentage of the total analyzed sleep-stage sequence occupied by each stage. Please clearly report all five final percentage values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract five sleep-stage percentage values from the agent's report:\n1) W stage percentage\n2) N1 stage percentage\n3) N2 stage percentage\n4) N3 stage percentage\n5) R/REM stage percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"w_stage_ratio_percent\", \"n1_stage_ratio_percent\", \"n2_stage_ratio_percent\", \"n3_stage_ratio_percent\", and \"r_stage_ratio_percent\".\n5. All values must be FLOATS or null.\n6. Values are percentages; if the report says 25.5%, output 25.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"w_stage_ratio_percent\": , \"n1_stage_ratio_percent\": , \"n2_stage_ratio_percent\": , \"n3_stage_ratio_percent\": , \"r_stage_ratio_percent\": }", + "metrics": [ + { + "metric_id": "w_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "w_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 6.845238095238095, + "tolerance": 0.34226190476190477 + } + }, + { + "metric_id": "n1_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n1_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 7.341269841269841, + "tolerance": 0.3670634920634921 + } + }, + { + "metric_id": "n2_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n2_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 53.37301587301587, + "tolerance": 2.668650793650794 + } + }, + { + "metric_id": "n3_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n3_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 14.484126984126984, + "tolerance": 0.7242063492063493 + } + }, + { + "metric_id": "r_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "r_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 17.956349206349206, + "tolerance": 0.8978174603174603 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case14/case14_17.json b/sleep_assessment/cases/case14/case14_17.json new file mode 100644 index 0000000000000000000000000000000000000000..00ac722e914bc86bb85f5a9f7c15a94419e85a40 --- /dev/null +++ b/sleep_assessment/cases/case14/case14_17.json @@ -0,0 +1,67 @@ +{ + "meta_info": { + "case_id": "MASSSS3_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_02.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the percentage distribution of the five sleep stages W, N1, N2, N3, and R. Report the percentage of the total analyzed sleep-stage sequence occupied by each stage. Please clearly report all five final percentage values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract five sleep-stage percentage values from the agent's report:\n1) W stage percentage\n2) N1 stage percentage\n3) N2 stage percentage\n4) N3 stage percentage\n5) R/REM stage percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"w_stage_ratio_percent\", \"n1_stage_ratio_percent\", \"n2_stage_ratio_percent\", \"n3_stage_ratio_percent\", and \"r_stage_ratio_percent\".\n5. All values must be FLOATS or null.\n6. Values are percentages; if the report says 25.5%, output 25.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"w_stage_ratio_percent\": , \"n1_stage_ratio_percent\": , \"n2_stage_ratio_percent\": , \"n3_stage_ratio_percent\": , \"r_stage_ratio_percent\": }", + "metrics": [ + { + "metric_id": "w_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "w_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 10.914179104477611, + "tolerance": 0.5457089552238806 + } + }, + { + "metric_id": "n1_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n1_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 5.7835820895522385, + "tolerance": 0.2891791044776119 + } + }, + { + "metric_id": "n2_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n2_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 44.49626865671642, + "tolerance": 2.224813432835821 + } + }, + { + "metric_id": "n3_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n3_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 16.23134328358209, + "tolerance": 0.8115671641791045 + } + }, + { + "metric_id": "r_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "r_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 22.574626865671643, + "tolerance": 1.1287313432835822 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case14/case14_18.json b/sleep_assessment/cases/case14/case14_18.json new file mode 100644 index 0000000000000000000000000000000000000000..487c2b431efd7eb61edb56a0e231fc3230f0ec17 --- /dev/null +++ b/sleep_assessment/cases/case14/case14_18.json @@ -0,0 +1,67 @@ +{ + "meta_info": { + "case_id": "MASSSS3_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_03.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the percentage distribution of the five sleep stages W, N1, N2, N3, and R. Report the percentage of the total analyzed sleep-stage sequence occupied by each stage. Please clearly report all five final percentage values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract five sleep-stage percentage values from the agent's report:\n1) W stage percentage\n2) N1 stage percentage\n3) N2 stage percentage\n4) N3 stage percentage\n5) R/REM stage percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"w_stage_ratio_percent\", \"n1_stage_ratio_percent\", \"n2_stage_ratio_percent\", \"n3_stage_ratio_percent\", and \"r_stage_ratio_percent\".\n5. All values must be FLOATS or null.\n6. Values are percentages; if the report says 25.5%, output 25.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"w_stage_ratio_percent\": , \"n1_stage_ratio_percent\": , \"n2_stage_ratio_percent\": , \"n3_stage_ratio_percent\": , \"r_stage_ratio_percent\": }", + "metrics": [ + { + "metric_id": "w_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "w_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 5.445026178010472, + "tolerance": 0.27225130890052357 + } + }, + { + "metric_id": "n1_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n1_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 5.968586387434555, + "tolerance": 0.29842931937172773 + } + }, + { + "metric_id": "n2_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n2_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 47.539267015706805, + "tolerance": 2.3769633507853403 + } + }, + { + "metric_id": "n3_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n3_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 12.984293193717278, + "tolerance": 0.6492146596858639 + } + }, + { + "metric_id": "r_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "r_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 28.06282722513089, + "tolerance": 1.4031413612565444 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case14/case14_19.json b/sleep_assessment/cases/case14/case14_19.json new file mode 100644 index 0000000000000000000000000000000000000000..36fa3c76cd3335b8b6ca6a4f9f8159427fa0525b --- /dev/null +++ b/sleep_assessment/cases/case14/case14_19.json @@ -0,0 +1,67 @@ +{ + "meta_info": { + "case_id": "MASSSS3_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_04.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the percentage distribution of the five sleep stages W, N1, N2, N3, and R. Report the percentage of the total analyzed sleep-stage sequence occupied by each stage. Please clearly report all five final percentage values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract five sleep-stage percentage values from the agent's report:\n1) W stage percentage\n2) N1 stage percentage\n3) N2 stage percentage\n4) N3 stage percentage\n5) R/REM stage percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"w_stage_ratio_percent\", \"n1_stage_ratio_percent\", \"n2_stage_ratio_percent\", \"n3_stage_ratio_percent\", and \"r_stage_ratio_percent\".\n5. All values must be FLOATS or null.\n6. Values are percentages; if the report says 25.5%, output 25.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"w_stage_ratio_percent\": , \"n1_stage_ratio_percent\": , \"n2_stage_ratio_percent\": , \"n3_stage_ratio_percent\": , \"r_stage_ratio_percent\": }", + "metrics": [ + { + "metric_id": "w_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "w_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 1.206140350877193, + "tolerance": 0.060307017543859656 + } + }, + { + "metric_id": "n1_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n1_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 5.2631578947368425, + "tolerance": 0.26315789473684215 + } + }, + { + "metric_id": "n2_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n2_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 59.64912280701754, + "tolerance": 2.982456140350877 + } + }, + { + "metric_id": "n3_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n3_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 7.017543859649122, + "tolerance": 0.3508771929824561 + } + }, + { + "metric_id": "r_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "r_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 26.864035087719298, + "tolerance": 1.343201754385965 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case14/case14_20.json b/sleep_assessment/cases/case14/case14_20.json new file mode 100644 index 0000000000000000000000000000000000000000..fa95e14ae55fb8a73301b891c8cf42b0f29b4b44 --- /dev/null +++ b/sleep_assessment/cases/case14/case14_20.json @@ -0,0 +1,67 @@ +{ + "meta_info": { + "case_id": "MASSSS3_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_05.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the percentage distribution of the five sleep stages W, N1, N2, N3, and R. Report the percentage of the total analyzed sleep-stage sequence occupied by each stage. Please clearly report all five final percentage values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract five sleep-stage percentage values from the agent's report:\n1) W stage percentage\n2) N1 stage percentage\n3) N2 stage percentage\n4) N3 stage percentage\n5) R/REM stage percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"w_stage_ratio_percent\", \"n1_stage_ratio_percent\", \"n2_stage_ratio_percent\", \"n3_stage_ratio_percent\", and \"r_stage_ratio_percent\".\n5. All values must be FLOATS or null.\n6. Values are percentages; if the report says 25.5%, output 25.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"w_stage_ratio_percent\": , \"n1_stage_ratio_percent\": , \"n2_stage_ratio_percent\": , \"n3_stage_ratio_percent\": , \"r_stage_ratio_percent\": }", + "metrics": [ + { + "metric_id": "w_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "w_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 4.9579045837231055, + "tolerance": 0.24789522918615528 + } + }, + { + "metric_id": "n1_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n1_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 7.015902712815715, + "tolerance": 0.3507951356407858 + } + }, + { + "metric_id": "n2_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n2_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 43.87277829747428, + "tolerance": 2.193638914873714 + } + }, + { + "metric_id": "n3_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n3_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 18.24134705332086, + "tolerance": 0.9120673526660431 + } + }, + { + "metric_id": "r_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "r_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 25.912067352666043, + "tolerance": 1.2956033676333023 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case14/case14_21.json b/sleep_assessment/cases/case14/case14_21.json new file mode 100644 index 0000000000000000000000000000000000000000..39d3ce5c95d38b104cd56a5f5568e9c1f2ef6806 --- /dev/null +++ b/sleep_assessment/cases/case14/case14_21.json @@ -0,0 +1,67 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_01.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the percentage distribution of the five sleep stages W, N1, N2, N3, and R. Report the percentage of the total analyzed sleep-stage sequence occupied by each stage. Please clearly report all five final percentage values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract five sleep-stage percentage values from the agent's report:\n1) W stage percentage\n2) N1 stage percentage\n3) N2 stage percentage\n4) N3 stage percentage\n5) R/REM stage percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"w_stage_ratio_percent\", \"n1_stage_ratio_percent\", \"n2_stage_ratio_percent\", \"n3_stage_ratio_percent\", and \"r_stage_ratio_percent\".\n5. All values must be FLOATS or null.\n6. Values are percentages; if the report says 25.5%, output 25.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"w_stage_ratio_percent\": , \"n1_stage_ratio_percent\": , \"n2_stage_ratio_percent\": , \"n3_stage_ratio_percent\": , \"r_stage_ratio_percent\": }", + "metrics": [ + { + "metric_id": "w_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "w_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 43.43195266272189, + "tolerance": 2.1715976331360944 + } + }, + { + "metric_id": "n1_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n1_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 6.863905325443787, + "tolerance": 0.3431952662721894 + } + }, + { + "metric_id": "n2_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n2_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 31.597633136094675, + "tolerance": 1.5798816568047338 + } + }, + { + "metric_id": "n3_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n3_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 8.520710059171599, + "tolerance": 0.42603550295857995 + } + }, + { + "metric_id": "r_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "r_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 9.585798816568047, + "tolerance": 0.47928994082840237 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case14/case14_22.json b/sleep_assessment/cases/case14/case14_22.json new file mode 100644 index 0000000000000000000000000000000000000000..fede014f03745629c2ca170e977de41bd3cfbb04 --- /dev/null +++ b/sleep_assessment/cases/case14/case14_22.json @@ -0,0 +1,67 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_02.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the percentage distribution of the five sleep stages W, N1, N2, N3, and R. Report the percentage of the total analyzed sleep-stage sequence occupied by each stage. Please clearly report all five final percentage values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract five sleep-stage percentage values from the agent's report:\n1) W stage percentage\n2) N1 stage percentage\n3) N2 stage percentage\n4) N3 stage percentage\n5) R/REM stage percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"w_stage_ratio_percent\", \"n1_stage_ratio_percent\", \"n2_stage_ratio_percent\", \"n3_stage_ratio_percent\", and \"r_stage_ratio_percent\".\n5. All values must be FLOATS or null.\n6. Values are percentages; if the report says 25.5%, output 25.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"w_stage_ratio_percent\": , \"n1_stage_ratio_percent\": , \"n2_stage_ratio_percent\": , \"n3_stage_ratio_percent\": , \"r_stage_ratio_percent\": }", + "metrics": [ + { + "metric_id": "w_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "w_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 40.7979407979408, + "tolerance": 2.0398970398970397 + } + }, + { + "metric_id": "n1_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n1_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 17.245817245817246, + "tolerance": 0.8622908622908624 + } + }, + { + "metric_id": "n2_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n2_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 39.12483912483913, + "tolerance": 1.9562419562419564 + } + }, + { + "metric_id": "n3_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n3_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 2.7027027027027026, + "tolerance": 0.13513513513513514 + } + }, + { + "metric_id": "r_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "r_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 0.1287001287001287, + "tolerance": 0.0064350064350064354 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case14/case14_23.json b/sleep_assessment/cases/case14/case14_23.json new file mode 100644 index 0000000000000000000000000000000000000000..0005b4866171df5d0d6a0e77eaddbd53f9708b60 --- /dev/null +++ b/sleep_assessment/cases/case14/case14_23.json @@ -0,0 +1,67 @@ +{ + "meta_info": { + "case_id": "Physionet2018_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_03.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the percentage distribution of the five sleep stages W, N1, N2, N3, and R. Report the percentage of the total analyzed sleep-stage sequence occupied by each stage. Please clearly report all five final percentage values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract five sleep-stage percentage values from the agent's report:\n1) W stage percentage\n2) N1 stage percentage\n3) N2 stage percentage\n4) N3 stage percentage\n5) R/REM stage percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"w_stage_ratio_percent\", \"n1_stage_ratio_percent\", \"n2_stage_ratio_percent\", \"n3_stage_ratio_percent\", and \"r_stage_ratio_percent\".\n5. All values must be FLOATS or null.\n6. Values are percentages; if the report says 25.5%, output 25.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"w_stage_ratio_percent\": , \"n1_stage_ratio_percent\": , \"n2_stage_ratio_percent\": , \"n3_stage_ratio_percent\": , \"r_stage_ratio_percent\": }", + "metrics": [ + { + "metric_id": "w_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "w_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 53.78323108384458, + "tolerance": 2.6891615541922294 + } + }, + { + "metric_id": "n1_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n1_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 5.828220858895706, + "tolerance": 0.2914110429447853 + } + }, + { + "metric_id": "n2_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n2_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 33.94683026584867, + "tolerance": 1.6973415132924334 + } + }, + { + "metric_id": "n3_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n3_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 6.339468302658487, + "tolerance": 0.3169734151329244 + } + }, + { + "metric_id": "r_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "r_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 0.10224948875255624, + "tolerance": 0.005112474437627812 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case14/case14_24.json b/sleep_assessment/cases/case14/case14_24.json new file mode 100644 index 0000000000000000000000000000000000000000..7416c299504d52882fcc43db4946940515ef9a1f --- /dev/null +++ b/sleep_assessment/cases/case14/case14_24.json @@ -0,0 +1,67 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_04.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the percentage distribution of the five sleep stages W, N1, N2, N3, and R. Report the percentage of the total analyzed sleep-stage sequence occupied by each stage. Please clearly report all five final percentage values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract five sleep-stage percentage values from the agent's report:\n1) W stage percentage\n2) N1 stage percentage\n3) N2 stage percentage\n4) N3 stage percentage\n5) R/REM stage percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"w_stage_ratio_percent\", \"n1_stage_ratio_percent\", \"n2_stage_ratio_percent\", \"n3_stage_ratio_percent\", and \"r_stage_ratio_percent\".\n5. All values must be FLOATS or null.\n6. Values are percentages; if the report says 25.5%, output 25.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"w_stage_ratio_percent\": , \"n1_stage_ratio_percent\": , \"n2_stage_ratio_percent\": , \"n3_stage_ratio_percent\": , \"r_stage_ratio_percent\": }", + "metrics": [ + { + "metric_id": "w_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "w_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 9.597156398104266, + "tolerance": 0.4798578199052133 + } + }, + { + "metric_id": "n1_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n1_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 19.075829383886255, + "tolerance": 0.9537914691943128 + } + }, + { + "metric_id": "n2_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n2_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 40.28436018957346, + "tolerance": 2.014218009478673 + } + }, + { + "metric_id": "n3_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n3_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 20.379146919431278, + "tolerance": 1.018957345971564 + } + }, + { + "metric_id": "r_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "r_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 10.663507109004739, + "tolerance": 0.533175355450237 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case14/case14_25.json b/sleep_assessment/cases/case14/case14_25.json new file mode 100644 index 0000000000000000000000000000000000000000..fcfe34550818f2c19d5167dfe8d0d5d3656cb85e --- /dev/null +++ b/sleep_assessment/cases/case14/case14_25.json @@ -0,0 +1,67 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_05.npy", + "instruction": "Please use the provided sleep-stage label file to calculate the percentage distribution of the five sleep stages W, N1, N2, N3, and R. Report the percentage of the total analyzed sleep-stage sequence occupied by each stage. Please clearly report all five final percentage values." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract five sleep-stage percentage values from the agent's report:\n1) W stage percentage\n2) N1 stage percentage\n3) N2 stage percentage\n4) N3 stage percentage\n5) R/REM stage percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"w_stage_ratio_percent\", \"n1_stage_ratio_percent\", \"n2_stage_ratio_percent\", \"n3_stage_ratio_percent\", and \"r_stage_ratio_percent\".\n5. All values must be FLOATS or null.\n6. Values are percentages; if the report says 25.5%, output 25.5.\n7. If one value is missing or cannot be found, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"w_stage_ratio_percent\": , \"n1_stage_ratio_percent\": , \"n2_stage_ratio_percent\": , \"n3_stage_ratio_percent\": , \"r_stage_ratio_percent\": }", + "metrics": [ + { + "metric_id": "w_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "w_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 6.152241918665276, + "tolerance": 0.30761209593326383 + } + }, + { + "metric_id": "n1_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n1_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 34.82794577685089, + "tolerance": 1.7413972888425446 + } + }, + { + "metric_id": "n2_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n2_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 38.06047966631908, + "tolerance": 1.9030239833159541 + } + }, + { + "metric_id": "n3_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "n3_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 10.323253388946819, + "tolerance": 0.516162669447341 + } + }, + { + "metric_id": "r_stage_ratio_percent_accuracy", + "type": "numeric_check", + "target_key": "r_stage_ratio_percent", + "weight": 20, + "params": { + "gt_value": 10.636079249217936, + "tolerance": 0.5318039624608968 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case15/case15_01.json b/sleep_assessment/cases/case15/case15_01.json new file mode 100644 index 0000000000000000000000000000000000000000..757e543700336ee71be349b8b2e540963aaa7701 --- /dev/null +++ b/sleep_assessment/cases/case15/case15_01.json @@ -0,0 +1,125 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Please use the provided sleep-stage label file to determine whether each of the following sleep-structure abnormalities is present: A. N3 absent, B. REM absent, C. excessive wake proportion, D. low deep sleep (N3 deficit), E. low REM proportion, F. N2 dominance, G. early REM (short REM latency), H. abnormal N3 distribution (not front-loaded), I. abnormal REM distribution (not back-loaded), J. high fragmentation, K. excessive awakenings, L. short continuous sleep. Please return the abnormality letters that are present." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nThere are EXACTLY 12 candidate abnormality options: A, B, C, D, E, F, G, H, I, J, K, and L. Read the agent's report and determine for each letter whether the report indicates that the abnormality is present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY these 12 keys: \"A\", \"B\", \"C\", \"D\", \"E\", \"F\", \"G\", \"H\", \"I\", \"J\", \"K\", \"L\".\n5. Each value must be EXACTLY either \"Normal\" or \"Abnormal\".\n6. If the report explicitly lists or clearly states that a given abnormality exists, output \"Abnormal\" for that letter.\n7. Otherwise output \"Normal\" for that letter.\n8. Do NOT perform medical reasoning beyond extracting what the report says.\n\n### OUTPUT TEMPLATE\n{\"A\": \"Normal|Abnormal\", \"B\": \"Normal|Abnormal\", \"C\": \"Normal|Abnormal\", \"D\": \"Normal|Abnormal\", \"E\": \"Normal|Abnormal\", \"F\": \"Normal|Abnormal\", \"G\": \"Normal|Abnormal\", \"H\": \"Normal|Abnormal\", \"I\": \"Normal|Abnormal\", \"J\": \"Normal|Abnormal\", \"K\": \"Normal|Abnormal\", \"L\": \"Normal|Abnormal\"}", + "metrics": [ + { + "metric_id": "A_abnormality_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "B_abnormality_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "C_abnormality_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "D_abnormality_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "E_abnormality_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "F_abnormality_accuracy", + "type": "categorical_check", + "target_key": "F", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "G_abnormality_accuracy", + "type": "categorical_check", + "target_key": "G", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "H_abnormality_accuracy", + "type": "categorical_check", + "target_key": "H", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "I_abnormality_accuracy", + "type": "categorical_check", + "target_key": "I", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "J_abnormality_accuracy", + "type": "categorical_check", + "target_key": "J", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "K_abnormality_accuracy", + "type": "categorical_check", + "target_key": "K", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "L_abnormality_accuracy", + "type": "categorical_check", + "target_key": "L", + "weight": 8.3333333337, + "params": { + "gt_value": "Normal" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case15/case15_02.json b/sleep_assessment/cases/case15/case15_02.json new file mode 100644 index 0000000000000000000000000000000000000000..e4ed2485030491c3eb864f690b54626127c40eb3 --- /dev/null +++ b/sleep_assessment/cases/case15/case15_02.json @@ -0,0 +1,125 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_02.npy", + "instruction": "Please use the provided sleep-stage label file to determine whether each of the following sleep-structure abnormalities is present: A. N3 absent, B. REM absent, C. excessive wake proportion, D. low deep sleep (N3 deficit), E. low REM proportion, F. N2 dominance, G. early REM (short REM latency), H. abnormal N3 distribution (not front-loaded), I. abnormal REM distribution (not back-loaded), J. high fragmentation, K. excessive awakenings, L. short continuous sleep. Please return the abnormality letters that are present." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nThere are EXACTLY 12 candidate abnormality options: A, B, C, D, E, F, G, H, I, J, K, and L. Read the agent's report and determine for each letter whether the report indicates that the abnormality is present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY these 12 keys: \"A\", \"B\", \"C\", \"D\", \"E\", \"F\", \"G\", \"H\", \"I\", \"J\", \"K\", \"L\".\n5. Each value must be EXACTLY either \"Normal\" or \"Abnormal\".\n6. If the report explicitly lists or clearly states that a given abnormality exists, output \"Abnormal\" for that letter.\n7. Otherwise output \"Normal\" for that letter.\n8. Do NOT perform medical reasoning beyond extracting what the report says.\n\n### OUTPUT TEMPLATE\n{\"A\": \"Normal|Abnormal\", \"B\": \"Normal|Abnormal\", \"C\": \"Normal|Abnormal\", \"D\": \"Normal|Abnormal\", \"E\": \"Normal|Abnormal\", \"F\": \"Normal|Abnormal\", \"G\": \"Normal|Abnormal\", \"H\": \"Normal|Abnormal\", \"I\": \"Normal|Abnormal\", \"J\": \"Normal|Abnormal\", \"K\": \"Normal|Abnormal\", \"L\": \"Normal|Abnormal\"}", + "metrics": [ + { + "metric_id": "A_abnormality_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "B_abnormality_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "C_abnormality_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "D_abnormality_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "E_abnormality_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "F_abnormality_accuracy", + "type": "categorical_check", + "target_key": "F", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "G_abnormality_accuracy", + "type": "categorical_check", + "target_key": "G", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "H_abnormality_accuracy", + "type": "categorical_check", + "target_key": "H", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "I_abnormality_accuracy", + "type": "categorical_check", + "target_key": "I", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "J_abnormality_accuracy", + "type": "categorical_check", + "target_key": "J", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "K_abnormality_accuracy", + "type": "categorical_check", + "target_key": "K", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "L_abnormality_accuracy", + "type": "categorical_check", + "target_key": "L", + "weight": 8.3333333337, + "params": { + "gt_value": "Normal" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case15/case15_03.json b/sleep_assessment/cases/case15/case15_03.json new file mode 100644 index 0000000000000000000000000000000000000000..cd9fb5bb6b8d1826de6ba31a56d440a731468dd4 --- /dev/null +++ b/sleep_assessment/cases/case15/case15_03.json @@ -0,0 +1,125 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_03.npy", + "instruction": "Please use the provided sleep-stage label file to determine whether each of the following sleep-structure abnormalities is present: A. N3 absent, B. REM absent, C. excessive wake proportion, D. low deep sleep (N3 deficit), E. low REM proportion, F. N2 dominance, G. early REM (short REM latency), H. abnormal N3 distribution (not front-loaded), I. abnormal REM distribution (not back-loaded), J. high fragmentation, K. excessive awakenings, L. short continuous sleep. Please return the abnormality letters that are present." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nThere are EXACTLY 12 candidate abnormality options: A, B, C, D, E, F, G, H, I, J, K, and L. Read the agent's report and determine for each letter whether the report indicates that the abnormality is present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY these 12 keys: \"A\", \"B\", \"C\", \"D\", \"E\", \"F\", \"G\", \"H\", \"I\", \"J\", \"K\", \"L\".\n5. Each value must be EXACTLY either \"Normal\" or \"Abnormal\".\n6. If the report explicitly lists or clearly states that a given abnormality exists, output \"Abnormal\" for that letter.\n7. Otherwise output \"Normal\" for that letter.\n8. Do NOT perform medical reasoning beyond extracting what the report says.\n\n### OUTPUT TEMPLATE\n{\"A\": \"Normal|Abnormal\", \"B\": \"Normal|Abnormal\", \"C\": \"Normal|Abnormal\", \"D\": \"Normal|Abnormal\", \"E\": \"Normal|Abnormal\", \"F\": \"Normal|Abnormal\", \"G\": \"Normal|Abnormal\", \"H\": \"Normal|Abnormal\", \"I\": \"Normal|Abnormal\", \"J\": \"Normal|Abnormal\", \"K\": \"Normal|Abnormal\", \"L\": \"Normal|Abnormal\"}", + "metrics": [ + { + "metric_id": "A_abnormality_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "B_abnormality_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "C_abnormality_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "D_abnormality_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "E_abnormality_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "F_abnormality_accuracy", + "type": "categorical_check", + "target_key": "F", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "G_abnormality_accuracy", + "type": "categorical_check", + "target_key": "G", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "H_abnormality_accuracy", + "type": "categorical_check", + "target_key": "H", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "I_abnormality_accuracy", + "type": "categorical_check", + "target_key": "I", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "J_abnormality_accuracy", + "type": "categorical_check", + "target_key": "J", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "K_abnormality_accuracy", + "type": "categorical_check", + "target_key": "K", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "L_abnormality_accuracy", + "type": "categorical_check", + "target_key": "L", + "weight": 8.3333333337, + "params": { + "gt_value": "Normal" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case15/case15_04.json b/sleep_assessment/cases/case15/case15_04.json new file mode 100644 index 0000000000000000000000000000000000000000..d576a65508a5dcc347322b2bce1d7540f3bfb28c --- /dev/null +++ b/sleep_assessment/cases/case15/case15_04.json @@ -0,0 +1,125 @@ +{ + "meta_info": { + "case_id": "ISRUC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_04.npy", + "instruction": "Please use the provided sleep-stage label file to determine whether each of the following sleep-structure abnormalities is present: A. N3 absent, B. REM absent, C. excessive wake proportion, D. low deep sleep (N3 deficit), E. low REM proportion, F. N2 dominance, G. early REM (short REM latency), H. abnormal N3 distribution (not front-loaded), I. abnormal REM distribution (not back-loaded), J. high fragmentation, K. excessive awakenings, L. short continuous sleep. Please return the abnormality letters that are present." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nThere are EXACTLY 12 candidate abnormality options: A, B, C, D, E, F, G, H, I, J, K, and L. Read the agent's report and determine for each letter whether the report indicates that the abnormality is present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY these 12 keys: \"A\", \"B\", \"C\", \"D\", \"E\", \"F\", \"G\", \"H\", \"I\", \"J\", \"K\", \"L\".\n5. Each value must be EXACTLY either \"Normal\" or \"Abnormal\".\n6. If the report explicitly lists or clearly states that a given abnormality exists, output \"Abnormal\" for that letter.\n7. Otherwise output \"Normal\" for that letter.\n8. Do NOT perform medical reasoning beyond extracting what the report says.\n\n### OUTPUT TEMPLATE\n{\"A\": \"Normal|Abnormal\", \"B\": \"Normal|Abnormal\", \"C\": \"Normal|Abnormal\", \"D\": \"Normal|Abnormal\", \"E\": \"Normal|Abnormal\", \"F\": \"Normal|Abnormal\", \"G\": \"Normal|Abnormal\", \"H\": \"Normal|Abnormal\", \"I\": \"Normal|Abnormal\", \"J\": \"Normal|Abnormal\", \"K\": \"Normal|Abnormal\", \"L\": \"Normal|Abnormal\"}", + "metrics": [ + { + "metric_id": "A_abnormality_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "B_abnormality_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "C_abnormality_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "D_abnormality_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "E_abnormality_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "F_abnormality_accuracy", + "type": "categorical_check", + "target_key": "F", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "G_abnormality_accuracy", + "type": "categorical_check", + "target_key": "G", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "H_abnormality_accuracy", + "type": "categorical_check", + "target_key": "H", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "I_abnormality_accuracy", + "type": "categorical_check", + "target_key": "I", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "J_abnormality_accuracy", + "type": "categorical_check", + "target_key": "J", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "K_abnormality_accuracy", + "type": "categorical_check", + "target_key": "K", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "L_abnormality_accuracy", + "type": "categorical_check", + "target_key": "L", + "weight": 8.3333333337, + "params": { + "gt_value": "Normal" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case15/case15_05.json b/sleep_assessment/cases/case15/case15_05.json new file mode 100644 index 0000000000000000000000000000000000000000..56111bf68a3f71ddff92a6565cc4864adca4324a --- /dev/null +++ b/sleep_assessment/cases/case15/case15_05.json @@ -0,0 +1,125 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "label_path": "data/sleep/ISRUC_05.npy", + "instruction": "Please use the provided sleep-stage label file to determine whether each of the following sleep-structure abnormalities is present: A. N3 absent, B. REM absent, C. excessive wake proportion, D. low deep sleep (N3 deficit), E. low REM proportion, F. N2 dominance, G. early REM (short REM latency), H. abnormal N3 distribution (not front-loaded), I. abnormal REM distribution (not back-loaded), J. high fragmentation, K. excessive awakenings, L. short continuous sleep. Please return the abnormality letters that are present." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nThere are EXACTLY 12 candidate abnormality options: A, B, C, D, E, F, G, H, I, J, K, and L. Read the agent's report and determine for each letter whether the report indicates that the abnormality is present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY these 12 keys: \"A\", \"B\", \"C\", \"D\", \"E\", \"F\", \"G\", \"H\", \"I\", \"J\", \"K\", \"L\".\n5. Each value must be EXACTLY either \"Normal\" or \"Abnormal\".\n6. If the report explicitly lists or clearly states that a given abnormality exists, output \"Abnormal\" for that letter.\n7. Otherwise output \"Normal\" for that letter.\n8. Do NOT perform medical reasoning beyond extracting what the report says.\n\n### OUTPUT TEMPLATE\n{\"A\": \"Normal|Abnormal\", \"B\": \"Normal|Abnormal\", \"C\": \"Normal|Abnormal\", \"D\": \"Normal|Abnormal\", \"E\": \"Normal|Abnormal\", \"F\": \"Normal|Abnormal\", \"G\": \"Normal|Abnormal\", \"H\": \"Normal|Abnormal\", \"I\": \"Normal|Abnormal\", \"J\": \"Normal|Abnormal\", \"K\": \"Normal|Abnormal\", \"L\": \"Normal|Abnormal\"}", + "metrics": [ + { + "metric_id": "A_abnormality_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "B_abnormality_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "C_abnormality_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "D_abnormality_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "E_abnormality_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "F_abnormality_accuracy", + "type": "categorical_check", + "target_key": "F", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "G_abnormality_accuracy", + "type": "categorical_check", + "target_key": "G", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "H_abnormality_accuracy", + "type": "categorical_check", + "target_key": "H", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "I_abnormality_accuracy", + "type": "categorical_check", + "target_key": "I", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "J_abnormality_accuracy", + "type": "categorical_check", + "target_key": "J", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "K_abnormality_accuracy", + "type": "categorical_check", + "target_key": "K", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "L_abnormality_accuracy", + "type": "categorical_check", + "target_key": "L", + "weight": 8.3333333337, + "params": { + "gt_value": "Normal" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case15/case15_06.json b/sleep_assessment/cases/case15/case15_06.json new file mode 100644 index 0000000000000000000000000000000000000000..fee6ae815cd426ac68b5f6c17283ccc771bd6f1d --- /dev/null +++ b/sleep_assessment/cases/case15/case15_06.json @@ -0,0 +1,125 @@ +{ + "meta_info": { + "case_id": "HMC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_01.npy", + "instruction": "Please use the provided sleep-stage label file to determine whether each of the following sleep-structure abnormalities is present: A. N3 absent, B. REM absent, C. excessive wake proportion, D. low deep sleep (N3 deficit), E. low REM proportion, F. N2 dominance, G. early REM (short REM latency), H. abnormal N3 distribution (not front-loaded), I. abnormal REM distribution (not back-loaded), J. high fragmentation, K. excessive awakenings, L. short continuous sleep. Please return the abnormality letters that are present." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nThere are EXACTLY 12 candidate abnormality options: A, B, C, D, E, F, G, H, I, J, K, and L. Read the agent's report and determine for each letter whether the report indicates that the abnormality is present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY these 12 keys: \"A\", \"B\", \"C\", \"D\", \"E\", \"F\", \"G\", \"H\", \"I\", \"J\", \"K\", \"L\".\n5. Each value must be EXACTLY either \"Normal\" or \"Abnormal\".\n6. If the report explicitly lists or clearly states that a given abnormality exists, output \"Abnormal\" for that letter.\n7. Otherwise output \"Normal\" for that letter.\n8. Do NOT perform medical reasoning beyond extracting what the report says.\n\n### OUTPUT TEMPLATE\n{\"A\": \"Normal|Abnormal\", \"B\": \"Normal|Abnormal\", \"C\": \"Normal|Abnormal\", \"D\": \"Normal|Abnormal\", \"E\": \"Normal|Abnormal\", \"F\": \"Normal|Abnormal\", \"G\": \"Normal|Abnormal\", \"H\": \"Normal|Abnormal\", \"I\": \"Normal|Abnormal\", \"J\": \"Normal|Abnormal\", \"K\": \"Normal|Abnormal\", \"L\": \"Normal|Abnormal\"}", + "metrics": [ + { + "metric_id": "A_abnormality_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "B_abnormality_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "C_abnormality_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "D_abnormality_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "E_abnormality_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "F_abnormality_accuracy", + "type": "categorical_check", + "target_key": "F", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "G_abnormality_accuracy", + "type": "categorical_check", + "target_key": "G", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "H_abnormality_accuracy", + "type": "categorical_check", + "target_key": "H", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "I_abnormality_accuracy", + "type": "categorical_check", + "target_key": "I", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "J_abnormality_accuracy", + "type": "categorical_check", + "target_key": "J", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "K_abnormality_accuracy", + "type": "categorical_check", + "target_key": "K", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "L_abnormality_accuracy", + "type": "categorical_check", + "target_key": "L", + "weight": 8.3333333337, + "params": { + "gt_value": "Normal" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case15/case15_07.json b/sleep_assessment/cases/case15/case15_07.json new file mode 100644 index 0000000000000000000000000000000000000000..838f795e1cbf14ba41172fe9017de93029ef0f93 --- /dev/null +++ b/sleep_assessment/cases/case15/case15_07.json @@ -0,0 +1,125 @@ +{ + "meta_info": { + "case_id": "HMC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_02.npy", + "instruction": "Please use the provided sleep-stage label file to determine whether each of the following sleep-structure abnormalities is present: A. N3 absent, B. REM absent, C. excessive wake proportion, D. low deep sleep (N3 deficit), E. low REM proportion, F. N2 dominance, G. early REM (short REM latency), H. abnormal N3 distribution (not front-loaded), I. abnormal REM distribution (not back-loaded), J. high fragmentation, K. excessive awakenings, L. short continuous sleep. Please return the abnormality letters that are present." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nThere are EXACTLY 12 candidate abnormality options: A, B, C, D, E, F, G, H, I, J, K, and L. Read the agent's report and determine for each letter whether the report indicates that the abnormality is present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY these 12 keys: \"A\", \"B\", \"C\", \"D\", \"E\", \"F\", \"G\", \"H\", \"I\", \"J\", \"K\", \"L\".\n5. Each value must be EXACTLY either \"Normal\" or \"Abnormal\".\n6. If the report explicitly lists or clearly states that a given abnormality exists, output \"Abnormal\" for that letter.\n7. Otherwise output \"Normal\" for that letter.\n8. Do NOT perform medical reasoning beyond extracting what the report says.\n\n### OUTPUT TEMPLATE\n{\"A\": \"Normal|Abnormal\", \"B\": \"Normal|Abnormal\", \"C\": \"Normal|Abnormal\", \"D\": \"Normal|Abnormal\", \"E\": \"Normal|Abnormal\", \"F\": \"Normal|Abnormal\", \"G\": \"Normal|Abnormal\", \"H\": \"Normal|Abnormal\", \"I\": \"Normal|Abnormal\", \"J\": \"Normal|Abnormal\", \"K\": \"Normal|Abnormal\", \"L\": \"Normal|Abnormal\"}", + "metrics": [ + { + "metric_id": "A_abnormality_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "B_abnormality_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "C_abnormality_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "D_abnormality_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "E_abnormality_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "F_abnormality_accuracy", + "type": "categorical_check", + "target_key": "F", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "G_abnormality_accuracy", + "type": "categorical_check", + "target_key": "G", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "H_abnormality_accuracy", + "type": "categorical_check", + "target_key": "H", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "I_abnormality_accuracy", + "type": "categorical_check", + "target_key": "I", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "J_abnormality_accuracy", + "type": "categorical_check", + "target_key": "J", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "K_abnormality_accuracy", + "type": "categorical_check", + "target_key": "K", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "L_abnormality_accuracy", + "type": "categorical_check", + "target_key": "L", + "weight": 8.3333333337, + "params": { + "gt_value": "Normal" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case15/case15_08.json b/sleep_assessment/cases/case15/case15_08.json new file mode 100644 index 0000000000000000000000000000000000000000..f6e91865961d845b86347e02f1de17e1983dfcd4 --- /dev/null +++ b/sleep_assessment/cases/case15/case15_08.json @@ -0,0 +1,125 @@ +{ + "meta_info": { + "case_id": "HMC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_03.npy", + "instruction": "Please use the provided sleep-stage label file to determine whether each of the following sleep-structure abnormalities is present: A. N3 absent, B. REM absent, C. excessive wake proportion, D. low deep sleep (N3 deficit), E. low REM proportion, F. N2 dominance, G. early REM (short REM latency), H. abnormal N3 distribution (not front-loaded), I. abnormal REM distribution (not back-loaded), J. high fragmentation, K. excessive awakenings, L. short continuous sleep. Please return the abnormality letters that are present." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nThere are EXACTLY 12 candidate abnormality options: A, B, C, D, E, F, G, H, I, J, K, and L. Read the agent's report and determine for each letter whether the report indicates that the abnormality is present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY these 12 keys: \"A\", \"B\", \"C\", \"D\", \"E\", \"F\", \"G\", \"H\", \"I\", \"J\", \"K\", \"L\".\n5. Each value must be EXACTLY either \"Normal\" or \"Abnormal\".\n6. If the report explicitly lists or clearly states that a given abnormality exists, output \"Abnormal\" for that letter.\n7. Otherwise output \"Normal\" for that letter.\n8. Do NOT perform medical reasoning beyond extracting what the report says.\n\n### OUTPUT TEMPLATE\n{\"A\": \"Normal|Abnormal\", \"B\": \"Normal|Abnormal\", \"C\": \"Normal|Abnormal\", \"D\": \"Normal|Abnormal\", \"E\": \"Normal|Abnormal\", \"F\": \"Normal|Abnormal\", \"G\": \"Normal|Abnormal\", \"H\": \"Normal|Abnormal\", \"I\": \"Normal|Abnormal\", \"J\": \"Normal|Abnormal\", \"K\": \"Normal|Abnormal\", \"L\": \"Normal|Abnormal\"}", + "metrics": [ + { + "metric_id": "A_abnormality_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "B_abnormality_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "C_abnormality_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "D_abnormality_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "E_abnormality_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "F_abnormality_accuracy", + "type": "categorical_check", + "target_key": "F", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "G_abnormality_accuracy", + "type": "categorical_check", + "target_key": "G", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "H_abnormality_accuracy", + "type": "categorical_check", + "target_key": "H", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "I_abnormality_accuracy", + "type": "categorical_check", + "target_key": "I", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "J_abnormality_accuracy", + "type": "categorical_check", + "target_key": "J", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "K_abnormality_accuracy", + "type": "categorical_check", + "target_key": "K", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "L_abnormality_accuracy", + "type": "categorical_check", + "target_key": "L", + "weight": 8.3333333337, + "params": { + "gt_value": "Normal" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case15/case15_09.json b/sleep_assessment/cases/case15/case15_09.json new file mode 100644 index 0000000000000000000000000000000000000000..30084b9a1e0f303c91627b83c5b652ea35d72c14 --- /dev/null +++ b/sleep_assessment/cases/case15/case15_09.json @@ -0,0 +1,125 @@ +{ + "meta_info": { + "case_id": "HMC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_04.npy", + "instruction": "Please use the provided sleep-stage label file to determine whether each of the following sleep-structure abnormalities is present: A. N3 absent, B. REM absent, C. excessive wake proportion, D. low deep sleep (N3 deficit), E. low REM proportion, F. N2 dominance, G. early REM (short REM latency), H. abnormal N3 distribution (not front-loaded), I. abnormal REM distribution (not back-loaded), J. high fragmentation, K. excessive awakenings, L. short continuous sleep. Please return the abnormality letters that are present." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nThere are EXACTLY 12 candidate abnormality options: A, B, C, D, E, F, G, H, I, J, K, and L. Read the agent's report and determine for each letter whether the report indicates that the abnormality is present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY these 12 keys: \"A\", \"B\", \"C\", \"D\", \"E\", \"F\", \"G\", \"H\", \"I\", \"J\", \"K\", \"L\".\n5. Each value must be EXACTLY either \"Normal\" or \"Abnormal\".\n6. If the report explicitly lists or clearly states that a given abnormality exists, output \"Abnormal\" for that letter.\n7. Otherwise output \"Normal\" for that letter.\n8. Do NOT perform medical reasoning beyond extracting what the report says.\n\n### OUTPUT TEMPLATE\n{\"A\": \"Normal|Abnormal\", \"B\": \"Normal|Abnormal\", \"C\": \"Normal|Abnormal\", \"D\": \"Normal|Abnormal\", \"E\": \"Normal|Abnormal\", \"F\": \"Normal|Abnormal\", \"G\": \"Normal|Abnormal\", \"H\": \"Normal|Abnormal\", \"I\": \"Normal|Abnormal\", \"J\": \"Normal|Abnormal\", \"K\": \"Normal|Abnormal\", \"L\": \"Normal|Abnormal\"}", + "metrics": [ + { + "metric_id": "A_abnormality_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "B_abnormality_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "C_abnormality_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "D_abnormality_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "E_abnormality_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "F_abnormality_accuracy", + "type": "categorical_check", + "target_key": "F", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "G_abnormality_accuracy", + "type": "categorical_check", + "target_key": "G", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "H_abnormality_accuracy", + "type": "categorical_check", + "target_key": "H", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "I_abnormality_accuracy", + "type": "categorical_check", + "target_key": "I", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "J_abnormality_accuracy", + "type": "categorical_check", + "target_key": "J", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "K_abnormality_accuracy", + "type": "categorical_check", + "target_key": "K", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "L_abnormality_accuracy", + "type": "categorical_check", + "target_key": "L", + "weight": 8.3333333337, + "params": { + "gt_value": "Normal" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case15/case15_10.json b/sleep_assessment/cases/case15/case15_10.json new file mode 100644 index 0000000000000000000000000000000000000000..75a9b4903646b51e436969c11944042e678091e1 --- /dev/null +++ b/sleep_assessment/cases/case15/case15_10.json @@ -0,0 +1,125 @@ +{ + "meta_info": { + "case_id": "HMC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "HMC" + }, + "agent_input": { + "label_path": "data/sleep/HMC_05.npy", + "instruction": "Please use the provided sleep-stage label file to determine whether each of the following sleep-structure abnormalities is present: A. N3 absent, B. REM absent, C. excessive wake proportion, D. low deep sleep (N3 deficit), E. low REM proportion, F. N2 dominance, G. early REM (short REM latency), H. abnormal N3 distribution (not front-loaded), I. abnormal REM distribution (not back-loaded), J. high fragmentation, K. excessive awakenings, L. short continuous sleep. Please return the abnormality letters that are present." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nThere are EXACTLY 12 candidate abnormality options: A, B, C, D, E, F, G, H, I, J, K, and L. Read the agent's report and determine for each letter whether the report indicates that the abnormality is present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY these 12 keys: \"A\", \"B\", \"C\", \"D\", \"E\", \"F\", \"G\", \"H\", \"I\", \"J\", \"K\", \"L\".\n5. Each value must be EXACTLY either \"Normal\" or \"Abnormal\".\n6. If the report explicitly lists or clearly states that a given abnormality exists, output \"Abnormal\" for that letter.\n7. Otherwise output \"Normal\" for that letter.\n8. Do NOT perform medical reasoning beyond extracting what the report says.\n\n### OUTPUT TEMPLATE\n{\"A\": \"Normal|Abnormal\", \"B\": \"Normal|Abnormal\", \"C\": \"Normal|Abnormal\", \"D\": \"Normal|Abnormal\", \"E\": \"Normal|Abnormal\", \"F\": \"Normal|Abnormal\", \"G\": \"Normal|Abnormal\", \"H\": \"Normal|Abnormal\", \"I\": \"Normal|Abnormal\", \"J\": \"Normal|Abnormal\", \"K\": \"Normal|Abnormal\", \"L\": \"Normal|Abnormal\"}", + "metrics": [ + { + "metric_id": "A_abnormality_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "B_abnormality_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "C_abnormality_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "D_abnormality_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "E_abnormality_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "F_abnormality_accuracy", + "type": "categorical_check", + "target_key": "F", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "G_abnormality_accuracy", + "type": "categorical_check", + "target_key": "G", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "H_abnormality_accuracy", + "type": "categorical_check", + "target_key": "H", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "I_abnormality_accuracy", + "type": "categorical_check", + "target_key": "I", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "J_abnormality_accuracy", + "type": "categorical_check", + "target_key": "J", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "K_abnormality_accuracy", + "type": "categorical_check", + "target_key": "K", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "L_abnormality_accuracy", + "type": "categorical_check", + "target_key": "L", + "weight": 8.3333333337, + "params": { + "gt_value": "Abnormal" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case15/case15_11.json b/sleep_assessment/cases/case15/case15_11.json new file mode 100644 index 0000000000000000000000000000000000000000..c49cebd50fc0f71571f02326dde24a4260ac097b --- /dev/null +++ b/sleep_assessment/cases/case15/case15_11.json @@ -0,0 +1,125 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_01.npy", + "instruction": "Please use the provided sleep-stage label file to determine whether each of the following sleep-structure abnormalities is present: A. N3 absent, B. REM absent, C. excessive wake proportion, D. low deep sleep (N3 deficit), E. low REM proportion, F. N2 dominance, G. early REM (short REM latency), H. abnormal N3 distribution (not front-loaded), I. abnormal REM distribution (not back-loaded), J. high fragmentation, K. excessive awakenings, L. short continuous sleep. Please return the abnormality letters that are present." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nThere are EXACTLY 12 candidate abnormality options: A, B, C, D, E, F, G, H, I, J, K, and L. Read the agent's report and determine for each letter whether the report indicates that the abnormality is present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY these 12 keys: \"A\", \"B\", \"C\", \"D\", \"E\", \"F\", \"G\", \"H\", \"I\", \"J\", \"K\", \"L\".\n5. Each value must be EXACTLY either \"Normal\" or \"Abnormal\".\n6. If the report explicitly lists or clearly states that a given abnormality exists, output \"Abnormal\" for that letter.\n7. Otherwise output \"Normal\" for that letter.\n8. Do NOT perform medical reasoning beyond extracting what the report says.\n\n### OUTPUT TEMPLATE\n{\"A\": \"Normal|Abnormal\", \"B\": \"Normal|Abnormal\", \"C\": \"Normal|Abnormal\", \"D\": \"Normal|Abnormal\", \"E\": \"Normal|Abnormal\", \"F\": \"Normal|Abnormal\", \"G\": \"Normal|Abnormal\", \"H\": \"Normal|Abnormal\", \"I\": \"Normal|Abnormal\", \"J\": \"Normal|Abnormal\", \"K\": \"Normal|Abnormal\", \"L\": \"Normal|Abnormal\"}", + "metrics": [ + { + "metric_id": "A_abnormality_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "B_abnormality_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "C_abnormality_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "D_abnormality_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "E_abnormality_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "F_abnormality_accuracy", + "type": "categorical_check", + "target_key": "F", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "G_abnormality_accuracy", + "type": "categorical_check", + "target_key": "G", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "H_abnormality_accuracy", + "type": "categorical_check", + "target_key": "H", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "I_abnormality_accuracy", + "type": "categorical_check", + "target_key": "I", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "J_abnormality_accuracy", + "type": "categorical_check", + "target_key": "J", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "K_abnormality_accuracy", + "type": "categorical_check", + "target_key": "K", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "L_abnormality_accuracy", + "type": "categorical_check", + "target_key": "L", + "weight": 8.3333333337, + "params": { + "gt_value": "Normal" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case15/case15_12.json b/sleep_assessment/cases/case15/case15_12.json new file mode 100644 index 0000000000000000000000000000000000000000..6a119e10cd7f7f07c9f92c65b768fbc6bc67708f --- /dev/null +++ b/sleep_assessment/cases/case15/case15_12.json @@ -0,0 +1,125 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_02.npy", + "instruction": "Please use the provided sleep-stage label file to determine whether each of the following sleep-structure abnormalities is present: A. N3 absent, B. REM absent, C. excessive wake proportion, D. low deep sleep (N3 deficit), E. low REM proportion, F. N2 dominance, G. early REM (short REM latency), H. abnormal N3 distribution (not front-loaded), I. abnormal REM distribution (not back-loaded), J. high fragmentation, K. excessive awakenings, L. short continuous sleep. Please return the abnormality letters that are present." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nThere are EXACTLY 12 candidate abnormality options: A, B, C, D, E, F, G, H, I, J, K, and L. Read the agent's report and determine for each letter whether the report indicates that the abnormality is present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY these 12 keys: \"A\", \"B\", \"C\", \"D\", \"E\", \"F\", \"G\", \"H\", \"I\", \"J\", \"K\", \"L\".\n5. Each value must be EXACTLY either \"Normal\" or \"Abnormal\".\n6. If the report explicitly lists or clearly states that a given abnormality exists, output \"Abnormal\" for that letter.\n7. Otherwise output \"Normal\" for that letter.\n8. Do NOT perform medical reasoning beyond extracting what the report says.\n\n### OUTPUT TEMPLATE\n{\"A\": \"Normal|Abnormal\", \"B\": \"Normal|Abnormal\", \"C\": \"Normal|Abnormal\", \"D\": \"Normal|Abnormal\", \"E\": \"Normal|Abnormal\", \"F\": \"Normal|Abnormal\", \"G\": \"Normal|Abnormal\", \"H\": \"Normal|Abnormal\", \"I\": \"Normal|Abnormal\", \"J\": \"Normal|Abnormal\", \"K\": \"Normal|Abnormal\", \"L\": \"Normal|Abnormal\"}", + "metrics": [ + { + "metric_id": "A_abnormality_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "B_abnormality_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "C_abnormality_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "D_abnormality_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "E_abnormality_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "F_abnormality_accuracy", + "type": "categorical_check", + "target_key": "F", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "G_abnormality_accuracy", + "type": "categorical_check", + "target_key": "G", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "H_abnormality_accuracy", + "type": "categorical_check", + "target_key": "H", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "I_abnormality_accuracy", + "type": "categorical_check", + "target_key": "I", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "J_abnormality_accuracy", + "type": "categorical_check", + "target_key": "J", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "K_abnormality_accuracy", + "type": "categorical_check", + "target_key": "K", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "L_abnormality_accuracy", + "type": "categorical_check", + "target_key": "L", + "weight": 8.3333333337, + "params": { + "gt_value": "Normal" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case15/case15_13.json b/sleep_assessment/cases/case15/case15_13.json new file mode 100644 index 0000000000000000000000000000000000000000..3014d86c3ccc672d80d542554569f5f02728ddc3 --- /dev/null +++ b/sleep_assessment/cases/case15/case15_13.json @@ -0,0 +1,125 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_03.npy", + "instruction": "Please use the provided sleep-stage label file to determine whether each of the following sleep-structure abnormalities is present: A. N3 absent, B. REM absent, C. excessive wake proportion, D. low deep sleep (N3 deficit), E. low REM proportion, F. N2 dominance, G. early REM (short REM latency), H. abnormal N3 distribution (not front-loaded), I. abnormal REM distribution (not back-loaded), J. high fragmentation, K. excessive awakenings, L. short continuous sleep. Please return the abnormality letters that are present." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nThere are EXACTLY 12 candidate abnormality options: A, B, C, D, E, F, G, H, I, J, K, and L. Read the agent's report and determine for each letter whether the report indicates that the abnormality is present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY these 12 keys: \"A\", \"B\", \"C\", \"D\", \"E\", \"F\", \"G\", \"H\", \"I\", \"J\", \"K\", \"L\".\n5. Each value must be EXACTLY either \"Normal\" or \"Abnormal\".\n6. If the report explicitly lists or clearly states that a given abnormality exists, output \"Abnormal\" for that letter.\n7. Otherwise output \"Normal\" for that letter.\n8. Do NOT perform medical reasoning beyond extracting what the report says.\n\n### OUTPUT TEMPLATE\n{\"A\": \"Normal|Abnormal\", \"B\": \"Normal|Abnormal\", \"C\": \"Normal|Abnormal\", \"D\": \"Normal|Abnormal\", \"E\": \"Normal|Abnormal\", \"F\": \"Normal|Abnormal\", \"G\": \"Normal|Abnormal\", \"H\": \"Normal|Abnormal\", \"I\": \"Normal|Abnormal\", \"J\": \"Normal|Abnormal\", \"K\": \"Normal|Abnormal\", \"L\": \"Normal|Abnormal\"}", + "metrics": [ + { + "metric_id": "A_abnormality_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "B_abnormality_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "C_abnormality_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "D_abnormality_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "E_abnormality_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "F_abnormality_accuracy", + "type": "categorical_check", + "target_key": "F", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "G_abnormality_accuracy", + "type": "categorical_check", + "target_key": "G", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "H_abnormality_accuracy", + "type": "categorical_check", + "target_key": "H", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "I_abnormality_accuracy", + "type": "categorical_check", + "target_key": "I", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "J_abnormality_accuracy", + "type": "categorical_check", + "target_key": "J", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "K_abnormality_accuracy", + "type": "categorical_check", + "target_key": "K", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "L_abnormality_accuracy", + "type": "categorical_check", + "target_key": "L", + "weight": 8.3333333337, + "params": { + "gt_value": "Abnormal" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case15/case15_14.json b/sleep_assessment/cases/case15/case15_14.json new file mode 100644 index 0000000000000000000000000000000000000000..2549cbcd5cb1d3d71524d7287fe728a4ff2529ec --- /dev/null +++ b/sleep_assessment/cases/case15/case15_14.json @@ -0,0 +1,125 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_04.npy", + "instruction": "Please use the provided sleep-stage label file to determine whether each of the following sleep-structure abnormalities is present: A. N3 absent, B. REM absent, C. excessive wake proportion, D. low deep sleep (N3 deficit), E. low REM proportion, F. N2 dominance, G. early REM (short REM latency), H. abnormal N3 distribution (not front-loaded), I. abnormal REM distribution (not back-loaded), J. high fragmentation, K. excessive awakenings, L. short continuous sleep. Please return the abnormality letters that are present." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nThere are EXACTLY 12 candidate abnormality options: A, B, C, D, E, F, G, H, I, J, K, and L. Read the agent's report and determine for each letter whether the report indicates that the abnormality is present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY these 12 keys: \"A\", \"B\", \"C\", \"D\", \"E\", \"F\", \"G\", \"H\", \"I\", \"J\", \"K\", \"L\".\n5. Each value must be EXACTLY either \"Normal\" or \"Abnormal\".\n6. If the report explicitly lists or clearly states that a given abnormality exists, output \"Abnormal\" for that letter.\n7. Otherwise output \"Normal\" for that letter.\n8. Do NOT perform medical reasoning beyond extracting what the report says.\n\n### OUTPUT TEMPLATE\n{\"A\": \"Normal|Abnormal\", \"B\": \"Normal|Abnormal\", \"C\": \"Normal|Abnormal\", \"D\": \"Normal|Abnormal\", \"E\": \"Normal|Abnormal\", \"F\": \"Normal|Abnormal\", \"G\": \"Normal|Abnormal\", \"H\": \"Normal|Abnormal\", \"I\": \"Normal|Abnormal\", \"J\": \"Normal|Abnormal\", \"K\": \"Normal|Abnormal\", \"L\": \"Normal|Abnormal\"}", + "metrics": [ + { + "metric_id": "A_abnormality_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "B_abnormality_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "C_abnormality_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "D_abnormality_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "E_abnormality_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "F_abnormality_accuracy", + "type": "categorical_check", + "target_key": "F", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "G_abnormality_accuracy", + "type": "categorical_check", + "target_key": "G", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "H_abnormality_accuracy", + "type": "categorical_check", + "target_key": "H", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "I_abnormality_accuracy", + "type": "categorical_check", + "target_key": "I", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "J_abnormality_accuracy", + "type": "categorical_check", + "target_key": "J", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "K_abnormality_accuracy", + "type": "categorical_check", + "target_key": "K", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "L_abnormality_accuracy", + "type": "categorical_check", + "target_key": "L", + "weight": 8.3333333337, + "params": { + "gt_value": "Abnormal" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case15/case15_15.json b/sleep_assessment/cases/case15/case15_15.json new file mode 100644 index 0000000000000000000000000000000000000000..c3184b3866fc6cbe73b9123bb491f535328c4c50 --- /dev/null +++ b/sleep_assessment/cases/case15/case15_15.json @@ -0,0 +1,125 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "label_path": "data/sleep/SHHS1_05.npy", + "instruction": "Please use the provided sleep-stage label file to determine whether each of the following sleep-structure abnormalities is present: A. N3 absent, B. REM absent, C. excessive wake proportion, D. low deep sleep (N3 deficit), E. low REM proportion, F. N2 dominance, G. early REM (short REM latency), H. abnormal N3 distribution (not front-loaded), I. abnormal REM distribution (not back-loaded), J. high fragmentation, K. excessive awakenings, L. short continuous sleep. Please return the abnormality letters that are present." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nThere are EXACTLY 12 candidate abnormality options: A, B, C, D, E, F, G, H, I, J, K, and L. Read the agent's report and determine for each letter whether the report indicates that the abnormality is present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY these 12 keys: \"A\", \"B\", \"C\", \"D\", \"E\", \"F\", \"G\", \"H\", \"I\", \"J\", \"K\", \"L\".\n5. Each value must be EXACTLY either \"Normal\" or \"Abnormal\".\n6. If the report explicitly lists or clearly states that a given abnormality exists, output \"Abnormal\" for that letter.\n7. Otherwise output \"Normal\" for that letter.\n8. Do NOT perform medical reasoning beyond extracting what the report says.\n\n### OUTPUT TEMPLATE\n{\"A\": \"Normal|Abnormal\", \"B\": \"Normal|Abnormal\", \"C\": \"Normal|Abnormal\", \"D\": \"Normal|Abnormal\", \"E\": \"Normal|Abnormal\", \"F\": \"Normal|Abnormal\", \"G\": \"Normal|Abnormal\", \"H\": \"Normal|Abnormal\", \"I\": \"Normal|Abnormal\", \"J\": \"Normal|Abnormal\", \"K\": \"Normal|Abnormal\", \"L\": \"Normal|Abnormal\"}", + "metrics": [ + { + "metric_id": "A_abnormality_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "B_abnormality_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "C_abnormality_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "D_abnormality_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "E_abnormality_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "F_abnormality_accuracy", + "type": "categorical_check", + "target_key": "F", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "G_abnormality_accuracy", + "type": "categorical_check", + "target_key": "G", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "H_abnormality_accuracy", + "type": "categorical_check", + "target_key": "H", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "I_abnormality_accuracy", + "type": "categorical_check", + "target_key": "I", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "J_abnormality_accuracy", + "type": "categorical_check", + "target_key": "J", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "K_abnormality_accuracy", + "type": "categorical_check", + "target_key": "K", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "L_abnormality_accuracy", + "type": "categorical_check", + "target_key": "L", + "weight": 8.3333333337, + "params": { + "gt_value": "Normal" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case15/case15_16.json b/sleep_assessment/cases/case15/case15_16.json new file mode 100644 index 0000000000000000000000000000000000000000..32453c1567a139fbe3a4997dd36bef479e5a6ca8 --- /dev/null +++ b/sleep_assessment/cases/case15/case15_16.json @@ -0,0 +1,125 @@ +{ + "meta_info": { + "case_id": "MASSSS3_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_01.npy", + "instruction": "Please use the provided sleep-stage label file to determine whether each of the following sleep-structure abnormalities is present: A. N3 absent, B. REM absent, C. excessive wake proportion, D. low deep sleep (N3 deficit), E. low REM proportion, F. N2 dominance, G. early REM (short REM latency), H. abnormal N3 distribution (not front-loaded), I. abnormal REM distribution (not back-loaded), J. high fragmentation, K. excessive awakenings, L. short continuous sleep. Please return the abnormality letters that are present." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nThere are EXACTLY 12 candidate abnormality options: A, B, C, D, E, F, G, H, I, J, K, and L. Read the agent's report and determine for each letter whether the report indicates that the abnormality is present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY these 12 keys: \"A\", \"B\", \"C\", \"D\", \"E\", \"F\", \"G\", \"H\", \"I\", \"J\", \"K\", \"L\".\n5. Each value must be EXACTLY either \"Normal\" or \"Abnormal\".\n6. If the report explicitly lists or clearly states that a given abnormality exists, output \"Abnormal\" for that letter.\n7. Otherwise output \"Normal\" for that letter.\n8. Do NOT perform medical reasoning beyond extracting what the report says.\n\n### OUTPUT TEMPLATE\n{\"A\": \"Normal|Abnormal\", \"B\": \"Normal|Abnormal\", \"C\": \"Normal|Abnormal\", \"D\": \"Normal|Abnormal\", \"E\": \"Normal|Abnormal\", \"F\": \"Normal|Abnormal\", \"G\": \"Normal|Abnormal\", \"H\": \"Normal|Abnormal\", \"I\": \"Normal|Abnormal\", \"J\": \"Normal|Abnormal\", \"K\": \"Normal|Abnormal\", \"L\": \"Normal|Abnormal\"}", + "metrics": [ + { + "metric_id": "A_abnormality_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "B_abnormality_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "C_abnormality_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "D_abnormality_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "E_abnormality_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "F_abnormality_accuracy", + "type": "categorical_check", + "target_key": "F", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "G_abnormality_accuracy", + "type": "categorical_check", + "target_key": "G", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "H_abnormality_accuracy", + "type": "categorical_check", + "target_key": "H", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "I_abnormality_accuracy", + "type": "categorical_check", + "target_key": "I", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "J_abnormality_accuracy", + "type": "categorical_check", + "target_key": "J", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "K_abnormality_accuracy", + "type": "categorical_check", + "target_key": "K", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "L_abnormality_accuracy", + "type": "categorical_check", + "target_key": "L", + "weight": 8.3333333337, + "params": { + "gt_value": "Normal" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case15/case15_17.json b/sleep_assessment/cases/case15/case15_17.json new file mode 100644 index 0000000000000000000000000000000000000000..b3ac3c99c17f4242c40f1442f86fbee384f97ec7 --- /dev/null +++ b/sleep_assessment/cases/case15/case15_17.json @@ -0,0 +1,125 @@ +{ + "meta_info": { + "case_id": "MASSSS3_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_02.npy", + "instruction": "Please use the provided sleep-stage label file to determine whether each of the following sleep-structure abnormalities is present: A. N3 absent, B. REM absent, C. excessive wake proportion, D. low deep sleep (N3 deficit), E. low REM proportion, F. N2 dominance, G. early REM (short REM latency), H. abnormal N3 distribution (not front-loaded), I. abnormal REM distribution (not back-loaded), J. high fragmentation, K. excessive awakenings, L. short continuous sleep. Please return the abnormality letters that are present." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nThere are EXACTLY 12 candidate abnormality options: A, B, C, D, E, F, G, H, I, J, K, and L. Read the agent's report and determine for each letter whether the report indicates that the abnormality is present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY these 12 keys: \"A\", \"B\", \"C\", \"D\", \"E\", \"F\", \"G\", \"H\", \"I\", \"J\", \"K\", \"L\".\n5. Each value must be EXACTLY either \"Normal\" or \"Abnormal\".\n6. If the report explicitly lists or clearly states that a given abnormality exists, output \"Abnormal\" for that letter.\n7. Otherwise output \"Normal\" for that letter.\n8. Do NOT perform medical reasoning beyond extracting what the report says.\n\n### OUTPUT TEMPLATE\n{\"A\": \"Normal|Abnormal\", \"B\": \"Normal|Abnormal\", \"C\": \"Normal|Abnormal\", \"D\": \"Normal|Abnormal\", \"E\": \"Normal|Abnormal\", \"F\": \"Normal|Abnormal\", \"G\": \"Normal|Abnormal\", \"H\": \"Normal|Abnormal\", \"I\": \"Normal|Abnormal\", \"J\": \"Normal|Abnormal\", \"K\": \"Normal|Abnormal\", \"L\": \"Normal|Abnormal\"}", + "metrics": [ + { + "metric_id": "A_abnormality_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "B_abnormality_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "C_abnormality_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "D_abnormality_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "E_abnormality_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "F_abnormality_accuracy", + "type": "categorical_check", + "target_key": "F", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "G_abnormality_accuracy", + "type": "categorical_check", + "target_key": "G", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "H_abnormality_accuracy", + "type": "categorical_check", + "target_key": "H", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "I_abnormality_accuracy", + "type": "categorical_check", + "target_key": "I", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "J_abnormality_accuracy", + "type": "categorical_check", + "target_key": "J", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "K_abnormality_accuracy", + "type": "categorical_check", + "target_key": "K", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "L_abnormality_accuracy", + "type": "categorical_check", + "target_key": "L", + "weight": 8.3333333337, + "params": { + "gt_value": "Normal" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case15/case15_18.json b/sleep_assessment/cases/case15/case15_18.json new file mode 100644 index 0000000000000000000000000000000000000000..f838d83d9760c2ee89a3372065af2cb4f3e29698 --- /dev/null +++ b/sleep_assessment/cases/case15/case15_18.json @@ -0,0 +1,125 @@ +{ + "meta_info": { + "case_id": "MASSSS3_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_03.npy", + "instruction": "Please use the provided sleep-stage label file to determine whether each of the following sleep-structure abnormalities is present: A. N3 absent, B. REM absent, C. excessive wake proportion, D. low deep sleep (N3 deficit), E. low REM proportion, F. N2 dominance, G. early REM (short REM latency), H. abnormal N3 distribution (not front-loaded), I. abnormal REM distribution (not back-loaded), J. high fragmentation, K. excessive awakenings, L. short continuous sleep. Please return the abnormality letters that are present." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nThere are EXACTLY 12 candidate abnormality options: A, B, C, D, E, F, G, H, I, J, K, and L. Read the agent's report and determine for each letter whether the report indicates that the abnormality is present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY these 12 keys: \"A\", \"B\", \"C\", \"D\", \"E\", \"F\", \"G\", \"H\", \"I\", \"J\", \"K\", \"L\".\n5. Each value must be EXACTLY either \"Normal\" or \"Abnormal\".\n6. If the report explicitly lists or clearly states that a given abnormality exists, output \"Abnormal\" for that letter.\n7. Otherwise output \"Normal\" for that letter.\n8. Do NOT perform medical reasoning beyond extracting what the report says.\n\n### OUTPUT TEMPLATE\n{\"A\": \"Normal|Abnormal\", \"B\": \"Normal|Abnormal\", \"C\": \"Normal|Abnormal\", \"D\": \"Normal|Abnormal\", \"E\": \"Normal|Abnormal\", \"F\": \"Normal|Abnormal\", \"G\": \"Normal|Abnormal\", \"H\": \"Normal|Abnormal\", \"I\": \"Normal|Abnormal\", \"J\": \"Normal|Abnormal\", \"K\": \"Normal|Abnormal\", \"L\": \"Normal|Abnormal\"}", + "metrics": [ + { + "metric_id": "A_abnormality_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "B_abnormality_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "C_abnormality_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "D_abnormality_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "E_abnormality_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "F_abnormality_accuracy", + "type": "categorical_check", + "target_key": "F", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "G_abnormality_accuracy", + "type": "categorical_check", + "target_key": "G", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "H_abnormality_accuracy", + "type": "categorical_check", + "target_key": "H", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "I_abnormality_accuracy", + "type": "categorical_check", + "target_key": "I", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "J_abnormality_accuracy", + "type": "categorical_check", + "target_key": "J", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "K_abnormality_accuracy", + "type": "categorical_check", + "target_key": "K", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "L_abnormality_accuracy", + "type": "categorical_check", + "target_key": "L", + "weight": 8.3333333337, + "params": { + "gt_value": "Normal" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case15/case15_19.json b/sleep_assessment/cases/case15/case15_19.json new file mode 100644 index 0000000000000000000000000000000000000000..db0f5c05bbd969897beac7a09fea00876467e51f --- /dev/null +++ b/sleep_assessment/cases/case15/case15_19.json @@ -0,0 +1,125 @@ +{ + "meta_info": { + "case_id": "MASSSS3_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_04.npy", + "instruction": "Please use the provided sleep-stage label file to determine whether each of the following sleep-structure abnormalities is present: A. N3 absent, B. REM absent, C. excessive wake proportion, D. low deep sleep (N3 deficit), E. low REM proportion, F. N2 dominance, G. early REM (short REM latency), H. abnormal N3 distribution (not front-loaded), I. abnormal REM distribution (not back-loaded), J. high fragmentation, K. excessive awakenings, L. short continuous sleep. Please return the abnormality letters that are present." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nThere are EXACTLY 12 candidate abnormality options: A, B, C, D, E, F, G, H, I, J, K, and L. Read the agent's report and determine for each letter whether the report indicates that the abnormality is present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY these 12 keys: \"A\", \"B\", \"C\", \"D\", \"E\", \"F\", \"G\", \"H\", \"I\", \"J\", \"K\", \"L\".\n5. Each value must be EXACTLY either \"Normal\" or \"Abnormal\".\n6. If the report explicitly lists or clearly states that a given abnormality exists, output \"Abnormal\" for that letter.\n7. Otherwise output \"Normal\" for that letter.\n8. Do NOT perform medical reasoning beyond extracting what the report says.\n\n### OUTPUT TEMPLATE\n{\"A\": \"Normal|Abnormal\", \"B\": \"Normal|Abnormal\", \"C\": \"Normal|Abnormal\", \"D\": \"Normal|Abnormal\", \"E\": \"Normal|Abnormal\", \"F\": \"Normal|Abnormal\", \"G\": \"Normal|Abnormal\", \"H\": \"Normal|Abnormal\", \"I\": \"Normal|Abnormal\", \"J\": \"Normal|Abnormal\", \"K\": \"Normal|Abnormal\", \"L\": \"Normal|Abnormal\"}", + "metrics": [ + { + "metric_id": "A_abnormality_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "B_abnormality_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "C_abnormality_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "D_abnormality_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "E_abnormality_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "F_abnormality_accuracy", + "type": "categorical_check", + "target_key": "F", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "G_abnormality_accuracy", + "type": "categorical_check", + "target_key": "G", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "H_abnormality_accuracy", + "type": "categorical_check", + "target_key": "H", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "I_abnormality_accuracy", + "type": "categorical_check", + "target_key": "I", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "J_abnormality_accuracy", + "type": "categorical_check", + "target_key": "J", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "K_abnormality_accuracy", + "type": "categorical_check", + "target_key": "K", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "L_abnormality_accuracy", + "type": "categorical_check", + "target_key": "L", + "weight": 8.3333333337, + "params": { + "gt_value": "Normal" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case15/case15_20.json b/sleep_assessment/cases/case15/case15_20.json new file mode 100644 index 0000000000000000000000000000000000000000..f2073f22b434c512f7ccf545076be30b81258238 --- /dev/null +++ b/sleep_assessment/cases/case15/case15_20.json @@ -0,0 +1,125 @@ +{ + "meta_info": { + "case_id": "MASSSS3_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "label_path": "data/sleep/MASSSS3_05.npy", + "instruction": "Please use the provided sleep-stage label file to determine whether each of the following sleep-structure abnormalities is present: A. N3 absent, B. REM absent, C. excessive wake proportion, D. low deep sleep (N3 deficit), E. low REM proportion, F. N2 dominance, G. early REM (short REM latency), H. abnormal N3 distribution (not front-loaded), I. abnormal REM distribution (not back-loaded), J. high fragmentation, K. excessive awakenings, L. short continuous sleep. Please return the abnormality letters that are present." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nThere are EXACTLY 12 candidate abnormality options: A, B, C, D, E, F, G, H, I, J, K, and L. Read the agent's report and determine for each letter whether the report indicates that the abnormality is present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY these 12 keys: \"A\", \"B\", \"C\", \"D\", \"E\", \"F\", \"G\", \"H\", \"I\", \"J\", \"K\", \"L\".\n5. Each value must be EXACTLY either \"Normal\" or \"Abnormal\".\n6. If the report explicitly lists or clearly states that a given abnormality exists, output \"Abnormal\" for that letter.\n7. Otherwise output \"Normal\" for that letter.\n8. Do NOT perform medical reasoning beyond extracting what the report says.\n\n### OUTPUT TEMPLATE\n{\"A\": \"Normal|Abnormal\", \"B\": \"Normal|Abnormal\", \"C\": \"Normal|Abnormal\", \"D\": \"Normal|Abnormal\", \"E\": \"Normal|Abnormal\", \"F\": \"Normal|Abnormal\", \"G\": \"Normal|Abnormal\", \"H\": \"Normal|Abnormal\", \"I\": \"Normal|Abnormal\", \"J\": \"Normal|Abnormal\", \"K\": \"Normal|Abnormal\", \"L\": \"Normal|Abnormal\"}", + "metrics": [ + { + "metric_id": "A_abnormality_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "B_abnormality_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "C_abnormality_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "D_abnormality_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "E_abnormality_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "F_abnormality_accuracy", + "type": "categorical_check", + "target_key": "F", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "G_abnormality_accuracy", + "type": "categorical_check", + "target_key": "G", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "H_abnormality_accuracy", + "type": "categorical_check", + "target_key": "H", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "I_abnormality_accuracy", + "type": "categorical_check", + "target_key": "I", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "J_abnormality_accuracy", + "type": "categorical_check", + "target_key": "J", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "K_abnormality_accuracy", + "type": "categorical_check", + "target_key": "K", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "L_abnormality_accuracy", + "type": "categorical_check", + "target_key": "L", + "weight": 8.3333333337, + "params": { + "gt_value": "Normal" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case15/case15_21.json b/sleep_assessment/cases/case15/case15_21.json new file mode 100644 index 0000000000000000000000000000000000000000..706690a762ed54d65b78410ff316beede8fea806 --- /dev/null +++ b/sleep_assessment/cases/case15/case15_21.json @@ -0,0 +1,125 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_01.npy", + "instruction": "Please use the provided sleep-stage label file to determine whether each of the following sleep-structure abnormalities is present: A. N3 absent, B. REM absent, C. excessive wake proportion, D. low deep sleep (N3 deficit), E. low REM proportion, F. N2 dominance, G. early REM (short REM latency), H. abnormal N3 distribution (not front-loaded), I. abnormal REM distribution (not back-loaded), J. high fragmentation, K. excessive awakenings, L. short continuous sleep. Please return the abnormality letters that are present." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nThere are EXACTLY 12 candidate abnormality options: A, B, C, D, E, F, G, H, I, J, K, and L. Read the agent's report and determine for each letter whether the report indicates that the abnormality is present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY these 12 keys: \"A\", \"B\", \"C\", \"D\", \"E\", \"F\", \"G\", \"H\", \"I\", \"J\", \"K\", \"L\".\n5. Each value must be EXACTLY either \"Normal\" or \"Abnormal\".\n6. If the report explicitly lists or clearly states that a given abnormality exists, output \"Abnormal\" for that letter.\n7. Otherwise output \"Normal\" for that letter.\n8. Do NOT perform medical reasoning beyond extracting what the report says.\n\n### OUTPUT TEMPLATE\n{\"A\": \"Normal|Abnormal\", \"B\": \"Normal|Abnormal\", \"C\": \"Normal|Abnormal\", \"D\": \"Normal|Abnormal\", \"E\": \"Normal|Abnormal\", \"F\": \"Normal|Abnormal\", \"G\": \"Normal|Abnormal\", \"H\": \"Normal|Abnormal\", \"I\": \"Normal|Abnormal\", \"J\": \"Normal|Abnormal\", \"K\": \"Normal|Abnormal\", \"L\": \"Normal|Abnormal\"}", + "metrics": [ + { + "metric_id": "A_abnormality_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "B_abnormality_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "C_abnormality_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "D_abnormality_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "E_abnormality_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "F_abnormality_accuracy", + "type": "categorical_check", + "target_key": "F", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "G_abnormality_accuracy", + "type": "categorical_check", + "target_key": "G", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "H_abnormality_accuracy", + "type": "categorical_check", + "target_key": "H", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "I_abnormality_accuracy", + "type": "categorical_check", + "target_key": "I", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "J_abnormality_accuracy", + "type": "categorical_check", + "target_key": "J", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "K_abnormality_accuracy", + "type": "categorical_check", + "target_key": "K", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "L_abnormality_accuracy", + "type": "categorical_check", + "target_key": "L", + "weight": 8.3333333337, + "params": { + "gt_value": "Abnormal" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case15/case15_22.json b/sleep_assessment/cases/case15/case15_22.json new file mode 100644 index 0000000000000000000000000000000000000000..e1eae7ca8a369be0b60bd5dbe7a63b263c04f02b --- /dev/null +++ b/sleep_assessment/cases/case15/case15_22.json @@ -0,0 +1,125 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_02.npy", + "instruction": "Please use the provided sleep-stage label file to determine whether each of the following sleep-structure abnormalities is present: A. N3 absent, B. REM absent, C. excessive wake proportion, D. low deep sleep (N3 deficit), E. low REM proportion, F. N2 dominance, G. early REM (short REM latency), H. abnormal N3 distribution (not front-loaded), I. abnormal REM distribution (not back-loaded), J. high fragmentation, K. excessive awakenings, L. short continuous sleep. Please return the abnormality letters that are present." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nThere are EXACTLY 12 candidate abnormality options: A, B, C, D, E, F, G, H, I, J, K, and L. Read the agent's report and determine for each letter whether the report indicates that the abnormality is present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY these 12 keys: \"A\", \"B\", \"C\", \"D\", \"E\", \"F\", \"G\", \"H\", \"I\", \"J\", \"K\", \"L\".\n5. Each value must be EXACTLY either \"Normal\" or \"Abnormal\".\n6. If the report explicitly lists or clearly states that a given abnormality exists, output \"Abnormal\" for that letter.\n7. Otherwise output \"Normal\" for that letter.\n8. Do NOT perform medical reasoning beyond extracting what the report says.\n\n### OUTPUT TEMPLATE\n{\"A\": \"Normal|Abnormal\", \"B\": \"Normal|Abnormal\", \"C\": \"Normal|Abnormal\", \"D\": \"Normal|Abnormal\", \"E\": \"Normal|Abnormal\", \"F\": \"Normal|Abnormal\", \"G\": \"Normal|Abnormal\", \"H\": \"Normal|Abnormal\", \"I\": \"Normal|Abnormal\", \"J\": \"Normal|Abnormal\", \"K\": \"Normal|Abnormal\", \"L\": \"Normal|Abnormal\"}", + "metrics": [ + { + "metric_id": "A_abnormality_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "B_abnormality_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "C_abnormality_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "D_abnormality_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "E_abnormality_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "F_abnormality_accuracy", + "type": "categorical_check", + "target_key": "F", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "G_abnormality_accuracy", + "type": "categorical_check", + "target_key": "G", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "H_abnormality_accuracy", + "type": "categorical_check", + "target_key": "H", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "I_abnormality_accuracy", + "type": "categorical_check", + "target_key": "I", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "J_abnormality_accuracy", + "type": "categorical_check", + "target_key": "J", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "K_abnormality_accuracy", + "type": "categorical_check", + "target_key": "K", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "L_abnormality_accuracy", + "type": "categorical_check", + "target_key": "L", + "weight": 8.3333333337, + "params": { + "gt_value": "Normal" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case15/case15_23.json b/sleep_assessment/cases/case15/case15_23.json new file mode 100644 index 0000000000000000000000000000000000000000..30a91275e689e2679f6c312ae53e1f80770819d5 --- /dev/null +++ b/sleep_assessment/cases/case15/case15_23.json @@ -0,0 +1,125 @@ +{ + "meta_info": { + "case_id": "Physionet2018_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_03.npy", + "instruction": "Please use the provided sleep-stage label file to determine whether each of the following sleep-structure abnormalities is present: A. N3 absent, B. REM absent, C. excessive wake proportion, D. low deep sleep (N3 deficit), E. low REM proportion, F. N2 dominance, G. early REM (short REM latency), H. abnormal N3 distribution (not front-loaded), I. abnormal REM distribution (not back-loaded), J. high fragmentation, K. excessive awakenings, L. short continuous sleep. Please return the abnormality letters that are present." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nThere are EXACTLY 12 candidate abnormality options: A, B, C, D, E, F, G, H, I, J, K, and L. Read the agent's report and determine for each letter whether the report indicates that the abnormality is present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY these 12 keys: \"A\", \"B\", \"C\", \"D\", \"E\", \"F\", \"G\", \"H\", \"I\", \"J\", \"K\", \"L\".\n5. Each value must be EXACTLY either \"Normal\" or \"Abnormal\".\n6. If the report explicitly lists or clearly states that a given abnormality exists, output \"Abnormal\" for that letter.\n7. Otherwise output \"Normal\" for that letter.\n8. Do NOT perform medical reasoning beyond extracting what the report says.\n\n### OUTPUT TEMPLATE\n{\"A\": \"Normal|Abnormal\", \"B\": \"Normal|Abnormal\", \"C\": \"Normal|Abnormal\", \"D\": \"Normal|Abnormal\", \"E\": \"Normal|Abnormal\", \"F\": \"Normal|Abnormal\", \"G\": \"Normal|Abnormal\", \"H\": \"Normal|Abnormal\", \"I\": \"Normal|Abnormal\", \"J\": \"Normal|Abnormal\", \"K\": \"Normal|Abnormal\", \"L\": \"Normal|Abnormal\"}", + "metrics": [ + { + "metric_id": "A_abnormality_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "B_abnormality_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "C_abnormality_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "D_abnormality_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "E_abnormality_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "F_abnormality_accuracy", + "type": "categorical_check", + "target_key": "F", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "G_abnormality_accuracy", + "type": "categorical_check", + "target_key": "G", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "H_abnormality_accuracy", + "type": "categorical_check", + "target_key": "H", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "I_abnormality_accuracy", + "type": "categorical_check", + "target_key": "I", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "J_abnormality_accuracy", + "type": "categorical_check", + "target_key": "J", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "K_abnormality_accuracy", + "type": "categorical_check", + "target_key": "K", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "L_abnormality_accuracy", + "type": "categorical_check", + "target_key": "L", + "weight": 8.3333333337, + "params": { + "gt_value": "Abnormal" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case15/case15_24.json b/sleep_assessment/cases/case15/case15_24.json new file mode 100644 index 0000000000000000000000000000000000000000..936566688b9232c889a5f8fc550ac336ca82d0d7 --- /dev/null +++ b/sleep_assessment/cases/case15/case15_24.json @@ -0,0 +1,125 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_04.npy", + "instruction": "Please use the provided sleep-stage label file to determine whether each of the following sleep-structure abnormalities is present: A. N3 absent, B. REM absent, C. excessive wake proportion, D. low deep sleep (N3 deficit), E. low REM proportion, F. N2 dominance, G. early REM (short REM latency), H. abnormal N3 distribution (not front-loaded), I. abnormal REM distribution (not back-loaded), J. high fragmentation, K. excessive awakenings, L. short continuous sleep. Please return the abnormality letters that are present." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nThere are EXACTLY 12 candidate abnormality options: A, B, C, D, E, F, G, H, I, J, K, and L. Read the agent's report and determine for each letter whether the report indicates that the abnormality is present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY these 12 keys: \"A\", \"B\", \"C\", \"D\", \"E\", \"F\", \"G\", \"H\", \"I\", \"J\", \"K\", \"L\".\n5. Each value must be EXACTLY either \"Normal\" or \"Abnormal\".\n6. If the report explicitly lists or clearly states that a given abnormality exists, output \"Abnormal\" for that letter.\n7. Otherwise output \"Normal\" for that letter.\n8. Do NOT perform medical reasoning beyond extracting what the report says.\n\n### OUTPUT TEMPLATE\n{\"A\": \"Normal|Abnormal\", \"B\": \"Normal|Abnormal\", \"C\": \"Normal|Abnormal\", \"D\": \"Normal|Abnormal\", \"E\": \"Normal|Abnormal\", \"F\": \"Normal|Abnormal\", \"G\": \"Normal|Abnormal\", \"H\": \"Normal|Abnormal\", \"I\": \"Normal|Abnormal\", \"J\": \"Normal|Abnormal\", \"K\": \"Normal|Abnormal\", \"L\": \"Normal|Abnormal\"}", + "metrics": [ + { + "metric_id": "A_abnormality_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "B_abnormality_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "C_abnormality_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "D_abnormality_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "E_abnormality_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "F_abnormality_accuracy", + "type": "categorical_check", + "target_key": "F", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "G_abnormality_accuracy", + "type": "categorical_check", + "target_key": "G", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "H_abnormality_accuracy", + "type": "categorical_check", + "target_key": "H", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "I_abnormality_accuracy", + "type": "categorical_check", + "target_key": "I", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "J_abnormality_accuracy", + "type": "categorical_check", + "target_key": "J", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "K_abnormality_accuracy", + "type": "categorical_check", + "target_key": "K", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "L_abnormality_accuracy", + "type": "categorical_check", + "target_key": "L", + "weight": 8.3333333337, + "params": { + "gt_value": "Normal" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case15/case15_25.json b/sleep_assessment/cases/case15/case15_25.json new file mode 100644 index 0000000000000000000000000000000000000000..8f78130932f84270e0b3b3a51d5e85785e46c5e0 --- /dev/null +++ b/sleep_assessment/cases/case15/case15_25.json @@ -0,0 +1,125 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "label_path": "data/sleep/Physionet2018_05.npy", + "instruction": "Please use the provided sleep-stage label file to determine whether each of the following sleep-structure abnormalities is present: A. N3 absent, B. REM absent, C. excessive wake proportion, D. low deep sleep (N3 deficit), E. low REM proportion, F. N2 dominance, G. early REM (short REM latency), H. abnormal N3 distribution (not front-loaded), I. abnormal REM distribution (not back-loaded), J. high fragmentation, K. excessive awakenings, L. short continuous sleep. Please return the abnormality letters that are present." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nThere are EXACTLY 12 candidate abnormality options: A, B, C, D, E, F, G, H, I, J, K, and L. Read the agent's report and determine for each letter whether the report indicates that the abnormality is present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY these 12 keys: \"A\", \"B\", \"C\", \"D\", \"E\", \"F\", \"G\", \"H\", \"I\", \"J\", \"K\", \"L\".\n5. Each value must be EXACTLY either \"Normal\" or \"Abnormal\".\n6. If the report explicitly lists or clearly states that a given abnormality exists, output \"Abnormal\" for that letter.\n7. Otherwise output \"Normal\" for that letter.\n8. Do NOT perform medical reasoning beyond extracting what the report says.\n\n### OUTPUT TEMPLATE\n{\"A\": \"Normal|Abnormal\", \"B\": \"Normal|Abnormal\", \"C\": \"Normal|Abnormal\", \"D\": \"Normal|Abnormal\", \"E\": \"Normal|Abnormal\", \"F\": \"Normal|Abnormal\", \"G\": \"Normal|Abnormal\", \"H\": \"Normal|Abnormal\", \"I\": \"Normal|Abnormal\", \"J\": \"Normal|Abnormal\", \"K\": \"Normal|Abnormal\", \"L\": \"Normal|Abnormal\"}", + "metrics": [ + { + "metric_id": "A_abnormality_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "B_abnormality_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "C_abnormality_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "D_abnormality_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "E_abnormality_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "F_abnormality_accuracy", + "type": "categorical_check", + "target_key": "F", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "G_abnormality_accuracy", + "type": "categorical_check", + "target_key": "G", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "H_abnormality_accuracy", + "type": "categorical_check", + "target_key": "H", + "weight": 8.3333333333, + "params": { + "gt_value": "Abnormal" + } + }, + { + "metric_id": "I_abnormality_accuracy", + "type": "categorical_check", + "target_key": "I", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "J_abnormality_accuracy", + "type": "categorical_check", + "target_key": "J", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "K_abnormality_accuracy", + "type": "categorical_check", + "target_key": "K", + "weight": 8.3333333333, + "params": { + "gt_value": "Normal" + } + }, + { + "metric_id": "L_abnormality_accuracy", + "type": "categorical_check", + "target_key": "L", + "weight": 8.3333333337, + "params": { + "gt_value": "Normal" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case16/case16_01.json b/sleep_assessment/cases/case16/case16_01.json new file mode 100644 index 0000000000000000000000000000000000000000..cea6449a1f62fa96834db39cc882c8244e2af1c3 --- /dev/null +++ b/sleep_assessment/cases/case16/case16_01.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "instruction": "Please use the provided sleep file to analyze the segment from minute 218.5 to minute 221.5. Determine the dominant frequency band of this segment and the sleep stage of this segment. Report the dominant band using one of: delta, theta, alpha, beta, gamma, None. Report the sleep stage using one of: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the dominant frequency band of the specified segment\n2) the sleep stage of the specified segment\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"dominant_band\" and \"sleep_stage\".\n5. \"dominant_band\" must be exactly one of: \"delta\", \"theta\", \"alpha\", \"beta\", \"gamma\", or null.\n6. \"sleep_stage\" must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\", or null.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"dominant_band\": <\"delta\"|\"theta\"|\"alpha\"|\"beta\"|\"gamma\"|null>, \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "dominant_band_accuracy", + "type": "categorical_check", + "target_key": "dominant_band", + "weight": 50, + "params": { + "gt_value": "delta" + } + }, + { + "metric_id": "sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 50, + "params": { + "gt_value": "N1" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case16/case16_02.json b/sleep_assessment/cases/case16/case16_02.json new file mode 100644 index 0000000000000000000000000000000000000000..e3633cc60589a5c753d5baa8ff16ac3e45731c4b --- /dev/null +++ b/sleep_assessment/cases/case16/case16_02.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "instruction": "Please use the provided sleep file to analyze the segment from minute 464.0 to minute 467.0. Determine the dominant frequency band of this segment and the sleep stage of this segment. Report the dominant band using one of: delta, theta, alpha, beta, gamma, None. Report the sleep stage using one of: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the dominant frequency band of the specified segment\n2) the sleep stage of the specified segment\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"dominant_band\" and \"sleep_stage\".\n5. \"dominant_band\" must be exactly one of: \"delta\", \"theta\", \"alpha\", \"beta\", \"gamma\", or null.\n6. \"sleep_stage\" must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\", or null.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"dominant_band\": <\"delta\"|\"theta\"|\"alpha\"|\"beta\"|\"gamma\"|null>, \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "dominant_band_accuracy", + "type": "categorical_check", + "target_key": "dominant_band", + "weight": 50, + "params": { + "gt_value": "delta" + } + }, + { + "metric_id": "sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 50, + "params": { + "gt_value": "R" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case16/case16_03.json b/sleep_assessment/cases/case16/case16_03.json new file mode 100644 index 0000000000000000000000000000000000000000..3b61f96a5b70caae7b161c9c3a1920d9ae4bb190 --- /dev/null +++ b/sleep_assessment/cases/case16/case16_03.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "instruction": "Please use the provided sleep file to analyze the segment from minute 0.5 to minute 3.5. Determine the dominant frequency band of this segment and the sleep stage of this segment. Report the dominant band using one of: delta, theta, alpha, beta, gamma, None. Report the sleep stage using one of: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the dominant frequency band of the specified segment\n2) the sleep stage of the specified segment\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"dominant_band\" and \"sleep_stage\".\n5. \"dominant_band\" must be exactly one of: \"delta\", \"theta\", \"alpha\", \"beta\", \"gamma\", or null.\n6. \"sleep_stage\" must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\", or null.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"dominant_band\": <\"delta\"|\"theta\"|\"alpha\"|\"beta\"|\"gamma\"|null>, \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "dominant_band_accuracy", + "type": "categorical_check", + "target_key": "dominant_band", + "weight": 50, + "params": { + "gt_value": "alpha" + } + }, + { + "metric_id": "sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 50, + "params": { + "gt_value": "W" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case16/case16_04.json b/sleep_assessment/cases/case16/case16_04.json new file mode 100644 index 0000000000000000000000000000000000000000..f5bb0851e4b4d7586e5f1d27e08be46d355e9f25 --- /dev/null +++ b/sleep_assessment/cases/case16/case16_04.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "ISRUC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_04.edf", + "instruction": "Please use the provided sleep file to analyze the segment from minute 42.5 to minute 45.5. Determine the dominant frequency band of this segment and the sleep stage of this segment. Report the dominant band using one of: delta, theta, alpha, beta, gamma, None. Report the sleep stage using one of: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the dominant frequency band of the specified segment\n2) the sleep stage of the specified segment\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"dominant_band\" and \"sleep_stage\".\n5. \"dominant_band\" must be exactly one of: \"delta\", \"theta\", \"alpha\", \"beta\", \"gamma\", or null.\n6. \"sleep_stage\" must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\", or null.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"dominant_band\": <\"delta\"|\"theta\"|\"alpha\"|\"beta\"|\"gamma\"|null>, \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "dominant_band_accuracy", + "type": "categorical_check", + "target_key": "dominant_band", + "weight": 50, + "params": { + "gt_value": "delta" + } + }, + { + "metric_id": "sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 50, + "params": { + "gt_value": "N2" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case16/case16_05.json b/sleep_assessment/cases/case16/case16_05.json new file mode 100644 index 0000000000000000000000000000000000000000..f8f14291d22a0ac8b74dbd6a5010e42a63d1a232 --- /dev/null +++ b/sleep_assessment/cases/case16/case16_05.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "instruction": "Please use the provided sleep file to analyze the segment from minute 196.5 to minute 199.5. Determine the dominant frequency band of this segment and the sleep stage of this segment. Report the dominant band using one of: delta, theta, alpha, beta, gamma, None. Report the sleep stage using one of: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the dominant frequency band of the specified segment\n2) the sleep stage of the specified segment\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"dominant_band\" and \"sleep_stage\".\n5. \"dominant_band\" must be exactly one of: \"delta\", \"theta\", \"alpha\", \"beta\", \"gamma\", or null.\n6. \"sleep_stage\" must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\", or null.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"dominant_band\": <\"delta\"|\"theta\"|\"alpha\"|\"beta\"|\"gamma\"|null>, \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "dominant_band_accuracy", + "type": "categorical_check", + "target_key": "dominant_band", + "weight": 50, + "params": { + "gt_value": "delta" + } + }, + { + "metric_id": "sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 50, + "params": { + "gt_value": "N3" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case16/case16_06.json b/sleep_assessment/cases/case16/case16_06.json new file mode 100644 index 0000000000000000000000000000000000000000..4576cf243bcdee97f00ad324b03ce26973b0d81f --- /dev/null +++ b/sleep_assessment/cases/case16/case16_06.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "HMC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_01.edf", + "instruction": "Please use the provided sleep file to analyze the segment from minute 120.5 to minute 123.5. Determine the dominant frequency band of this segment and the sleep stage of this segment. Report the dominant band using one of: delta, theta, alpha, beta, gamma, None. Report the sleep stage using one of: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the dominant frequency band of the specified segment\n2) the sleep stage of the specified segment\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"dominant_band\" and \"sleep_stage\".\n5. \"dominant_band\" must be exactly one of: \"delta\", \"theta\", \"alpha\", \"beta\", \"gamma\", or null.\n6. \"sleep_stage\" must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\", or null.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"dominant_band\": <\"delta\"|\"theta\"|\"alpha\"|\"beta\"|\"gamma\"|null>, \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "dominant_band_accuracy", + "type": "categorical_check", + "target_key": "dominant_band", + "weight": 50, + "params": { + "gt_value": "delta" + } + }, + { + "metric_id": "sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 50, + "params": { + "gt_value": "N1" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case16/case16_07.json b/sleep_assessment/cases/case16/case16_07.json new file mode 100644 index 0000000000000000000000000000000000000000..14fa27264706abaafb629bef4f705d4fba806e63 --- /dev/null +++ b/sleep_assessment/cases/case16/case16_07.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "HMC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_02.edf", + "instruction": "Please use the provided sleep file to analyze the segment from minute 303.5 to minute 306.5. Determine the dominant frequency band of this segment and the sleep stage of this segment. Report the dominant band using one of: delta, theta, alpha, beta, gamma, None. Report the sleep stage using one of: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the dominant frequency band of the specified segment\n2) the sleep stage of the specified segment\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"dominant_band\" and \"sleep_stage\".\n5. \"dominant_band\" must be exactly one of: \"delta\", \"theta\", \"alpha\", \"beta\", \"gamma\", or null.\n6. \"sleep_stage\" must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\", or null.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"dominant_band\": <\"delta\"|\"theta\"|\"alpha\"|\"beta\"|\"gamma\"|null>, \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "dominant_band_accuracy", + "type": "categorical_check", + "target_key": "dominant_band", + "weight": 50, + "params": { + "gt_value": "delta" + } + }, + { + "metric_id": "sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 50, + "params": { + "gt_value": "R" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case16/case16_08.json b/sleep_assessment/cases/case16/case16_08.json new file mode 100644 index 0000000000000000000000000000000000000000..8a5b6984f4c695ef7142d166c194f4dca0bab7f2 --- /dev/null +++ b/sleep_assessment/cases/case16/case16_08.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "HMC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_03.edf", + "instruction": "Please use the provided sleep file to analyze the segment from minute 234.0 to minute 237.0. Determine the dominant frequency band of this segment and the sleep stage of this segment. Report the dominant band using one of: delta, theta, alpha, beta, gamma, None. Report the sleep stage using one of: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the dominant frequency band of the specified segment\n2) the sleep stage of the specified segment\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"dominant_band\" and \"sleep_stage\".\n5. \"dominant_band\" must be exactly one of: \"delta\", \"theta\", \"alpha\", \"beta\", \"gamma\", or null.\n6. \"sleep_stage\" must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\", or null.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"dominant_band\": <\"delta\"|\"theta\"|\"alpha\"|\"beta\"|\"gamma\"|null>, \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "dominant_band_accuracy", + "type": "categorical_check", + "target_key": "dominant_band", + "weight": 50, + "params": { + "gt_value": "delta" + } + }, + { + "metric_id": "sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 50, + "params": { + "gt_value": "N3" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case16/case16_09.json b/sleep_assessment/cases/case16/case16_09.json new file mode 100644 index 0000000000000000000000000000000000000000..4a8a985552963d5c7e53a3065c4f008b6ee82717 --- /dev/null +++ b/sleep_assessment/cases/case16/case16_09.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "HMC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_04.edf", + "instruction": "Please use the provided sleep file to analyze the segment from minute 416.5 to minute 419.5. Determine the dominant frequency band of this segment and the sleep stage of this segment. Report the dominant band using one of: delta, theta, alpha, beta, gamma, None. Report the sleep stage using one of: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the dominant frequency band of the specified segment\n2) the sleep stage of the specified segment\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"dominant_band\" and \"sleep_stage\".\n5. \"dominant_band\" must be exactly one of: \"delta\", \"theta\", \"alpha\", \"beta\", \"gamma\", or null.\n6. \"sleep_stage\" must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\", or null.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"dominant_band\": <\"delta\"|\"theta\"|\"alpha\"|\"beta\"|\"gamma\"|null>, \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "dominant_band_accuracy", + "type": "categorical_check", + "target_key": "dominant_band", + "weight": 50, + "params": { + "gt_value": "delta" + } + }, + { + "metric_id": "sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 50, + "params": { + "gt_value": "N2" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case16/case16_10.json b/sleep_assessment/cases/case16/case16_10.json new file mode 100644 index 0000000000000000000000000000000000000000..1092a48854c3dd8b8627d87cf297b27aed16bc6d --- /dev/null +++ b/sleep_assessment/cases/case16/case16_10.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "HMC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_05.edf", + "instruction": "Please use the provided sleep file to analyze the segment from minute 398.5 to minute 401.5. Determine the dominant frequency band of this segment and the sleep stage of this segment. Report the dominant band using one of: delta, theta, alpha, beta, gamma, None. Report the sleep stage using one of: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the dominant frequency band of the specified segment\n2) the sleep stage of the specified segment\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"dominant_band\" and \"sleep_stage\".\n5. \"dominant_band\" must be exactly one of: \"delta\", \"theta\", \"alpha\", \"beta\", \"gamma\", or null.\n6. \"sleep_stage\" must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\", or null.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"dominant_band\": <\"delta\"|\"theta\"|\"alpha\"|\"beta\"|\"gamma\"|null>, \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "dominant_band_accuracy", + "type": "categorical_check", + "target_key": "dominant_band", + "weight": 50, + "params": { + "gt_value": "delta" + } + }, + { + "metric_id": "sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 50, + "params": { + "gt_value": "W" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case16/case16_11.json b/sleep_assessment/cases/case16/case16_11.json new file mode 100644 index 0000000000000000000000000000000000000000..f1a513e6d2e8a152bd094e40171b973590f4a31b --- /dev/null +++ b/sleep_assessment/cases/case16/case16_11.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "instruction": "Please use the provided sleep file to analyze the segment from minute 149.0 to minute 152.0. Determine the dominant frequency band of this segment and the sleep stage of this segment. Report the dominant band using one of: delta, theta, alpha, beta, gamma, None. Report the sleep stage using one of: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the dominant frequency band of the specified segment\n2) the sleep stage of the specified segment\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"dominant_band\" and \"sleep_stage\".\n5. \"dominant_band\" must be exactly one of: \"delta\", \"theta\", \"alpha\", \"beta\", \"gamma\", or null.\n6. \"sleep_stage\" must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\", or null.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"dominant_band\": <\"delta\"|\"theta\"|\"alpha\"|\"beta\"|\"gamma\"|null>, \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "dominant_band_accuracy", + "type": "categorical_check", + "target_key": "dominant_band", + "weight": 50, + "params": { + "gt_value": "delta" + } + }, + { + "metric_id": "sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 50, + "params": { + "gt_value": "N2" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case16/case16_12.json b/sleep_assessment/cases/case16/case16_12.json new file mode 100644 index 0000000000000000000000000000000000000000..fb09c848b4abb9a2770e6f13c8e1037445e6bf5e --- /dev/null +++ b/sleep_assessment/cases/case16/case16_12.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "instruction": "Please use the provided sleep file to analyze the segment from minute 536.5 to minute 539.5. Determine the dominant frequency band of this segment and the sleep stage of this segment. Report the dominant band using one of: delta, theta, alpha, beta, gamma, None. Report the sleep stage using one of: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the dominant frequency band of the specified segment\n2) the sleep stage of the specified segment\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"dominant_band\" and \"sleep_stage\".\n5. \"dominant_band\" must be exactly one of: \"delta\", \"theta\", \"alpha\", \"beta\", \"gamma\", or null.\n6. \"sleep_stage\" must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\", or null.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"dominant_band\": <\"delta\"|\"theta\"|\"alpha\"|\"beta\"|\"gamma\"|null>, \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "dominant_band_accuracy", + "type": "categorical_check", + "target_key": "dominant_band", + "weight": 50, + "params": { + "gt_value": "alpha" + } + }, + { + "metric_id": "sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 50, + "params": { + "gt_value": "W" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case16/case16_13.json b/sleep_assessment/cases/case16/case16_13.json new file mode 100644 index 0000000000000000000000000000000000000000..1acd834fd7f75abd55717a769d5718444f90597c --- /dev/null +++ b/sleep_assessment/cases/case16/case16_13.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "instruction": "Please use the provided sleep file to analyze the segment from minute 440.0 to minute 443.0. Determine the dominant frequency band of this segment and the sleep stage of this segment. Report the dominant band using one of: delta, theta, alpha, beta, gamma, None. Report the sleep stage using one of: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the dominant frequency band of the specified segment\n2) the sleep stage of the specified segment\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"dominant_band\" and \"sleep_stage\".\n5. \"dominant_band\" must be exactly one of: \"delta\", \"theta\", \"alpha\", \"beta\", \"gamma\", or null.\n6. \"sleep_stage\" must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\", or null.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"dominant_band\": <\"delta\"|\"theta\"|\"alpha\"|\"beta\"|\"gamma\"|null>, \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "dominant_band_accuracy", + "type": "categorical_check", + "target_key": "dominant_band", + "weight": 50, + "params": { + "gt_value": "delta" + } + }, + { + "metric_id": "sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 50, + "params": { + "gt_value": "N1" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case16/case16_14.json b/sleep_assessment/cases/case16/case16_14.json new file mode 100644 index 0000000000000000000000000000000000000000..c0a3c8d42ce43ec2dd025b62dbc4e8c829ef5896 --- /dev/null +++ b/sleep_assessment/cases/case16/case16_14.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "instruction": "Please use the provided sleep file to analyze the segment from minute 149.5 to minute 152.5. Determine the dominant frequency band of this segment and the sleep stage of this segment. Report the dominant band using one of: delta, theta, alpha, beta, gamma, None. Report the sleep stage using one of: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the dominant frequency band of the specified segment\n2) the sleep stage of the specified segment\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"dominant_band\" and \"sleep_stage\".\n5. \"dominant_band\" must be exactly one of: \"delta\", \"theta\", \"alpha\", \"beta\", \"gamma\", or null.\n6. \"sleep_stage\" must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\", or null.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"dominant_band\": <\"delta\"|\"theta\"|\"alpha\"|\"beta\"|\"gamma\"|null>, \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "dominant_band_accuracy", + "type": "categorical_check", + "target_key": "dominant_band", + "weight": 50, + "params": { + "gt_value": "theta" + } + }, + { + "metric_id": "sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 50, + "params": { + "gt_value": "R" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case16/case16_15.json b/sleep_assessment/cases/case16/case16_15.json new file mode 100644 index 0000000000000000000000000000000000000000..597003093e033ce293c883b988cda2f2a15997a6 --- /dev/null +++ b/sleep_assessment/cases/case16/case16_15.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "instruction": "Please use the provided sleep file to analyze the segment from minute 118.5 to minute 121.5. Determine the dominant frequency band of this segment and the sleep stage of this segment. Report the dominant band using one of: delta, theta, alpha, beta, gamma, None. Report the sleep stage using one of: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the dominant frequency band of the specified segment\n2) the sleep stage of the specified segment\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"dominant_band\" and \"sleep_stage\".\n5. \"dominant_band\" must be exactly one of: \"delta\", \"theta\", \"alpha\", \"beta\", \"gamma\", or null.\n6. \"sleep_stage\" must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\", or null.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"dominant_band\": <\"delta\"|\"theta\"|\"alpha\"|\"beta\"|\"gamma\"|null>, \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "dominant_band_accuracy", + "type": "categorical_check", + "target_key": "dominant_band", + "weight": 50, + "params": { + "gt_value": "delta" + } + }, + { + "metric_id": "sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 50, + "params": { + "gt_value": "N3" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case16/case16_16.json b/sleep_assessment/cases/case16/case16_16.json new file mode 100644 index 0000000000000000000000000000000000000000..01f798110dade8e9e3745c01e5bcb8aa895ef444 --- /dev/null +++ b/sleep_assessment/cases/case16/case16_16.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "MASSSS3_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_01.edf", + "instruction": "Please use the provided sleep file to analyze the segment from minute 343.5 to minute 346.5. Determine the dominant frequency band of this segment and the sleep stage of this segment. Report the dominant band using one of: delta, theta, alpha, beta, gamma, None. Report the sleep stage using one of: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the dominant frequency band of the specified segment\n2) the sleep stage of the specified segment\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"dominant_band\" and \"sleep_stage\".\n5. \"dominant_band\" must be exactly one of: \"delta\", \"theta\", \"alpha\", \"beta\", \"gamma\", or null.\n6. \"sleep_stage\" must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\", or null.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"dominant_band\": <\"delta\"|\"theta\"|\"alpha\"|\"beta\"|\"gamma\"|null>, \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "dominant_band_accuracy", + "type": "categorical_check", + "target_key": "dominant_band", + "weight": 50, + "params": { + "gt_value": "delta" + } + }, + { + "metric_id": "sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 50, + "params": { + "gt_value": "N2" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case16/case16_17.json b/sleep_assessment/cases/case16/case16_17.json new file mode 100644 index 0000000000000000000000000000000000000000..ab8ebd90ca227bd7458b103aa53958ed850a2531 --- /dev/null +++ b/sleep_assessment/cases/case16/case16_17.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "MASSSS3_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_02.edf", + "instruction": "Please use the provided sleep file to analyze the segment from minute 407.5 to minute 410.5. Determine the dominant frequency band of this segment and the sleep stage of this segment. Report the dominant band using one of: delta, theta, alpha, beta, gamma, None. Report the sleep stage using one of: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the dominant frequency band of the specified segment\n2) the sleep stage of the specified segment\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"dominant_band\" and \"sleep_stage\".\n5. \"dominant_band\" must be exactly one of: \"delta\", \"theta\", \"alpha\", \"beta\", \"gamma\", or null.\n6. \"sleep_stage\" must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\", or null.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"dominant_band\": <\"delta\"|\"theta\"|\"alpha\"|\"beta\"|\"gamma\"|null>, \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "dominant_band_accuracy", + "type": "categorical_check", + "target_key": "dominant_band", + "weight": 50, + "params": { + "gt_value": "theta" + } + }, + { + "metric_id": "sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 50, + "params": { + "gt_value": "W" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case16/case16_18.json b/sleep_assessment/cases/case16/case16_18.json new file mode 100644 index 0000000000000000000000000000000000000000..a66444ec3cfbc86b92d0d1ad3c1eb03d733e238a --- /dev/null +++ b/sleep_assessment/cases/case16/case16_18.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "MASSSS3_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_03.edf", + "instruction": "Please use the provided sleep file to analyze the segment from minute 176.5 to minute 179.5. Determine the dominant frequency band of this segment and the sleep stage of this segment. Report the dominant band using one of: delta, theta, alpha, beta, gamma, None. Report the sleep stage using one of: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the dominant frequency band of the specified segment\n2) the sleep stage of the specified segment\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"dominant_band\" and \"sleep_stage\".\n5. \"dominant_band\" must be exactly one of: \"delta\", \"theta\", \"alpha\", \"beta\", \"gamma\", or null.\n6. \"sleep_stage\" must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\", or null.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"dominant_band\": <\"delta\"|\"theta\"|\"alpha\"|\"beta\"|\"gamma\"|null>, \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "dominant_band_accuracy", + "type": "categorical_check", + "target_key": "dominant_band", + "weight": 50, + "params": { + "gt_value": "delta" + } + }, + { + "metric_id": "sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 50, + "params": { + "gt_value": "R" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case16/case16_19.json b/sleep_assessment/cases/case16/case16_19.json new file mode 100644 index 0000000000000000000000000000000000000000..63f27e20fc546d08c2315e59416d1c5a7ac3645b --- /dev/null +++ b/sleep_assessment/cases/case16/case16_19.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "MASSSS3_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_04.edf", + "instruction": "Please use the provided sleep file to analyze the segment from minute 5.0 to minute 8.0. Determine the dominant frequency band of this segment and the sleep stage of this segment. Report the dominant band using one of: delta, theta, alpha, beta, gamma, None. Report the sleep stage using one of: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the dominant frequency band of the specified segment\n2) the sleep stage of the specified segment\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"dominant_band\" and \"sleep_stage\".\n5. \"dominant_band\" must be exactly one of: \"delta\", \"theta\", \"alpha\", \"beta\", \"gamma\", or null.\n6. \"sleep_stage\" must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\", or null.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"dominant_band\": <\"delta\"|\"theta\"|\"alpha\"|\"beta\"|\"gamma\"|null>, \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "dominant_band_accuracy", + "type": "categorical_check", + "target_key": "dominant_band", + "weight": 50, + "params": { + "gt_value": "delta" + } + }, + { + "metric_id": "sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 50, + "params": { + "gt_value": "N1" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case16/case16_20.json b/sleep_assessment/cases/case16/case16_20.json new file mode 100644 index 0000000000000000000000000000000000000000..0908358a0c5b773b0c7897caf14b28abef865766 --- /dev/null +++ b/sleep_assessment/cases/case16/case16_20.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "MASSSS3_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_05.edf", + "instruction": "Please use the provided sleep file to analyze the segment from minute 167.0 to minute 170.0. Determine the dominant frequency band of this segment and the sleep stage of this segment. Report the dominant band using one of: delta, theta, alpha, beta, gamma, None. Report the sleep stage using one of: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the dominant frequency band of the specified segment\n2) the sleep stage of the specified segment\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"dominant_band\" and \"sleep_stage\".\n5. \"dominant_band\" must be exactly one of: \"delta\", \"theta\", \"alpha\", \"beta\", \"gamma\", or null.\n6. \"sleep_stage\" must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\", or null.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"dominant_band\": <\"delta\"|\"theta\"|\"alpha\"|\"beta\"|\"gamma\"|null>, \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "dominant_band_accuracy", + "type": "categorical_check", + "target_key": "dominant_band", + "weight": 50, + "params": { + "gt_value": "delta" + } + }, + { + "metric_id": "sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 50, + "params": { + "gt_value": "N3" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case16/case16_21.json b/sleep_assessment/cases/case16/case16_21.json new file mode 100644 index 0000000000000000000000000000000000000000..eacfbdc50ecf11cfd78cd9cc5428fd6714b0cd0e --- /dev/null +++ b/sleep_assessment/cases/case16/case16_21.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "instruction": "Please use the provided sleep file to analyze the segment from minute 25.0 to minute 28.0. Determine the dominant frequency band of this segment and the sleep stage of this segment. Report the dominant band using one of: delta, theta, alpha, beta, gamma, None. Report the sleep stage using one of: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the dominant frequency band of the specified segment\n2) the sleep stage of the specified segment\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"dominant_band\" and \"sleep_stage\".\n5. \"dominant_band\" must be exactly one of: \"delta\", \"theta\", \"alpha\", \"beta\", \"gamma\", or null.\n6. \"sleep_stage\" must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\", or null.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"dominant_band\": <\"delta\"|\"theta\"|\"alpha\"|\"beta\"|\"gamma\"|null>, \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "dominant_band_accuracy", + "type": "categorical_check", + "target_key": "dominant_band", + "weight": 50, + "params": { + "gt_value": "delta" + } + }, + { + "metric_id": "sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 50, + "params": { + "gt_value": "N2" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case16/case16_22.json b/sleep_assessment/cases/case16/case16_22.json new file mode 100644 index 0000000000000000000000000000000000000000..c6fd6425b1474770aaad579fccd2a2e062496821 --- /dev/null +++ b/sleep_assessment/cases/case16/case16_22.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "instruction": "Please use the provided sleep file to analyze the segment from minute 225.5 to minute 228.5. Determine the dominant frequency band of this segment and the sleep stage of this segment. Report the dominant band using one of: delta, theta, alpha, beta, gamma, None. Report the sleep stage using one of: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the dominant frequency band of the specified segment\n2) the sleep stage of the specified segment\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"dominant_band\" and \"sleep_stage\".\n5. \"dominant_band\" must be exactly one of: \"delta\", \"theta\", \"alpha\", \"beta\", \"gamma\", or null.\n6. \"sleep_stage\" must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\", or null.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"dominant_band\": <\"delta\"|\"theta\"|\"alpha\"|\"beta\"|\"gamma\"|null>, \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "dominant_band_accuracy", + "type": "categorical_check", + "target_key": "dominant_band", + "weight": 50, + "params": { + "gt_value": "beta" + } + }, + { + "metric_id": "sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 50, + "params": { + "gt_value": "N1" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case16/case16_23.json b/sleep_assessment/cases/case16/case16_23.json new file mode 100644 index 0000000000000000000000000000000000000000..55df033003be34a2855ddd13ab4f1308843836cf --- /dev/null +++ b/sleep_assessment/cases/case16/case16_23.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "Physionet2018_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_03.edf", + "instruction": "Please use the provided sleep file to analyze the segment from minute 420.0 to minute 423.0. Determine the dominant frequency band of this segment and the sleep stage of this segment. Report the dominant band using one of: delta, theta, alpha, beta, gamma, None. Report the sleep stage using one of: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the dominant frequency band of the specified segment\n2) the sleep stage of the specified segment\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"dominant_band\" and \"sleep_stage\".\n5. \"dominant_band\" must be exactly one of: \"delta\", \"theta\", \"alpha\", \"beta\", \"gamma\", or null.\n6. \"sleep_stage\" must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\", or null.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"dominant_band\": <\"delta\"|\"theta\"|\"alpha\"|\"beta\"|\"gamma\"|null>, \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "dominant_band_accuracy", + "type": "categorical_check", + "target_key": "dominant_band", + "weight": 50, + "params": { + "gt_value": "beta" + } + }, + { + "metric_id": "sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 50, + "params": { + "gt_value": "W" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case16/case16_24.json b/sleep_assessment/cases/case16/case16_24.json new file mode 100644 index 0000000000000000000000000000000000000000..1594cb380c70ac58c361c8fc915ffd628346f7a3 --- /dev/null +++ b/sleep_assessment/cases/case16/case16_24.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "instruction": "Please use the provided sleep file to analyze the segment from minute 121.0 to minute 124.0. Determine the dominant frequency band of this segment and the sleep stage of this segment. Report the dominant band using one of: delta, theta, alpha, beta, gamma, None. Report the sleep stage using one of: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the dominant frequency band of the specified segment\n2) the sleep stage of the specified segment\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"dominant_band\" and \"sleep_stage\".\n5. \"dominant_band\" must be exactly one of: \"delta\", \"theta\", \"alpha\", \"beta\", \"gamma\", or null.\n6. \"sleep_stage\" must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\", or null.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"dominant_band\": <\"delta\"|\"theta\"|\"alpha\"|\"beta\"|\"gamma\"|null>, \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "dominant_band_accuracy", + "type": "categorical_check", + "target_key": "dominant_band", + "weight": 50, + "params": { + "gt_value": "delta" + } + }, + { + "metric_id": "sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 50, + "params": { + "gt_value": "N3" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case16/case16_25.json b/sleep_assessment/cases/case16/case16_25.json new file mode 100644 index 0000000000000000000000000000000000000000..970108ad41c061470014820c7a122ecad4adcd49 --- /dev/null +++ b/sleep_assessment/cases/case16/case16_25.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "instruction": "Please use the provided sleep file to analyze the segment from minute 370.5 to minute 373.5. Determine the dominant frequency band of this segment and the sleep stage of this segment. Report the dominant band using one of: delta, theta, alpha, beta, gamma, None. Report the sleep stage using one of: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the dominant frequency band of the specified segment\n2) the sleep stage of the specified segment\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"dominant_band\" and \"sleep_stage\".\n5. \"dominant_band\" must be exactly one of: \"delta\", \"theta\", \"alpha\", \"beta\", \"gamma\", or null.\n6. \"sleep_stage\" must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\", or null.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"dominant_band\": <\"delta\"|\"theta\"|\"alpha\"|\"beta\"|\"gamma\"|null>, \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "dominant_band_accuracy", + "type": "categorical_check", + "target_key": "dominant_band", + "weight": 50, + "params": { + "gt_value": "delta" + } + }, + { + "metric_id": "sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 50, + "params": { + "gt_value": "R" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case17/case17_01.json b/sleep_assessment/cases/case17/case17_01.json new file mode 100644 index 0000000000000000000000000000000000000000..ad1c88ce13f962528875705ee5b6a0477a0e6e9c --- /dev/null +++ b/sleep_assessment/cases/case17/case17_01.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "instruction": "Please use the provided sleep file to determine the sleep stages of the 10 consecutive epochs from epoch 638 to epoch 647. Return the corresponding 10-label sleep-stage sequence using only: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the 10-label sleep-stage sequence from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY one key named \"sleep_stage_sequence\".\n5. The value of \"sleep_stage_sequence\" must be a JSON array with EXACTLY 10 elements.\n6. Every element must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\".\n7. Do NOT output Chinese labels, integers, booleans, or null placeholders inside the list.\n8. If the report is unclear, still return the JSON key, but use an empty array only if no sequence can be extracted.\n\n### OUTPUT TEMPLATE\n{\"sleep_stage_sequence\": [\"W\", \"N1\", \"N2\", \"N3\", \"R\", \"W\", \"N1\", \"N2\", \"N3\", \"R\"]}", + "metrics": [ + { + "metric_id": "sleep_stage_sequence_accuracy", + "type": "sequence_match_check", + "target_key": "sleep_stage_sequence", + "weight": 100, + "params": { + "gt_value": [ + "N3", + "N2", + "N3", + "N3", + "N3", + "W", + "N1", + "N1", + "N2", + "N2" + ], + "match_mode": "per_hit_fraction_order" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case17/case17_02.json b/sleep_assessment/cases/case17/case17_02.json new file mode 100644 index 0000000000000000000000000000000000000000..641c9f30e7855cd08c8808334c131402ac31d4b0 --- /dev/null +++ b/sleep_assessment/cases/case17/case17_02.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "instruction": "Please use the provided sleep file to determine the sleep stages of the 10 consecutive epochs from epoch 495 to epoch 504. Return the corresponding 10-label sleep-stage sequence using only: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the 10-label sleep-stage sequence from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY one key named \"sleep_stage_sequence\".\n5. The value of \"sleep_stage_sequence\" must be a JSON array with EXACTLY 10 elements.\n6. Every element must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\".\n7. Do NOT output Chinese labels, integers, booleans, or null placeholders inside the list.\n8. If the report is unclear, still return the JSON key, but use an empty array only if no sequence can be extracted.\n\n### OUTPUT TEMPLATE\n{\"sleep_stage_sequence\": [\"W\", \"N1\", \"N2\", \"N3\", \"R\", \"W\", \"N1\", \"N2\", \"N3\", \"R\"]}", + "metrics": [ + { + "metric_id": "sleep_stage_sequence_accuracy", + "type": "sequence_match_check", + "target_key": "sleep_stage_sequence", + "weight": 100, + "params": { + "gt_value": [ + "R", + "W", + "W", + "N1", + "N2", + "N1", + "N2", + "N2", + "N1", + "N2" + ], + "match_mode": "per_hit_fraction_order" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case17/case17_03.json b/sleep_assessment/cases/case17/case17_03.json new file mode 100644 index 0000000000000000000000000000000000000000..c82f3228fb20ca1ccde8ba3fe35617184b385bcd --- /dev/null +++ b/sleep_assessment/cases/case17/case17_03.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "instruction": "Please use the provided sleep file to determine the sleep stages of the 10 consecutive epochs from epoch 849 to epoch 858. Return the corresponding 10-label sleep-stage sequence using only: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the 10-label sleep-stage sequence from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY one key named \"sleep_stage_sequence\".\n5. The value of \"sleep_stage_sequence\" must be a JSON array with EXACTLY 10 elements.\n6. Every element must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\".\n7. Do NOT output Chinese labels, integers, booleans, or null placeholders inside the list.\n8. If the report is unclear, still return the JSON key, but use an empty array only if no sequence can be extracted.\n\n### OUTPUT TEMPLATE\n{\"sleep_stage_sequence\": [\"W\", \"N1\", \"N2\", \"N3\", \"R\", \"W\", \"N1\", \"N2\", \"N3\", \"R\"]}", + "metrics": [ + { + "metric_id": "sleep_stage_sequence_accuracy", + "type": "sequence_match_check", + "target_key": "sleep_stage_sequence", + "weight": 100, + "params": { + "gt_value": [ + "N3", + "W", + "W", + "N1", + "N1", + "N1", + "N1", + "N2", + "W", + "N1" + ], + "match_mode": "per_hit_fraction_order" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case17/case17_04.json b/sleep_assessment/cases/case17/case17_04.json new file mode 100644 index 0000000000000000000000000000000000000000..9e53a19e50f8002a33742872e1593849de9dc6e9 --- /dev/null +++ b/sleep_assessment/cases/case17/case17_04.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "ISRUC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_04.edf", + "instruction": "Please use the provided sleep file to determine the sleep stages of the 10 consecutive epochs from epoch 82 to epoch 91. Return the corresponding 10-label sleep-stage sequence using only: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the 10-label sleep-stage sequence from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY one key named \"sleep_stage_sequence\".\n5. The value of \"sleep_stage_sequence\" must be a JSON array with EXACTLY 10 elements.\n6. Every element must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\".\n7. Do NOT output Chinese labels, integers, booleans, or null placeholders inside the list.\n8. If the report is unclear, still return the JSON key, but use an empty array only if no sequence can be extracted.\n\n### OUTPUT TEMPLATE\n{\"sleep_stage_sequence\": [\"W\", \"N1\", \"N2\", \"N3\", \"R\", \"W\", \"N1\", \"N2\", \"N3\", \"R\"]}", + "metrics": [ + { + "metric_id": "sleep_stage_sequence_accuracy", + "type": "sequence_match_check", + "target_key": "sleep_stage_sequence", + "weight": 100, + "params": { + "gt_value": [ + "N3", + "W", + "N1", + "N2", + "N2", + "N2", + "N2", + "N2", + "N2", + "N1" + ], + "match_mode": "per_hit_fraction_order" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case17/case17_05.json b/sleep_assessment/cases/case17/case17_05.json new file mode 100644 index 0000000000000000000000000000000000000000..0f2270f2ddfa1670a2f8294ad5313f38cfab92db --- /dev/null +++ b/sleep_assessment/cases/case17/case17_05.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "instruction": "Please use the provided sleep file to determine the sleep stages of the 10 consecutive epochs from epoch 599 to epoch 608. Return the corresponding 10-label sleep-stage sequence using only: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the 10-label sleep-stage sequence from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY one key named \"sleep_stage_sequence\".\n5. The value of \"sleep_stage_sequence\" must be a JSON array with EXACTLY 10 elements.\n6. Every element must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\".\n7. Do NOT output Chinese labels, integers, booleans, or null placeholders inside the list.\n8. If the report is unclear, still return the JSON key, but use an empty array only if no sequence can be extracted.\n\n### OUTPUT TEMPLATE\n{\"sleep_stage_sequence\": [\"W\", \"N1\", \"N2\", \"N3\", \"R\", \"W\", \"N1\", \"N2\", \"N3\", \"R\"]}", + "metrics": [ + { + "metric_id": "sleep_stage_sequence_accuracy", + "type": "sequence_match_check", + "target_key": "sleep_stage_sequence", + "weight": 100, + "params": { + "gt_value": [ + "N2", + "R", + "R", + "R", + "N1", + "N1", + "N1", + "N1", + "W", + "N1" + ], + "match_mode": "per_hit_fraction_order" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case17/case17_06.json b/sleep_assessment/cases/case17/case17_06.json new file mode 100644 index 0000000000000000000000000000000000000000..a465b60a4ab86cf881d11ba30118a7089570fd01 --- /dev/null +++ b/sleep_assessment/cases/case17/case17_06.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "HMC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_01.edf", + "instruction": "Please use the provided sleep file to determine the sleep stages of the 10 consecutive epochs from epoch 647 to epoch 656. Return the corresponding 10-label sleep-stage sequence using only: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the 10-label sleep-stage sequence from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY one key named \"sleep_stage_sequence\".\n5. The value of \"sleep_stage_sequence\" must be a JSON array with EXACTLY 10 elements.\n6. Every element must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\".\n7. Do NOT output Chinese labels, integers, booleans, or null placeholders inside the list.\n8. If the report is unclear, still return the JSON key, but use an empty array only if no sequence can be extracted.\n\n### OUTPUT TEMPLATE\n{\"sleep_stage_sequence\": [\"W\", \"N1\", \"N2\", \"N3\", \"R\", \"W\", \"N1\", \"N2\", \"N3\", \"R\"]}", + "metrics": [ + { + "metric_id": "sleep_stage_sequence_accuracy", + "type": "sequence_match_check", + "target_key": "sleep_stage_sequence", + "weight": 100, + "params": { + "gt_value": [ + "N3", + "N2", + "N3", + "N2", + "N2", + "W", + "W", + "W", + "N1", + "N2" + ], + "match_mode": "per_hit_fraction_order" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case17/case17_07.json b/sleep_assessment/cases/case17/case17_07.json new file mode 100644 index 0000000000000000000000000000000000000000..912ec14887323148aeee220ab07decee14fedbc5 --- /dev/null +++ b/sleep_assessment/cases/case17/case17_07.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "HMC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_02.edf", + "instruction": "Please use the provided sleep file to determine the sleep stages of the 10 consecutive epochs from epoch 5 to epoch 14. Return the corresponding 10-label sleep-stage sequence using only: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the 10-label sleep-stage sequence from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY one key named \"sleep_stage_sequence\".\n5. The value of \"sleep_stage_sequence\" must be a JSON array with EXACTLY 10 elements.\n6. Every element must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\".\n7. Do NOT output Chinese labels, integers, booleans, or null placeholders inside the list.\n8. If the report is unclear, still return the JSON key, but use an empty array only if no sequence can be extracted.\n\n### OUTPUT TEMPLATE\n{\"sleep_stage_sequence\": [\"W\", \"N1\", \"N2\", \"N3\", \"R\", \"W\", \"N1\", \"N2\", \"N3\", \"R\"]}", + "metrics": [ + { + "metric_id": "sleep_stage_sequence_accuracy", + "type": "sequence_match_check", + "target_key": "sleep_stage_sequence", + "weight": 100, + "params": { + "gt_value": [ + "W", + "N1", + "N1", + "N1", + "N2", + "N2", + "N2", + "N3", + "N2", + "N3" + ], + "match_mode": "per_hit_fraction_order" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case17/case17_08.json b/sleep_assessment/cases/case17/case17_08.json new file mode 100644 index 0000000000000000000000000000000000000000..85fc16a3c4dce2c7dce2e39a99b30d6e8bc82d2c --- /dev/null +++ b/sleep_assessment/cases/case17/case17_08.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "HMC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_03.edf", + "instruction": "Please use the provided sleep file to determine the sleep stages of the 10 consecutive epochs from epoch 81 to epoch 90. Return the corresponding 10-label sleep-stage sequence using only: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the 10-label sleep-stage sequence from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY one key named \"sleep_stage_sequence\".\n5. The value of \"sleep_stage_sequence\" must be a JSON array with EXACTLY 10 elements.\n6. Every element must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\".\n7. Do NOT output Chinese labels, integers, booleans, or null placeholders inside the list.\n8. If the report is unclear, still return the JSON key, but use an empty array only if no sequence can be extracted.\n\n### OUTPUT TEMPLATE\n{\"sleep_stage_sequence\": [\"W\", \"N1\", \"N2\", \"N3\", \"R\", \"W\", \"N1\", \"N2\", \"N3\", \"R\"]}", + "metrics": [ + { + "metric_id": "sleep_stage_sequence_accuracy", + "type": "sequence_match_check", + "target_key": "sleep_stage_sequence", + "weight": 100, + "params": { + "gt_value": [ + "N3", + "N3", + "N1", + "N2", + "N2", + "N3", + "N3", + "W", + "W", + "W" + ], + "match_mode": "per_hit_fraction_order" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case17/case17_09.json b/sleep_assessment/cases/case17/case17_09.json new file mode 100644 index 0000000000000000000000000000000000000000..8673051442b30a1a446cb638e3a2b02b3e0857b2 --- /dev/null +++ b/sleep_assessment/cases/case17/case17_09.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "HMC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_04.edf", + "instruction": "Please use the provided sleep file to determine the sleep stages of the 10 consecutive epochs from epoch 110 to epoch 119. Return the corresponding 10-label sleep-stage sequence using only: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the 10-label sleep-stage sequence from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY one key named \"sleep_stage_sequence\".\n5. The value of \"sleep_stage_sequence\" must be a JSON array with EXACTLY 10 elements.\n6. Every element must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\".\n7. Do NOT output Chinese labels, integers, booleans, or null placeholders inside the list.\n8. If the report is unclear, still return the JSON key, but use an empty array only if no sequence can be extracted.\n\n### OUTPUT TEMPLATE\n{\"sleep_stage_sequence\": [\"W\", \"N1\", \"N2\", \"N3\", \"R\", \"W\", \"N1\", \"N2\", \"N3\", \"R\"]}", + "metrics": [ + { + "metric_id": "sleep_stage_sequence_accuracy", + "type": "sequence_match_check", + "target_key": "sleep_stage_sequence", + "weight": 100, + "params": { + "gt_value": [ + "R", + "R", + "R", + "W", + "W", + "W", + "N1", + "N1", + "N2", + "N2" + ], + "match_mode": "per_hit_fraction_order" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case17/case17_10.json b/sleep_assessment/cases/case17/case17_10.json new file mode 100644 index 0000000000000000000000000000000000000000..6f75fe66b3edabdb97f53d6f7f09e84ddf347db0 --- /dev/null +++ b/sleep_assessment/cases/case17/case17_10.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "HMC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_05.edf", + "instruction": "Please use the provided sleep file to determine the sleep stages of the 10 consecutive epochs from epoch 926 to epoch 935. Return the corresponding 10-label sleep-stage sequence using only: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the 10-label sleep-stage sequence from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY one key named \"sleep_stage_sequence\".\n5. The value of \"sleep_stage_sequence\" must be a JSON array with EXACTLY 10 elements.\n6. Every element must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\".\n7. Do NOT output Chinese labels, integers, booleans, or null placeholders inside the list.\n8. If the report is unclear, still return the JSON key, but use an empty array only if no sequence can be extracted.\n\n### OUTPUT TEMPLATE\n{\"sleep_stage_sequence\": [\"W\", \"N1\", \"N2\", \"N3\", \"R\", \"W\", \"N1\", \"N2\", \"N3\", \"R\"]}", + "metrics": [ + { + "metric_id": "sleep_stage_sequence_accuracy", + "type": "sequence_match_check", + "target_key": "sleep_stage_sequence", + "weight": 100, + "params": { + "gt_value": [ + "R", + "N1", + "N2", + "R", + "R", + "W", + "N1", + "N1", + "N2", + "N1" + ], + "match_mode": "per_hit_fraction_order" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case17/case17_11.json b/sleep_assessment/cases/case17/case17_11.json new file mode 100644 index 0000000000000000000000000000000000000000..7686c84e77aa427037880d13da89aaa895f91efc --- /dev/null +++ b/sleep_assessment/cases/case17/case17_11.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "instruction": "Please use the provided sleep file to determine the sleep stages of the 10 consecutive epochs from epoch 452 to epoch 461. Return the corresponding 10-label sleep-stage sequence using only: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the 10-label sleep-stage sequence from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY one key named \"sleep_stage_sequence\".\n5. The value of \"sleep_stage_sequence\" must be a JSON array with EXACTLY 10 elements.\n6. Every element must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\".\n7. Do NOT output Chinese labels, integers, booleans, or null placeholders inside the list.\n8. If the report is unclear, still return the JSON key, but use an empty array only if no sequence can be extracted.\n\n### OUTPUT TEMPLATE\n{\"sleep_stage_sequence\": [\"W\", \"N1\", \"N2\", \"N3\", \"R\", \"W\", \"N1\", \"N2\", \"N3\", \"R\"]}", + "metrics": [ + { + "metric_id": "sleep_stage_sequence_accuracy", + "type": "sequence_match_check", + "target_key": "sleep_stage_sequence", + "weight": 100, + "params": { + "gt_value": [ + "N2", + "N3", + "N2", + "N2", + "N2", + "N2", + "W", + "N1", + "N2", + "R" + ], + "match_mode": "per_hit_fraction_order" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case17/case17_12.json b/sleep_assessment/cases/case17/case17_12.json new file mode 100644 index 0000000000000000000000000000000000000000..48231f1950b1357dde976f858e46688fa2e826cb --- /dev/null +++ b/sleep_assessment/cases/case17/case17_12.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "instruction": "Please use the provided sleep file to determine the sleep stages of the 10 consecutive epochs from epoch 29 to epoch 38. Return the corresponding 10-label sleep-stage sequence using only: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the 10-label sleep-stage sequence from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY one key named \"sleep_stage_sequence\".\n5. The value of \"sleep_stage_sequence\" must be a JSON array with EXACTLY 10 elements.\n6. Every element must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\".\n7. Do NOT output Chinese labels, integers, booleans, or null placeholders inside the list.\n8. If the report is unclear, still return the JSON key, but use an empty array only if no sequence can be extracted.\n\n### OUTPUT TEMPLATE\n{\"sleep_stage_sequence\": [\"W\", \"N1\", \"N2\", \"N3\", \"R\", \"W\", \"N1\", \"N2\", \"N3\", \"R\"]}", + "metrics": [ + { + "metric_id": "sleep_stage_sequence_accuracy", + "type": "sequence_match_check", + "target_key": "sleep_stage_sequence", + "weight": 100, + "params": { + "gt_value": [ + "N2", + "N3", + "N2", + "N2", + "N3", + "N2", + "N3", + "N2", + "W", + "N2" + ], + "match_mode": "per_hit_fraction_order" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case17/case17_13.json b/sleep_assessment/cases/case17/case17_13.json new file mode 100644 index 0000000000000000000000000000000000000000..7ef07c8d2f3524b1d27d4613c8c7a6ad1c498e8b --- /dev/null +++ b/sleep_assessment/cases/case17/case17_13.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "instruction": "Please use the provided sleep file to determine the sleep stages of the 10 consecutive epochs from epoch 532 to epoch 541. Return the corresponding 10-label sleep-stage sequence using only: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the 10-label sleep-stage sequence from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY one key named \"sleep_stage_sequence\".\n5. The value of \"sleep_stage_sequence\" must be a JSON array with EXACTLY 10 elements.\n6. Every element must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\".\n7. Do NOT output Chinese labels, integers, booleans, or null placeholders inside the list.\n8. If the report is unclear, still return the JSON key, but use an empty array only if no sequence can be extracted.\n\n### OUTPUT TEMPLATE\n{\"sleep_stage_sequence\": [\"W\", \"N1\", \"N2\", \"N3\", \"R\", \"W\", \"N1\", \"N2\", \"N3\", \"R\"]}", + "metrics": [ + { + "metric_id": "sleep_stage_sequence_accuracy", + "type": "sequence_match_check", + "target_key": "sleep_stage_sequence", + "weight": 100, + "params": { + "gt_value": [ + "R", + "R", + "N2", + "W", + "N2", + "W", + "W", + "N1", + "N1", + "N2" + ], + "match_mode": "per_hit_fraction_order" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case17/case17_14.json b/sleep_assessment/cases/case17/case17_14.json new file mode 100644 index 0000000000000000000000000000000000000000..49b105b151dea962db8736fae918beb2e8e1a61d --- /dev/null +++ b/sleep_assessment/cases/case17/case17_14.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "instruction": "Please use the provided sleep file to determine the sleep stages of the 10 consecutive epochs from epoch 120 to epoch 129. Return the corresponding 10-label sleep-stage sequence using only: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the 10-label sleep-stage sequence from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY one key named \"sleep_stage_sequence\".\n5. The value of \"sleep_stage_sequence\" must be a JSON array with EXACTLY 10 elements.\n6. Every element must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\".\n7. Do NOT output Chinese labels, integers, booleans, or null placeholders inside the list.\n8. If the report is unclear, still return the JSON key, but use an empty array only if no sequence can be extracted.\n\n### OUTPUT TEMPLATE\n{\"sleep_stage_sequence\": [\"W\", \"N1\", \"N2\", \"N3\", \"R\", \"W\", \"N1\", \"N2\", \"N3\", \"R\"]}", + "metrics": [ + { + "metric_id": "sleep_stage_sequence_accuracy", + "type": "sequence_match_check", + "target_key": "sleep_stage_sequence", + "weight": 100, + "params": { + "gt_value": [ + "N3", + "N3", + "N2", + "N3", + "W", + "N1", + "N2", + "N3", + "N2", + "N3" + ], + "match_mode": "per_hit_fraction_order" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case17/case17_15.json b/sleep_assessment/cases/case17/case17_15.json new file mode 100644 index 0000000000000000000000000000000000000000..9d6e124016b08763f037642ddfd13b5f6f3095be --- /dev/null +++ b/sleep_assessment/cases/case17/case17_15.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "instruction": "Please use the provided sleep file to determine the sleep stages of the 10 consecutive epochs from epoch 157 to epoch 166. Return the corresponding 10-label sleep-stage sequence using only: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the 10-label sleep-stage sequence from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY one key named \"sleep_stage_sequence\".\n5. The value of \"sleep_stage_sequence\" must be a JSON array with EXACTLY 10 elements.\n6. Every element must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\".\n7. Do NOT output Chinese labels, integers, booleans, or null placeholders inside the list.\n8. If the report is unclear, still return the JSON key, but use an empty array only if no sequence can be extracted.\n\n### OUTPUT TEMPLATE\n{\"sleep_stage_sequence\": [\"W\", \"N1\", \"N2\", \"N3\", \"R\", \"W\", \"N1\", \"N2\", \"N3\", \"R\"]}", + "metrics": [ + { + "metric_id": "sleep_stage_sequence_accuracy", + "type": "sequence_match_check", + "target_key": "sleep_stage_sequence", + "weight": 100, + "params": { + "gt_value": [ + "N3", + "N2", + "N3", + "N3", + "N3", + "W", + "W", + "N1", + "W", + "W" + ], + "match_mode": "per_hit_fraction_order" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case17/case17_16.json b/sleep_assessment/cases/case17/case17_16.json new file mode 100644 index 0000000000000000000000000000000000000000..693daf2fc58ea6f7cc8d4af35f02b1e15882469a --- /dev/null +++ b/sleep_assessment/cases/case17/case17_16.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "MASSSS3_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_01.edf", + "instruction": "Please use the provided sleep file to determine the sleep stages of the 10 consecutive epochs from epoch 5 to epoch 14. Return the corresponding 10-label sleep-stage sequence using only: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the 10-label sleep-stage sequence from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY one key named \"sleep_stage_sequence\".\n5. The value of \"sleep_stage_sequence\" must be a JSON array with EXACTLY 10 elements.\n6. Every element must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\".\n7. Do NOT output Chinese labels, integers, booleans, or null placeholders inside the list.\n8. If the report is unclear, still return the JSON key, but use an empty array only if no sequence can be extracted.\n\n### OUTPUT TEMPLATE\n{\"sleep_stage_sequence\": [\"W\", \"N1\", \"N2\", \"N3\", \"R\", \"W\", \"N1\", \"N2\", \"N3\", \"R\"]}", + "metrics": [ + { + "metric_id": "sleep_stage_sequence_accuracy", + "type": "sequence_match_check", + "target_key": "sleep_stage_sequence", + "weight": 100, + "params": { + "gt_value": [ + "W", + "W", + "N1", + "N1", + "N1", + "N1", + "N2", + "N2", + "N3", + "N2" + ], + "match_mode": "per_hit_fraction_order" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case17/case17_17.json b/sleep_assessment/cases/case17/case17_17.json new file mode 100644 index 0000000000000000000000000000000000000000..d18aefe119009187a92ee502393c3d4266470e58 --- /dev/null +++ b/sleep_assessment/cases/case17/case17_17.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "MASSSS3_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_02.edf", + "instruction": "Please use the provided sleep file to determine the sleep stages of the 10 consecutive epochs from epoch 192 to epoch 201. Return the corresponding 10-label sleep-stage sequence using only: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the 10-label sleep-stage sequence from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY one key named \"sleep_stage_sequence\".\n5. The value of \"sleep_stage_sequence\" must be a JSON array with EXACTLY 10 elements.\n6. Every element must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\".\n7. Do NOT output Chinese labels, integers, booleans, or null placeholders inside the list.\n8. If the report is unclear, still return the JSON key, but use an empty array only if no sequence can be extracted.\n\n### OUTPUT TEMPLATE\n{\"sleep_stage_sequence\": [\"W\", \"N1\", \"N2\", \"N3\", \"R\", \"W\", \"N1\", \"N2\", \"N3\", \"R\"]}", + "metrics": [ + { + "metric_id": "sleep_stage_sequence_accuracy", + "type": "sequence_match_check", + "target_key": "sleep_stage_sequence", + "weight": 100, + "params": { + "gt_value": [ + "N2", + "N3", + "N2", + "N3", + "W", + "W", + "N1", + "W", + "N1", + "N2" + ], + "match_mode": "per_hit_fraction_order" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case17/case17_18.json b/sleep_assessment/cases/case17/case17_18.json new file mode 100644 index 0000000000000000000000000000000000000000..2765d1786f1dc6a890a0504ab76e29671eeb0be4 --- /dev/null +++ b/sleep_assessment/cases/case17/case17_18.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "MASSSS3_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_03.edf", + "instruction": "Please use the provided sleep file to determine the sleep stages of the 10 consecutive epochs from epoch 745 to epoch 754. Return the corresponding 10-label sleep-stage sequence using only: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the 10-label sleep-stage sequence from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY one key named \"sleep_stage_sequence\".\n5. The value of \"sleep_stage_sequence\" must be a JSON array with EXACTLY 10 elements.\n6. Every element must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\".\n7. Do NOT output Chinese labels, integers, booleans, or null placeholders inside the list.\n8. If the report is unclear, still return the JSON key, but use an empty array only if no sequence can be extracted.\n\n### OUTPUT TEMPLATE\n{\"sleep_stage_sequence\": [\"W\", \"N1\", \"N2\", \"N3\", \"R\", \"W\", \"N1\", \"N2\", \"N3\", \"R\"]}", + "metrics": [ + { + "metric_id": "sleep_stage_sequence_accuracy", + "type": "sequence_match_check", + "target_key": "sleep_stage_sequence", + "weight": 100, + "params": { + "gt_value": [ + "R", + "R", + "W", + "N1", + "N1", + "N2", + "N2", + "N2", + "N2", + "W" + ], + "match_mode": "per_hit_fraction_order" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case17/case17_19.json b/sleep_assessment/cases/case17/case17_19.json new file mode 100644 index 0000000000000000000000000000000000000000..1ff0ce50b2434c787efe78d446b514565bbf35fb --- /dev/null +++ b/sleep_assessment/cases/case17/case17_19.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "MASSSS3_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_04.edf", + "instruction": "Please use the provided sleep file to determine the sleep stages of the 10 consecutive epochs from epoch 489 to epoch 498. Return the corresponding 10-label sleep-stage sequence using only: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the 10-label sleep-stage sequence from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY one key named \"sleep_stage_sequence\".\n5. The value of \"sleep_stage_sequence\" must be a JSON array with EXACTLY 10 elements.\n6. Every element must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\".\n7. Do NOT output Chinese labels, integers, booleans, or null placeholders inside the list.\n8. If the report is unclear, still return the JSON key, but use an empty array only if no sequence can be extracted.\n\n### OUTPUT TEMPLATE\n{\"sleep_stage_sequence\": [\"W\", \"N1\", \"N2\", \"N3\", \"R\", \"W\", \"N1\", \"N2\", \"N3\", \"R\"]}", + "metrics": [ + { + "metric_id": "sleep_stage_sequence_accuracy", + "type": "sequence_match_check", + "target_key": "sleep_stage_sequence", + "weight": 100, + "params": { + "gt_value": [ + "N1", + "N1", + "W", + "N1", + "N2", + "N2", + "N2", + "N2", + "N2", + "R" + ], + "match_mode": "per_hit_fraction_order" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case17/case17_20.json b/sleep_assessment/cases/case17/case17_20.json new file mode 100644 index 0000000000000000000000000000000000000000..74a54e5f2e71f988faf609b704712bfee894b2c6 --- /dev/null +++ b/sleep_assessment/cases/case17/case17_20.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "MASSSS3_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_05.edf", + "instruction": "Please use the provided sleep file to determine the sleep stages of the 10 consecutive epochs from epoch 711 to epoch 720. Return the corresponding 10-label sleep-stage sequence using only: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the 10-label sleep-stage sequence from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY one key named \"sleep_stage_sequence\".\n5. The value of \"sleep_stage_sequence\" must be a JSON array with EXACTLY 10 elements.\n6. Every element must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\".\n7. Do NOT output Chinese labels, integers, booleans, or null placeholders inside the list.\n8. If the report is unclear, still return the JSON key, but use an empty array only if no sequence can be extracted.\n\n### OUTPUT TEMPLATE\n{\"sleep_stage_sequence\": [\"W\", \"N1\", \"N2\", \"N3\", \"R\", \"W\", \"N1\", \"N2\", \"N3\", \"R\"]}", + "metrics": [ + { + "metric_id": "sleep_stage_sequence_accuracy", + "type": "sequence_match_check", + "target_key": "sleep_stage_sequence", + "weight": 100, + "params": { + "gt_value": [ + "N2", + "N3", + "N3", + "N2", + "N2", + "N3", + "N2", + "N3", + "W", + "N1" + ], + "match_mode": "per_hit_fraction_order" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case17/case17_21.json b/sleep_assessment/cases/case17/case17_21.json new file mode 100644 index 0000000000000000000000000000000000000000..50bf833683bd4e5109022db44f4a8c0253989df8 --- /dev/null +++ b/sleep_assessment/cases/case17/case17_21.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "instruction": "Please use the provided sleep file to determine the sleep stages of the 10 consecutive epochs from epoch 454 to epoch 463. Return the corresponding 10-label sleep-stage sequence using only: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the 10-label sleep-stage sequence from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY one key named \"sleep_stage_sequence\".\n5. The value of \"sleep_stage_sequence\" must be a JSON array with EXACTLY 10 elements.\n6. Every element must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\".\n7. Do NOT output Chinese labels, integers, booleans, or null placeholders inside the list.\n8. If the report is unclear, still return the JSON key, but use an empty array only if no sequence can be extracted.\n\n### OUTPUT TEMPLATE\n{\"sleep_stage_sequence\": [\"W\", \"N1\", \"N2\", \"N3\", \"R\", \"W\", \"N1\", \"N2\", \"N3\", \"R\"]}", + "metrics": [ + { + "metric_id": "sleep_stage_sequence_accuracy", + "type": "sequence_match_check", + "target_key": "sleep_stage_sequence", + "weight": 100, + "params": { + "gt_value": [ + "R", + "R", + "R", + "W", + "N1", + "N2", + "N2", + "N2", + "N1", + "N2" + ], + "match_mode": "per_hit_fraction_order" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case17/case17_22.json b/sleep_assessment/cases/case17/case17_22.json new file mode 100644 index 0000000000000000000000000000000000000000..3e3c1784dc842497259afedf72a752b3ed552c62 --- /dev/null +++ b/sleep_assessment/cases/case17/case17_22.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "instruction": "Please use the provided sleep file to determine the sleep stages of the 10 consecutive epochs from epoch 72 to epoch 81. Return the corresponding 10-label sleep-stage sequence using only: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the 10-label sleep-stage sequence from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY one key named \"sleep_stage_sequence\".\n5. The value of \"sleep_stage_sequence\" must be a JSON array with EXACTLY 10 elements.\n6. Every element must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\".\n7. Do NOT output Chinese labels, integers, booleans, or null placeholders inside the list.\n8. If the report is unclear, still return the JSON key, but use an empty array only if no sequence can be extracted.\n\n### OUTPUT TEMPLATE\n{\"sleep_stage_sequence\": [\"W\", \"N1\", \"N2\", \"N3\", \"R\", \"W\", \"N1\", \"N2\", \"N3\", \"R\"]}", + "metrics": [ + { + "metric_id": "sleep_stage_sequence_accuracy", + "type": "sequence_match_check", + "target_key": "sleep_stage_sequence", + "weight": 100, + "params": { + "gt_value": [ + "N2", + "N1", + "N1", + "N2", + "N2", + "W", + "N1", + "N2", + "N1", + "N2" + ], + "match_mode": "per_hit_fraction_order" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case17/case17_23.json b/sleep_assessment/cases/case17/case17_23.json new file mode 100644 index 0000000000000000000000000000000000000000..3f9f03aa2c6d23aed8cc6479c36d88426fe96141 --- /dev/null +++ b/sleep_assessment/cases/case17/case17_23.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "Physionet2018_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_03.edf", + "instruction": "Please use the provided sleep file to determine the sleep stages of the 10 consecutive epochs from epoch 242 to epoch 251. Return the corresponding 10-label sleep-stage sequence using only: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the 10-label sleep-stage sequence from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY one key named \"sleep_stage_sequence\".\n5. The value of \"sleep_stage_sequence\" must be a JSON array with EXACTLY 10 elements.\n6. Every element must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\".\n7. Do NOT output Chinese labels, integers, booleans, or null placeholders inside the list.\n8. If the report is unclear, still return the JSON key, but use an empty array only if no sequence can be extracted.\n\n### OUTPUT TEMPLATE\n{\"sleep_stage_sequence\": [\"W\", \"N1\", \"N2\", \"N3\", \"R\", \"W\", \"N1\", \"N2\", \"N3\", \"R\"]}", + "metrics": [ + { + "metric_id": "sleep_stage_sequence_accuracy", + "type": "sequence_match_check", + "target_key": "sleep_stage_sequence", + "weight": 100, + "params": { + "gt_value": [ + "N2", + "N2", + "N3", + "N2", + "N3", + "N2", + "W", + "N1", + "W", + "W" + ], + "match_mode": "per_hit_fraction_order" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case17/case17_24.json b/sleep_assessment/cases/case17/case17_24.json new file mode 100644 index 0000000000000000000000000000000000000000..9d49282f0d387aee97ead5cbd3b600ae28487df2 --- /dev/null +++ b/sleep_assessment/cases/case17/case17_24.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "instruction": "Please use the provided sleep file to determine the sleep stages of the 10 consecutive epochs from epoch 728 to epoch 737. Return the corresponding 10-label sleep-stage sequence using only: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the 10-label sleep-stage sequence from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY one key named \"sleep_stage_sequence\".\n5. The value of \"sleep_stage_sequence\" must be a JSON array with EXACTLY 10 elements.\n6. Every element must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\".\n7. Do NOT output Chinese labels, integers, booleans, or null placeholders inside the list.\n8. If the report is unclear, still return the JSON key, but use an empty array only if no sequence can be extracted.\n\n### OUTPUT TEMPLATE\n{\"sleep_stage_sequence\": [\"W\", \"N1\", \"N2\", \"N3\", \"R\", \"W\", \"N1\", \"N2\", \"N3\", \"R\"]}", + "metrics": [ + { + "metric_id": "sleep_stage_sequence_accuracy", + "type": "sequence_match_check", + "target_key": "sleep_stage_sequence", + "weight": 100, + "params": { + "gt_value": [ + "R", + "N1", + "N1", + "R", + "R", + "W", + "N1", + "N1", + "N2", + "N1" + ], + "match_mode": "per_hit_fraction_order" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case17/case17_25.json b/sleep_assessment/cases/case17/case17_25.json new file mode 100644 index 0000000000000000000000000000000000000000..261d7d1e063dcb959b98be8643e10542a44fb8e0 --- /dev/null +++ b/sleep_assessment/cases/case17/case17_25.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "instruction": "Please use the provided sleep file to determine the sleep stages of the 10 consecutive epochs from epoch 78 to epoch 87. Return the corresponding 10-label sleep-stage sequence using only: W, N1, N2, N3, R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the 10-label sleep-stage sequence from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON must contain EXACTLY one key named \"sleep_stage_sequence\".\n5. The value of \"sleep_stage_sequence\" must be a JSON array with EXACTLY 10 elements.\n6. Every element must be exactly one of: \"W\", \"N1\", \"N2\", \"N3\", \"R\".\n7. Do NOT output Chinese labels, integers, booleans, or null placeholders inside the list.\n8. If the report is unclear, still return the JSON key, but use an empty array only if no sequence can be extracted.\n\n### OUTPUT TEMPLATE\n{\"sleep_stage_sequence\": [\"W\", \"N1\", \"N2\", \"N3\", \"R\", \"W\", \"N1\", \"N2\", \"N3\", \"R\"]}", + "metrics": [ + { + "metric_id": "sleep_stage_sequence_accuracy", + "type": "sequence_match_check", + "target_key": "sleep_stage_sequence", + "weight": 100, + "params": { + "gt_value": [ + "N3", + "N3", + "N3", + "N3", + "N2", + "N2", + "N1", + "N2", + "N1", + "N1" + ], + "match_mode": "per_hit_fraction_order" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case18/case18_01.json b/sleep_assessment/cases/case18/case18_01.json new file mode 100644 index 0000000000000000000000000000000000000000..f9b40a390df96fcf1f54f9995277716c72ea7dcc --- /dev/null +++ b/sleep_assessment/cases/case18/case18_01.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "instruction": "Please use the provided sleep file to generate a sleep hypnogram, and save the image to the file_check directory with the filename ISRUC_01_case18_sleep_hypnogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep hypnogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_hypnogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_hypnogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "ISRUC_01_case18_sleep_hypnogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep hypnogram plotting task. Check the image only. The task requirement is to generate a sleep hypnogram, also known as a sleep-stage-over-time plot, from a raw sleep recording. Return pass only if the image is clearly a valid sleep hypnogram-style figure. Specifically verify: (1) the image shows sleep stages over time, epochs, or an overnight recording timeline; (2) the y-axis or visible labels represent discrete sleep stages such as W/Wake, N1, N2, N3, and REM/R; (3) the plotted content is a staircase line, stage trajectory, or equivalent discrete sleep-stage time series rather than continuous raw EEG amplitude; (4) the figure is not a PSD plot, spectrogram, heatmap, correlation matrix, bar chart, topomap, text report screenshot, or unrelated medical image; (5) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge whether the sleep-stage sequence is accurate, whether it matches ground-truth labels, or whether the clinical interpretation is correct. Do not require a specific color scheme or exact stage proportions. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case18/case18_02.json b/sleep_assessment/cases/case18/case18_02.json new file mode 100644 index 0000000000000000000000000000000000000000..22fbcc09a740066810f63a347969c8cd36a053dd --- /dev/null +++ b/sleep_assessment/cases/case18/case18_02.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "instruction": "Please use the provided sleep file to generate a sleep hypnogram, and save the image to the file_check directory with the filename ISRUC_02_case18_sleep_hypnogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep hypnogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_hypnogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_hypnogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "ISRUC_02_case18_sleep_hypnogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep hypnogram plotting task. Check the image only. The task requirement is to generate a sleep hypnogram, also known as a sleep-stage-over-time plot, from a raw sleep recording. Return pass only if the image is clearly a valid sleep hypnogram-style figure. Specifically verify: (1) the image shows sleep stages over time, epochs, or an overnight recording timeline; (2) the y-axis or visible labels represent discrete sleep stages such as W/Wake, N1, N2, N3, and REM/R; (3) the plotted content is a staircase line, stage trajectory, or equivalent discrete sleep-stage time series rather than continuous raw EEG amplitude; (4) the figure is not a PSD plot, spectrogram, heatmap, correlation matrix, bar chart, topomap, text report screenshot, or unrelated medical image; (5) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge whether the sleep-stage sequence is accurate, whether it matches ground-truth labels, or whether the clinical interpretation is correct. Do not require a specific color scheme or exact stage proportions. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case18/case18_03.json b/sleep_assessment/cases/case18/case18_03.json new file mode 100644 index 0000000000000000000000000000000000000000..8fe9f64c6b9beff9f0dcf2d0a3dcee7c409918f6 --- /dev/null +++ b/sleep_assessment/cases/case18/case18_03.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "instruction": "Please use the provided sleep file to generate a sleep hypnogram, and save the image to the file_check directory with the filename ISRUC_03_case18_sleep_hypnogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep hypnogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_hypnogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_hypnogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "ISRUC_03_case18_sleep_hypnogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep hypnogram plotting task. Check the image only. The task requirement is to generate a sleep hypnogram, also known as a sleep-stage-over-time plot, from a raw sleep recording. Return pass only if the image is clearly a valid sleep hypnogram-style figure. Specifically verify: (1) the image shows sleep stages over time, epochs, or an overnight recording timeline; (2) the y-axis or visible labels represent discrete sleep stages such as W/Wake, N1, N2, N3, and REM/R; (3) the plotted content is a staircase line, stage trajectory, or equivalent discrete sleep-stage time series rather than continuous raw EEG amplitude; (4) the figure is not a PSD plot, spectrogram, heatmap, correlation matrix, bar chart, topomap, text report screenshot, or unrelated medical image; (5) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge whether the sleep-stage sequence is accurate, whether it matches ground-truth labels, or whether the clinical interpretation is correct. Do not require a specific color scheme or exact stage proportions. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case18/case18_04.json b/sleep_assessment/cases/case18/case18_04.json new file mode 100644 index 0000000000000000000000000000000000000000..ac42aed3fb28a21d3cfe7e4d81f05c987249f1e4 --- /dev/null +++ b/sleep_assessment/cases/case18/case18_04.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "ISRUC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_04.edf", + "instruction": "Please use the provided sleep file to generate a sleep hypnogram, and save the image to the file_check directory with the filename ISRUC_04_case18_sleep_hypnogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep hypnogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_hypnogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_hypnogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "ISRUC_04_case18_sleep_hypnogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep hypnogram plotting task. Check the image only. The task requirement is to generate a sleep hypnogram, also known as a sleep-stage-over-time plot, from a raw sleep recording. Return pass only if the image is clearly a valid sleep hypnogram-style figure. Specifically verify: (1) the image shows sleep stages over time, epochs, or an overnight recording timeline; (2) the y-axis or visible labels represent discrete sleep stages such as W/Wake, N1, N2, N3, and REM/R; (3) the plotted content is a staircase line, stage trajectory, or equivalent discrete sleep-stage time series rather than continuous raw EEG amplitude; (4) the figure is not a PSD plot, spectrogram, heatmap, correlation matrix, bar chart, topomap, text report screenshot, or unrelated medical image; (5) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge whether the sleep-stage sequence is accurate, whether it matches ground-truth labels, or whether the clinical interpretation is correct. Do not require a specific color scheme or exact stage proportions. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case18/case18_05.json b/sleep_assessment/cases/case18/case18_05.json new file mode 100644 index 0000000000000000000000000000000000000000..a2036b104997f4a22f9fb6c55ed9766e61f5f129 --- /dev/null +++ b/sleep_assessment/cases/case18/case18_05.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "instruction": "Please use the provided sleep file to generate a sleep hypnogram, and save the image to the file_check directory with the filename ISRUC_05_case18_sleep_hypnogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep hypnogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_hypnogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_hypnogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "ISRUC_05_case18_sleep_hypnogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep hypnogram plotting task. Check the image only. The task requirement is to generate a sleep hypnogram, also known as a sleep-stage-over-time plot, from a raw sleep recording. Return pass only if the image is clearly a valid sleep hypnogram-style figure. Specifically verify: (1) the image shows sleep stages over time, epochs, or an overnight recording timeline; (2) the y-axis or visible labels represent discrete sleep stages such as W/Wake, N1, N2, N3, and REM/R; (3) the plotted content is a staircase line, stage trajectory, or equivalent discrete sleep-stage time series rather than continuous raw EEG amplitude; (4) the figure is not a PSD plot, spectrogram, heatmap, correlation matrix, bar chart, topomap, text report screenshot, or unrelated medical image; (5) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge whether the sleep-stage sequence is accurate, whether it matches ground-truth labels, or whether the clinical interpretation is correct. Do not require a specific color scheme or exact stage proportions. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case18/case18_06.json b/sleep_assessment/cases/case18/case18_06.json new file mode 100644 index 0000000000000000000000000000000000000000..c104e43b0d5db1a16f7c4b7331619ba891a23405 --- /dev/null +++ b/sleep_assessment/cases/case18/case18_06.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "HMC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_01.edf", + "instruction": "Please use the provided sleep file to generate a sleep hypnogram, and save the image to the file_check directory with the filename HMC_01_case18_sleep_hypnogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep hypnogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_hypnogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_hypnogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "HMC_01_case18_sleep_hypnogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep hypnogram plotting task. Check the image only. The task requirement is to generate a sleep hypnogram, also known as a sleep-stage-over-time plot, from a raw sleep recording. Return pass only if the image is clearly a valid sleep hypnogram-style figure. Specifically verify: (1) the image shows sleep stages over time, epochs, or an overnight recording timeline; (2) the y-axis or visible labels represent discrete sleep stages such as W/Wake, N1, N2, N3, and REM/R; (3) the plotted content is a staircase line, stage trajectory, or equivalent discrete sleep-stage time series rather than continuous raw EEG amplitude; (4) the figure is not a PSD plot, spectrogram, heatmap, correlation matrix, bar chart, topomap, text report screenshot, or unrelated medical image; (5) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge whether the sleep-stage sequence is accurate, whether it matches ground-truth labels, or whether the clinical interpretation is correct. Do not require a specific color scheme or exact stage proportions. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case18/case18_07.json b/sleep_assessment/cases/case18/case18_07.json new file mode 100644 index 0000000000000000000000000000000000000000..1fb904e59ea1b0510a3f00399334a94aec0e9b4d --- /dev/null +++ b/sleep_assessment/cases/case18/case18_07.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "HMC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_02.edf", + "instruction": "Please use the provided sleep file to generate a sleep hypnogram, and save the image to the file_check directory with the filename HMC_02_case18_sleep_hypnogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep hypnogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_hypnogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_hypnogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "HMC_02_case18_sleep_hypnogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep hypnogram plotting task. Check the image only. The task requirement is to generate a sleep hypnogram, also known as a sleep-stage-over-time plot, from a raw sleep recording. Return pass only if the image is clearly a valid sleep hypnogram-style figure. Specifically verify: (1) the image shows sleep stages over time, epochs, or an overnight recording timeline; (2) the y-axis or visible labels represent discrete sleep stages such as W/Wake, N1, N2, N3, and REM/R; (3) the plotted content is a staircase line, stage trajectory, or equivalent discrete sleep-stage time series rather than continuous raw EEG amplitude; (4) the figure is not a PSD plot, spectrogram, heatmap, correlation matrix, bar chart, topomap, text report screenshot, or unrelated medical image; (5) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge whether the sleep-stage sequence is accurate, whether it matches ground-truth labels, or whether the clinical interpretation is correct. Do not require a specific color scheme or exact stage proportions. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case18/case18_08.json b/sleep_assessment/cases/case18/case18_08.json new file mode 100644 index 0000000000000000000000000000000000000000..7475bf2ba7a7068c6b988649d25592e13fa8ad40 --- /dev/null +++ b/sleep_assessment/cases/case18/case18_08.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "HMC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_03.edf", + "instruction": "Please use the provided sleep file to generate a sleep hypnogram, and save the image to the file_check directory with the filename HMC_03_case18_sleep_hypnogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep hypnogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_hypnogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_hypnogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "HMC_03_case18_sleep_hypnogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep hypnogram plotting task. Check the image only. The task requirement is to generate a sleep hypnogram, also known as a sleep-stage-over-time plot, from a raw sleep recording. Return pass only if the image is clearly a valid sleep hypnogram-style figure. Specifically verify: (1) the image shows sleep stages over time, epochs, or an overnight recording timeline; (2) the y-axis or visible labels represent discrete sleep stages such as W/Wake, N1, N2, N3, and REM/R; (3) the plotted content is a staircase line, stage trajectory, or equivalent discrete sleep-stage time series rather than continuous raw EEG amplitude; (4) the figure is not a PSD plot, spectrogram, heatmap, correlation matrix, bar chart, topomap, text report screenshot, or unrelated medical image; (5) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge whether the sleep-stage sequence is accurate, whether it matches ground-truth labels, or whether the clinical interpretation is correct. Do not require a specific color scheme or exact stage proportions. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case18/case18_09.json b/sleep_assessment/cases/case18/case18_09.json new file mode 100644 index 0000000000000000000000000000000000000000..6580c33fa06bb5a1c923b6f5e490247aa8034838 --- /dev/null +++ b/sleep_assessment/cases/case18/case18_09.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "HMC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_04.edf", + "instruction": "Please use the provided sleep file to generate a sleep hypnogram, and save the image to the file_check directory with the filename HMC_04_case18_sleep_hypnogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep hypnogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_hypnogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_hypnogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "HMC_04_case18_sleep_hypnogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep hypnogram plotting task. Check the image only. The task requirement is to generate a sleep hypnogram, also known as a sleep-stage-over-time plot, from a raw sleep recording. Return pass only if the image is clearly a valid sleep hypnogram-style figure. Specifically verify: (1) the image shows sleep stages over time, epochs, or an overnight recording timeline; (2) the y-axis or visible labels represent discrete sleep stages such as W/Wake, N1, N2, N3, and REM/R; (3) the plotted content is a staircase line, stage trajectory, or equivalent discrete sleep-stage time series rather than continuous raw EEG amplitude; (4) the figure is not a PSD plot, spectrogram, heatmap, correlation matrix, bar chart, topomap, text report screenshot, or unrelated medical image; (5) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge whether the sleep-stage sequence is accurate, whether it matches ground-truth labels, or whether the clinical interpretation is correct. Do not require a specific color scheme or exact stage proportions. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case18/case18_10.json b/sleep_assessment/cases/case18/case18_10.json new file mode 100644 index 0000000000000000000000000000000000000000..cb0b42a9a2cd12af0476d7137b1ce08a9ab01180 --- /dev/null +++ b/sleep_assessment/cases/case18/case18_10.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "HMC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_05.edf", + "instruction": "Please use the provided sleep file to generate a sleep hypnogram, and save the image to the file_check directory with the filename HMC_05_case18_sleep_hypnogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep hypnogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_hypnogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_hypnogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "HMC_05_case18_sleep_hypnogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep hypnogram plotting task. Check the image only. The task requirement is to generate a sleep hypnogram, also known as a sleep-stage-over-time plot, from a raw sleep recording. Return pass only if the image is clearly a valid sleep hypnogram-style figure. Specifically verify: (1) the image shows sleep stages over time, epochs, or an overnight recording timeline; (2) the y-axis or visible labels represent discrete sleep stages such as W/Wake, N1, N2, N3, and REM/R; (3) the plotted content is a staircase line, stage trajectory, or equivalent discrete sleep-stage time series rather than continuous raw EEG amplitude; (4) the figure is not a PSD plot, spectrogram, heatmap, correlation matrix, bar chart, topomap, text report screenshot, or unrelated medical image; (5) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge whether the sleep-stage sequence is accurate, whether it matches ground-truth labels, or whether the clinical interpretation is correct. Do not require a specific color scheme or exact stage proportions. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case18/case18_11.json b/sleep_assessment/cases/case18/case18_11.json new file mode 100644 index 0000000000000000000000000000000000000000..337b70a30abfed971fd8b7c8fcdfed4a18342175 --- /dev/null +++ b/sleep_assessment/cases/case18/case18_11.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "instruction": "Please use the provided sleep file to generate a sleep hypnogram, and save the image to the file_check directory with the filename SHHS1_01_case18_sleep_hypnogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep hypnogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_hypnogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_hypnogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "SHHS1_01_case18_sleep_hypnogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep hypnogram plotting task. Check the image only. The task requirement is to generate a sleep hypnogram, also known as a sleep-stage-over-time plot, from a raw sleep recording. Return pass only if the image is clearly a valid sleep hypnogram-style figure. Specifically verify: (1) the image shows sleep stages over time, epochs, or an overnight recording timeline; (2) the y-axis or visible labels represent discrete sleep stages such as W/Wake, N1, N2, N3, and REM/R; (3) the plotted content is a staircase line, stage trajectory, or equivalent discrete sleep-stage time series rather than continuous raw EEG amplitude; (4) the figure is not a PSD plot, spectrogram, heatmap, correlation matrix, bar chart, topomap, text report screenshot, or unrelated medical image; (5) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge whether the sleep-stage sequence is accurate, whether it matches ground-truth labels, or whether the clinical interpretation is correct. Do not require a specific color scheme or exact stage proportions. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case18/case18_12.json b/sleep_assessment/cases/case18/case18_12.json new file mode 100644 index 0000000000000000000000000000000000000000..bca5b4d674a04611f3284fb0b0be3d18d1cebbb0 --- /dev/null +++ b/sleep_assessment/cases/case18/case18_12.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "instruction": "Please use the provided sleep file to generate a sleep hypnogram, and save the image to the file_check directory with the filename SHHS1_02_case18_sleep_hypnogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep hypnogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_hypnogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_hypnogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "SHHS1_02_case18_sleep_hypnogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep hypnogram plotting task. Check the image only. The task requirement is to generate a sleep hypnogram, also known as a sleep-stage-over-time plot, from a raw sleep recording. Return pass only if the image is clearly a valid sleep hypnogram-style figure. Specifically verify: (1) the image shows sleep stages over time, epochs, or an overnight recording timeline; (2) the y-axis or visible labels represent discrete sleep stages such as W/Wake, N1, N2, N3, and REM/R; (3) the plotted content is a staircase line, stage trajectory, or equivalent discrete sleep-stage time series rather than continuous raw EEG amplitude; (4) the figure is not a PSD plot, spectrogram, heatmap, correlation matrix, bar chart, topomap, text report screenshot, or unrelated medical image; (5) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge whether the sleep-stage sequence is accurate, whether it matches ground-truth labels, or whether the clinical interpretation is correct. Do not require a specific color scheme or exact stage proportions. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case18/case18_13.json b/sleep_assessment/cases/case18/case18_13.json new file mode 100644 index 0000000000000000000000000000000000000000..66bd5ba4d5bcd6e75b0e6e86e315e87293548ba3 --- /dev/null +++ b/sleep_assessment/cases/case18/case18_13.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "instruction": "Please use the provided sleep file to generate a sleep hypnogram, and save the image to the file_check directory with the filename SHHS1_03_case18_sleep_hypnogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep hypnogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_hypnogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_hypnogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "SHHS1_03_case18_sleep_hypnogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep hypnogram plotting task. Check the image only. The task requirement is to generate a sleep hypnogram, also known as a sleep-stage-over-time plot, from a raw sleep recording. Return pass only if the image is clearly a valid sleep hypnogram-style figure. Specifically verify: (1) the image shows sleep stages over time, epochs, or an overnight recording timeline; (2) the y-axis or visible labels represent discrete sleep stages such as W/Wake, N1, N2, N3, and REM/R; (3) the plotted content is a staircase line, stage trajectory, or equivalent discrete sleep-stage time series rather than continuous raw EEG amplitude; (4) the figure is not a PSD plot, spectrogram, heatmap, correlation matrix, bar chart, topomap, text report screenshot, or unrelated medical image; (5) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge whether the sleep-stage sequence is accurate, whether it matches ground-truth labels, or whether the clinical interpretation is correct. Do not require a specific color scheme or exact stage proportions. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case18/case18_14.json b/sleep_assessment/cases/case18/case18_14.json new file mode 100644 index 0000000000000000000000000000000000000000..9b38d2410519ac124936636dad00cc1cdc56c2a9 --- /dev/null +++ b/sleep_assessment/cases/case18/case18_14.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "instruction": "Please use the provided sleep file to generate a sleep hypnogram, and save the image to the file_check directory with the filename SHHS1_04_case18_sleep_hypnogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep hypnogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_hypnogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_hypnogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "SHHS1_04_case18_sleep_hypnogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep hypnogram plotting task. Check the image only. The task requirement is to generate a sleep hypnogram, also known as a sleep-stage-over-time plot, from a raw sleep recording. Return pass only if the image is clearly a valid sleep hypnogram-style figure. Specifically verify: (1) the image shows sleep stages over time, epochs, or an overnight recording timeline; (2) the y-axis or visible labels represent discrete sleep stages such as W/Wake, N1, N2, N3, and REM/R; (3) the plotted content is a staircase line, stage trajectory, or equivalent discrete sleep-stage time series rather than continuous raw EEG amplitude; (4) the figure is not a PSD plot, spectrogram, heatmap, correlation matrix, bar chart, topomap, text report screenshot, or unrelated medical image; (5) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge whether the sleep-stage sequence is accurate, whether it matches ground-truth labels, or whether the clinical interpretation is correct. Do not require a specific color scheme or exact stage proportions. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case18/case18_15.json b/sleep_assessment/cases/case18/case18_15.json new file mode 100644 index 0000000000000000000000000000000000000000..c08cb8de982c226b95223600d199b6bed1a692b1 --- /dev/null +++ b/sleep_assessment/cases/case18/case18_15.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "instruction": "Please use the provided sleep file to generate a sleep hypnogram, and save the image to the file_check directory with the filename SHHS1_05_case18_sleep_hypnogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep hypnogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_hypnogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_hypnogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "SHHS1_05_case18_sleep_hypnogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep hypnogram plotting task. Check the image only. The task requirement is to generate a sleep hypnogram, also known as a sleep-stage-over-time plot, from a raw sleep recording. Return pass only if the image is clearly a valid sleep hypnogram-style figure. Specifically verify: (1) the image shows sleep stages over time, epochs, or an overnight recording timeline; (2) the y-axis or visible labels represent discrete sleep stages such as W/Wake, N1, N2, N3, and REM/R; (3) the plotted content is a staircase line, stage trajectory, or equivalent discrete sleep-stage time series rather than continuous raw EEG amplitude; (4) the figure is not a PSD plot, spectrogram, heatmap, correlation matrix, bar chart, topomap, text report screenshot, or unrelated medical image; (5) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge whether the sleep-stage sequence is accurate, whether it matches ground-truth labels, or whether the clinical interpretation is correct. Do not require a specific color scheme or exact stage proportions. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case18/case18_16.json b/sleep_assessment/cases/case18/case18_16.json new file mode 100644 index 0000000000000000000000000000000000000000..12a912b751db86a321e6ce3c610a608711470ae1 --- /dev/null +++ b/sleep_assessment/cases/case18/case18_16.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "MASSSS3_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_01.edf", + "instruction": "Please use the provided sleep file to generate a sleep hypnogram, and save the image to the file_check directory with the filename MASSSS3_01_case18_sleep_hypnogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep hypnogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_hypnogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_hypnogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "MASSSS3_01_case18_sleep_hypnogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep hypnogram plotting task. Check the image only. The task requirement is to generate a sleep hypnogram, also known as a sleep-stage-over-time plot, from a raw sleep recording. Return pass only if the image is clearly a valid sleep hypnogram-style figure. Specifically verify: (1) the image shows sleep stages over time, epochs, or an overnight recording timeline; (2) the y-axis or visible labels represent discrete sleep stages such as W/Wake, N1, N2, N3, and REM/R; (3) the plotted content is a staircase line, stage trajectory, or equivalent discrete sleep-stage time series rather than continuous raw EEG amplitude; (4) the figure is not a PSD plot, spectrogram, heatmap, correlation matrix, bar chart, topomap, text report screenshot, or unrelated medical image; (5) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge whether the sleep-stage sequence is accurate, whether it matches ground-truth labels, or whether the clinical interpretation is correct. Do not require a specific color scheme or exact stage proportions. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case18/case18_17.json b/sleep_assessment/cases/case18/case18_17.json new file mode 100644 index 0000000000000000000000000000000000000000..5560fc2810de7d624d5795d82892a65e5c3fa240 --- /dev/null +++ b/sleep_assessment/cases/case18/case18_17.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "MASSSS3_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_02.edf", + "instruction": "Please use the provided sleep file to generate a sleep hypnogram, and save the image to the file_check directory with the filename MASSSS3_02_case18_sleep_hypnogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep hypnogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_hypnogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_hypnogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "MASSSS3_02_case18_sleep_hypnogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep hypnogram plotting task. Check the image only. The task requirement is to generate a sleep hypnogram, also known as a sleep-stage-over-time plot, from a raw sleep recording. Return pass only if the image is clearly a valid sleep hypnogram-style figure. Specifically verify: (1) the image shows sleep stages over time, epochs, or an overnight recording timeline; (2) the y-axis or visible labels represent discrete sleep stages such as W/Wake, N1, N2, N3, and REM/R; (3) the plotted content is a staircase line, stage trajectory, or equivalent discrete sleep-stage time series rather than continuous raw EEG amplitude; (4) the figure is not a PSD plot, spectrogram, heatmap, correlation matrix, bar chart, topomap, text report screenshot, or unrelated medical image; (5) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge whether the sleep-stage sequence is accurate, whether it matches ground-truth labels, or whether the clinical interpretation is correct. Do not require a specific color scheme or exact stage proportions. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case18/case18_18.json b/sleep_assessment/cases/case18/case18_18.json new file mode 100644 index 0000000000000000000000000000000000000000..d701d9125a424694d3063ca2bafa5fac5a040500 --- /dev/null +++ b/sleep_assessment/cases/case18/case18_18.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "MASSSS3_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_03.edf", + "instruction": "Please use the provided sleep file to generate a sleep hypnogram, and save the image to the file_check directory with the filename MASSSS3_03_case18_sleep_hypnogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep hypnogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_hypnogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_hypnogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "MASSSS3_03_case18_sleep_hypnogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep hypnogram plotting task. Check the image only. The task requirement is to generate a sleep hypnogram, also known as a sleep-stage-over-time plot, from a raw sleep recording. Return pass only if the image is clearly a valid sleep hypnogram-style figure. Specifically verify: (1) the image shows sleep stages over time, epochs, or an overnight recording timeline; (2) the y-axis or visible labels represent discrete sleep stages such as W/Wake, N1, N2, N3, and REM/R; (3) the plotted content is a staircase line, stage trajectory, or equivalent discrete sleep-stage time series rather than continuous raw EEG amplitude; (4) the figure is not a PSD plot, spectrogram, heatmap, correlation matrix, bar chart, topomap, text report screenshot, or unrelated medical image; (5) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge whether the sleep-stage sequence is accurate, whether it matches ground-truth labels, or whether the clinical interpretation is correct. Do not require a specific color scheme or exact stage proportions. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case18/case18_19.json b/sleep_assessment/cases/case18/case18_19.json new file mode 100644 index 0000000000000000000000000000000000000000..1fdd4093c92f0e4032c8e38b6df2026b5ff24216 --- /dev/null +++ b/sleep_assessment/cases/case18/case18_19.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "MASSSS3_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_04.edf", + "instruction": "Please use the provided sleep file to generate a sleep hypnogram, and save the image to the file_check directory with the filename MASSSS3_04_case18_sleep_hypnogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep hypnogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_hypnogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_hypnogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "MASSSS3_04_case18_sleep_hypnogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep hypnogram plotting task. Check the image only. The task requirement is to generate a sleep hypnogram, also known as a sleep-stage-over-time plot, from a raw sleep recording. Return pass only if the image is clearly a valid sleep hypnogram-style figure. Specifically verify: (1) the image shows sleep stages over time, epochs, or an overnight recording timeline; (2) the y-axis or visible labels represent discrete sleep stages such as W/Wake, N1, N2, N3, and REM/R; (3) the plotted content is a staircase line, stage trajectory, or equivalent discrete sleep-stage time series rather than continuous raw EEG amplitude; (4) the figure is not a PSD plot, spectrogram, heatmap, correlation matrix, bar chart, topomap, text report screenshot, or unrelated medical image; (5) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge whether the sleep-stage sequence is accurate, whether it matches ground-truth labels, or whether the clinical interpretation is correct. Do not require a specific color scheme or exact stage proportions. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case18/case18_20.json b/sleep_assessment/cases/case18/case18_20.json new file mode 100644 index 0000000000000000000000000000000000000000..fe1a936799235641cdf957f0886402fb4af44eec --- /dev/null +++ b/sleep_assessment/cases/case18/case18_20.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "MASSSS3_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_05.edf", + "instruction": "Please use the provided sleep file to generate a sleep hypnogram, and save the image to the file_check directory with the filename MASSSS3_05_case18_sleep_hypnogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep hypnogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_hypnogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_hypnogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "MASSSS3_05_case18_sleep_hypnogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep hypnogram plotting task. Check the image only. The task requirement is to generate a sleep hypnogram, also known as a sleep-stage-over-time plot, from a raw sleep recording. Return pass only if the image is clearly a valid sleep hypnogram-style figure. Specifically verify: (1) the image shows sleep stages over time, epochs, or an overnight recording timeline; (2) the y-axis or visible labels represent discrete sleep stages such as W/Wake, N1, N2, N3, and REM/R; (3) the plotted content is a staircase line, stage trajectory, or equivalent discrete sleep-stage time series rather than continuous raw EEG amplitude; (4) the figure is not a PSD plot, spectrogram, heatmap, correlation matrix, bar chart, topomap, text report screenshot, or unrelated medical image; (5) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge whether the sleep-stage sequence is accurate, whether it matches ground-truth labels, or whether the clinical interpretation is correct. Do not require a specific color scheme or exact stage proportions. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case18/case18_21.json b/sleep_assessment/cases/case18/case18_21.json new file mode 100644 index 0000000000000000000000000000000000000000..1aabadcb5632ef4e547b5751a38c355e614c8208 --- /dev/null +++ b/sleep_assessment/cases/case18/case18_21.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "instruction": "Please use the provided sleep file to generate a sleep hypnogram, and save the image to the file_check directory with the filename Physionet2018_01_case18_sleep_hypnogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep hypnogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_hypnogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_hypnogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "Physionet2018_01_case18_sleep_hypnogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep hypnogram plotting task. Check the image only. The task requirement is to generate a sleep hypnogram, also known as a sleep-stage-over-time plot, from a raw sleep recording. Return pass only if the image is clearly a valid sleep hypnogram-style figure. Specifically verify: (1) the image shows sleep stages over time, epochs, or an overnight recording timeline; (2) the y-axis or visible labels represent discrete sleep stages such as W/Wake, N1, N2, N3, and REM/R; (3) the plotted content is a staircase line, stage trajectory, or equivalent discrete sleep-stage time series rather than continuous raw EEG amplitude; (4) the figure is not a PSD plot, spectrogram, heatmap, correlation matrix, bar chart, topomap, text report screenshot, or unrelated medical image; (5) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge whether the sleep-stage sequence is accurate, whether it matches ground-truth labels, or whether the clinical interpretation is correct. Do not require a specific color scheme or exact stage proportions. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case18/case18_22.json b/sleep_assessment/cases/case18/case18_22.json new file mode 100644 index 0000000000000000000000000000000000000000..975dc59e41c57206358212fed5bf332098c56bad --- /dev/null +++ b/sleep_assessment/cases/case18/case18_22.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "instruction": "Please use the provided sleep file to generate a sleep hypnogram, and save the image to the file_check directory with the filename Physionet2018_02_case18_sleep_hypnogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep hypnogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_hypnogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_hypnogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "Physionet2018_02_case18_sleep_hypnogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep hypnogram plotting task. Check the image only. The task requirement is to generate a sleep hypnogram, also known as a sleep-stage-over-time plot, from a raw sleep recording. Return pass only if the image is clearly a valid sleep hypnogram-style figure. Specifically verify: (1) the image shows sleep stages over time, epochs, or an overnight recording timeline; (2) the y-axis or visible labels represent discrete sleep stages such as W/Wake, N1, N2, N3, and REM/R; (3) the plotted content is a staircase line, stage trajectory, or equivalent discrete sleep-stage time series rather than continuous raw EEG amplitude; (4) the figure is not a PSD plot, spectrogram, heatmap, correlation matrix, bar chart, topomap, text report screenshot, or unrelated medical image; (5) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge whether the sleep-stage sequence is accurate, whether it matches ground-truth labels, or whether the clinical interpretation is correct. Do not require a specific color scheme or exact stage proportions. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case18/case18_23.json b/sleep_assessment/cases/case18/case18_23.json new file mode 100644 index 0000000000000000000000000000000000000000..d262e316669c65fa9f58f3662b75355931050f87 --- /dev/null +++ b/sleep_assessment/cases/case18/case18_23.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "Physionet2018_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_03.edf", + "instruction": "Please use the provided sleep file to generate a sleep hypnogram, and save the image to the file_check directory with the filename Physionet2018_03_case18_sleep_hypnogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep hypnogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_hypnogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_hypnogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "Physionet2018_03_case18_sleep_hypnogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep hypnogram plotting task. Check the image only. The task requirement is to generate a sleep hypnogram, also known as a sleep-stage-over-time plot, from a raw sleep recording. Return pass only if the image is clearly a valid sleep hypnogram-style figure. Specifically verify: (1) the image shows sleep stages over time, epochs, or an overnight recording timeline; (2) the y-axis or visible labels represent discrete sleep stages such as W/Wake, N1, N2, N3, and REM/R; (3) the plotted content is a staircase line, stage trajectory, or equivalent discrete sleep-stage time series rather than continuous raw EEG amplitude; (4) the figure is not a PSD plot, spectrogram, heatmap, correlation matrix, bar chart, topomap, text report screenshot, or unrelated medical image; (5) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge whether the sleep-stage sequence is accurate, whether it matches ground-truth labels, or whether the clinical interpretation is correct. Do not require a specific color scheme or exact stage proportions. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case18/case18_24.json b/sleep_assessment/cases/case18/case18_24.json new file mode 100644 index 0000000000000000000000000000000000000000..7409600cdbdd8d3f87a337ef744ffc730cd8c14b --- /dev/null +++ b/sleep_assessment/cases/case18/case18_24.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "instruction": "Please use the provided sleep file to generate a sleep hypnogram, and save the image to the file_check directory with the filename Physionet2018_04_case18_sleep_hypnogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep hypnogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_hypnogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_hypnogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "Physionet2018_04_case18_sleep_hypnogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep hypnogram plotting task. Check the image only. The task requirement is to generate a sleep hypnogram, also known as a sleep-stage-over-time plot, from a raw sleep recording. Return pass only if the image is clearly a valid sleep hypnogram-style figure. Specifically verify: (1) the image shows sleep stages over time, epochs, or an overnight recording timeline; (2) the y-axis or visible labels represent discrete sleep stages such as W/Wake, N1, N2, N3, and REM/R; (3) the plotted content is a staircase line, stage trajectory, or equivalent discrete sleep-stage time series rather than continuous raw EEG amplitude; (4) the figure is not a PSD plot, spectrogram, heatmap, correlation matrix, bar chart, topomap, text report screenshot, or unrelated medical image; (5) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge whether the sleep-stage sequence is accurate, whether it matches ground-truth labels, or whether the clinical interpretation is correct. Do not require a specific color scheme or exact stage proportions. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case18/case18_25.json b/sleep_assessment/cases/case18/case18_25.json new file mode 100644 index 0000000000000000000000000000000000000000..30b7548005859dfa5d5268dce828c7bb92a1c4fe --- /dev/null +++ b/sleep_assessment/cases/case18/case18_25.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "instruction": "Please use the provided sleep file to generate a sleep hypnogram, and save the image to the file_check directory with the filename Physionet2018_05_case18_sleep_hypnogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep hypnogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_hypnogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_hypnogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "Physionet2018_05_case18_sleep_hypnogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep hypnogram plotting task. Check the image only. The task requirement is to generate a sleep hypnogram, also known as a sleep-stage-over-time plot, from a raw sleep recording. Return pass only if the image is clearly a valid sleep hypnogram-style figure. Specifically verify: (1) the image shows sleep stages over time, epochs, or an overnight recording timeline; (2) the y-axis or visible labels represent discrete sleep stages such as W/Wake, N1, N2, N3, and REM/R; (3) the plotted content is a staircase line, stage trajectory, or equivalent discrete sleep-stage time series rather than continuous raw EEG amplitude; (4) the figure is not a PSD plot, spectrogram, heatmap, correlation matrix, bar chart, topomap, text report screenshot, or unrelated medical image; (5) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge whether the sleep-stage sequence is accurate, whether it matches ground-truth labels, or whether the clinical interpretation is correct. Do not require a specific color scheme or exact stage proportions. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case19/case19_01.json b/sleep_assessment/cases/case19/case19_01.json new file mode 100644 index 0000000000000000000000000000000000000000..3073c63975ff9978c6685a15ca8b1723a1aca573 --- /dev/null +++ b/sleep_assessment/cases/case19/case19_01.json @@ -0,0 +1,44 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "instruction": "Please use the provided sleep file to complete preprocessing for sleep staging. First select EEG channels only. Then apply a 60Hz notch filter because this case treats all Sleep subset data sources as non-China datasets, apply a 0.5-40Hz bandpass filter(apply IIR filtering), set an average reference, and resample to 100Hz. Next split the processed signal into non-overlapping 30-second epochs and save the data as a .npy array with shape (epoch_num, channel_num, 3000). Discard any final segment shorter than 30 seconds. Save the file to the file_check directory with the filename ISRUC_01_case19_processed_data.npy." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved preprocessed NPY file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved NPY file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_processed_data.npy\"}", + "metrics": [ + { + "metric_id": "processed_sleep_epoch_npy_status", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "eeg_file_match", + "validator_config": { + "file_name": { + "value": "ISRUC_01_case19_processed_data.npy" + }, + "npy_shape": { + "value": [ + 880, + 6, + 3000 + ], + "weight": 50 + }, + "signal_rms": { + "value": 2.8576221746994855e-06, + "weight": 50, + "tolerance": 2.857622174699486e-07 + } + } + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case19/case19_02.json b/sleep_assessment/cases/case19/case19_02.json new file mode 100644 index 0000000000000000000000000000000000000000..9c4a50fcae1a5156a6812f9dbbe14a6eea7b82f0 --- /dev/null +++ b/sleep_assessment/cases/case19/case19_02.json @@ -0,0 +1,44 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "instruction": "Please use the provided sleep file to complete preprocessing for sleep staging. First select EEG channels only. Then apply a 60Hz notch filter because this case treats all Sleep subset data sources as non-China datasets, apply a 0.5-40Hz bandpass filter(apply IIR filtering), set an average reference, and resample to 100Hz. Next split the processed signal into non-overlapping 30-second epochs and save the data as a .npy array with shape (epoch_num, channel_num, 3000). Discard any final segment shorter than 30 seconds. Save the file to the file_check directory with the filename ISRUC_02_case19_processed_data.npy." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved preprocessed NPY file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved NPY file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_processed_data.npy\"}", + "metrics": [ + { + "metric_id": "processed_sleep_epoch_npy_status", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "eeg_file_match", + "validator_config": { + "file_name": { + "value": "ISRUC_02_case19_processed_data.npy" + }, + "npy_shape": { + "value": [ + 964, + 6, + 3000 + ], + "weight": 50 + }, + "signal_rms": { + "value": 7.150938501825713e-07, + "weight": 50, + "tolerance": 7.150938501825712e-08 + } + } + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case19/case19_03.json b/sleep_assessment/cases/case19/case19_03.json new file mode 100644 index 0000000000000000000000000000000000000000..0febd9fe81d40f64297bf8e3ddb1ec055c17f252 --- /dev/null +++ b/sleep_assessment/cases/case19/case19_03.json @@ -0,0 +1,44 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "instruction": "Please use the provided sleep file to complete preprocessing for sleep staging. First select EEG channels only. Then apply a 60Hz notch filter because this case treats all Sleep subset data sources as non-China datasets, apply a 0.5-40Hz bandpass filter(apply IIR filtering), set an average reference, and resample to 100Hz. Next split the processed signal into non-overlapping 30-second epochs and save the data as a .npy array with shape (epoch_num, channel_num, 3000). Discard any final segment shorter than 30 seconds. Save the file to the file_check directory with the filename ISRUC_03_case19_processed_data.npy." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved preprocessed NPY file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved NPY file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_processed_data.npy\"}", + "metrics": [ + { + "metric_id": "processed_sleep_epoch_npy_status", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "eeg_file_match", + "validator_config": { + "file_name": { + "value": "ISRUC_03_case19_processed_data.npy" + }, + "npy_shape": { + "value": [ + 943, + 6, + 3000 + ], + "weight": 50 + }, + "signal_rms": { + "value": 9.646151925984784e-07, + "weight": 50, + "tolerance": 9.646151925984785e-08 + } + } + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case19/case19_04.json b/sleep_assessment/cases/case19/case19_04.json new file mode 100644 index 0000000000000000000000000000000000000000..86474cb262104f2166c3b91d634105bf187365ad --- /dev/null +++ b/sleep_assessment/cases/case19/case19_04.json @@ -0,0 +1,44 @@ +{ + "meta_info": { + "case_id": "ISRUC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_04.edf", + "instruction": "Please use the provided sleep file to complete preprocessing for sleep staging. First select EEG channels only. Then apply a 60Hz notch filter because this case treats all Sleep subset data sources as non-China datasets, apply a 0.5-40Hz bandpass filter(apply IIR filtering), set an average reference, and resample to 100Hz. Next split the processed signal into non-overlapping 30-second epochs and save the data as a .npy array with shape (epoch_num, channel_num, 3000). Discard any final segment shorter than 30 seconds. Save the file to the file_check directory with the filename ISRUC_04_case19_processed_data.npy." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved preprocessed NPY file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved NPY file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_processed_data.npy\"}", + "metrics": [ + { + "metric_id": "processed_sleep_epoch_npy_status", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "eeg_file_match", + "validator_config": { + "file_name": { + "value": "ISRUC_04_case19_processed_data.npy" + }, + "npy_shape": { + "value": [ + 963, + 6, + 3000 + ], + "weight": 50 + }, + "signal_rms": { + "value": 1.6586367305902015e-06, + "weight": 50, + "tolerance": 1.6586367305902016e-07 + } + } + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case19/case19_05.json b/sleep_assessment/cases/case19/case19_05.json new file mode 100644 index 0000000000000000000000000000000000000000..f1306c5efbdfab2cb7ad5839f9612f5679e47417 --- /dev/null +++ b/sleep_assessment/cases/case19/case19_05.json @@ -0,0 +1,44 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "instruction": "Please use the provided sleep file to complete preprocessing for sleep staging. First select EEG channels only. Then apply a 60Hz notch filter because this case treats all Sleep subset data sources as non-China datasets, apply a 0.5-40Hz bandpass filter(apply IIR filtering), set an average reference, and resample to 100Hz. Next split the processed signal into non-overlapping 30-second epochs and save the data as a .npy array with shape (epoch_num, channel_num, 3000). Discard any final segment shorter than 30 seconds. Save the file to the file_check directory with the filename ISRUC_05_case19_processed_data.npy." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved preprocessed NPY file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved NPY file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_processed_data.npy\"}", + "metrics": [ + { + "metric_id": "processed_sleep_epoch_npy_status", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "eeg_file_match", + "validator_config": { + "file_name": { + "value": "ISRUC_05_case19_processed_data.npy" + }, + "npy_shape": { + "value": [ + 875, + 6, + 3000 + ], + "weight": 50 + }, + "signal_rms": { + "value": 9.007648216435279e-07, + "weight": 50, + "tolerance": 9.007648216435279e-08 + } + } + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case19/case19_06.json b/sleep_assessment/cases/case19/case19_06.json new file mode 100644 index 0000000000000000000000000000000000000000..2e45143eeeabe4918e031c0f3925a80a61291c7c --- /dev/null +++ b/sleep_assessment/cases/case19/case19_06.json @@ -0,0 +1,44 @@ +{ + "meta_info": { + "case_id": "HMC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_01.edf", + "instruction": "Please use the provided sleep file to complete preprocessing for sleep staging. First select EEG channels only. Then apply a 60Hz notch filter because this case treats all Sleep subset data sources as non-China datasets, apply a 0.5-40Hz bandpass filter(apply IIR filtering), set an average reference, and resample to 100Hz. Next split the processed signal into non-overlapping 30-second epochs and save the data as a .npy array with shape (epoch_num, channel_num, 3000). Discard any final segment shorter than 30 seconds. Save the file to the file_check directory with the filename HMC_01_case19_processed_data.npy." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved preprocessed NPY file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved NPY file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_processed_data.npy\"}", + "metrics": [ + { + "metric_id": "processed_sleep_epoch_npy_status", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "eeg_file_match", + "validator_config": { + "file_name": { + "value": "HMC_01_case19_processed_data.npy" + }, + "npy_shape": { + "value": [ + 855, + 4, + 3000 + ], + "weight": 50 + }, + "signal_rms": { + "value": 1.3020269313234744e-05, + "weight": 50, + "tolerance": 1.3020269313234744e-06 + } + } + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case19/case19_07.json b/sleep_assessment/cases/case19/case19_07.json new file mode 100644 index 0000000000000000000000000000000000000000..9464bd30f0c8d5e4e789a5e081817efef8f176c6 --- /dev/null +++ b/sleep_assessment/cases/case19/case19_07.json @@ -0,0 +1,44 @@ +{ + "meta_info": { + "case_id": "HMC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_02.edf", + "instruction": "Please use the provided sleep file to complete preprocessing for sleep staging. First select EEG channels only. Then apply a 60Hz notch filter because this case treats all Sleep subset data sources as non-China datasets, apply a 0.5-40Hz bandpass filter(apply IIR filtering), set an average reference, and resample to 100Hz. Next split the processed signal into non-overlapping 30-second epochs and save the data as a .npy array with shape (epoch_num, channel_num, 3000). Discard any final segment shorter than 30 seconds. Save the file to the file_check directory with the filename HMC_02_case19_processed_data.npy." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved preprocessed NPY file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved NPY file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_processed_data.npy\"}", + "metrics": [ + { + "metric_id": "processed_sleep_epoch_npy_status", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "eeg_file_match", + "validator_config": { + "file_name": { + "value": "HMC_02_case19_processed_data.npy" + }, + "npy_shape": { + "value": [ + 856, + 4, + 3000 + ], + "weight": 50 + }, + "signal_rms": { + "value": 1.2270847782234478e-05, + "weight": 50, + "tolerance": 1.2270847782234478e-06 + } + } + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case19/case19_08.json b/sleep_assessment/cases/case19/case19_08.json new file mode 100644 index 0000000000000000000000000000000000000000..f5085bace2ce3c797adae4932c01873ab36d426a --- /dev/null +++ b/sleep_assessment/cases/case19/case19_08.json @@ -0,0 +1,44 @@ +{ + "meta_info": { + "case_id": "HMC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_03.edf", + "instruction": "Please use the provided sleep file to complete preprocessing for sleep staging. First select EEG channels only. Then apply a 60Hz notch filter because this case treats all Sleep subset data sources as non-China datasets, apply a 0.5-40Hz bandpass filter(apply IIR filtering), set an average reference, and resample to 100Hz. Next split the processed signal into non-overlapping 30-second epochs and save the data as a .npy array with shape (epoch_num, channel_num, 3000). Discard any final segment shorter than 30 seconds. Save the file to the file_check directory with the filename HMC_03_case19_processed_data.npy." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved preprocessed NPY file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved NPY file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_processed_data.npy\"}", + "metrics": [ + { + "metric_id": "processed_sleep_epoch_npy_status", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "eeg_file_match", + "validator_config": { + "file_name": { + "value": "HMC_03_case19_processed_data.npy" + }, + "npy_shape": { + "value": [ + 954, + 4, + 3000 + ], + "weight": 50 + }, + "signal_rms": { + "value": 1.549550757606623e-05, + "weight": 50, + "tolerance": 1.5495507576066231e-06 + } + } + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case19/case19_09.json b/sleep_assessment/cases/case19/case19_09.json new file mode 100644 index 0000000000000000000000000000000000000000..b20a1d2e6d0f071328c543172d2a5bc10e7dd5c3 --- /dev/null +++ b/sleep_assessment/cases/case19/case19_09.json @@ -0,0 +1,44 @@ +{ + "meta_info": { + "case_id": "HMC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_04.edf", + "instruction": "Please use the provided sleep file to complete preprocessing for sleep staging. First select EEG channels only. Then apply a 60Hz notch filter because this case treats all Sleep subset data sources as non-China datasets, apply a 0.5-40Hz bandpass filter(apply IIR filtering), set an average reference, and resample to 100Hz. Next split the processed signal into non-overlapping 30-second epochs and save the data as a .npy array with shape (epoch_num, channel_num, 3000). Discard any final segment shorter than 30 seconds. Save the file to the file_check directory with the filename HMC_04_case19_processed_data.npy." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved preprocessed NPY file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved NPY file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_processed_data.npy\"}", + "metrics": [ + { + "metric_id": "processed_sleep_epoch_npy_status", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "eeg_file_match", + "validator_config": { + "file_name": { + "value": "HMC_04_case19_processed_data.npy" + }, + "npy_shape": { + "value": [ + 1016, + 4, + 3000 + ], + "weight": 50 + }, + "signal_rms": { + "value": 6.450915364145694e-06, + "weight": 50, + "tolerance": 6.450915364145695e-07 + } + } + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case19/case19_10.json b/sleep_assessment/cases/case19/case19_10.json new file mode 100644 index 0000000000000000000000000000000000000000..2b1de41a28a940c5edd425d6e9e446f450e4a18a --- /dev/null +++ b/sleep_assessment/cases/case19/case19_10.json @@ -0,0 +1,44 @@ +{ + "meta_info": { + "case_id": "HMC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_05.edf", + "instruction": "Please use the provided sleep file to complete preprocessing for sleep staging. First select EEG channels only. Then apply a 60Hz notch filter because this case treats all Sleep subset data sources as non-China datasets, apply a 0.5-40Hz bandpass filter(apply IIR filtering), set an average reference, and resample to 100Hz. Next split the processed signal into non-overlapping 30-second epochs and save the data as a .npy array with shape (epoch_num, channel_num, 3000). Discard any final segment shorter than 30 seconds. Save the file to the file_check directory with the filename HMC_05_case19_processed_data.npy." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved preprocessed NPY file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved NPY file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_processed_data.npy\"}", + "metrics": [ + { + "metric_id": "processed_sleep_epoch_npy_status", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "eeg_file_match", + "validator_config": { + "file_name": { + "value": "HMC_05_case19_processed_data.npy" + }, + "npy_shape": { + "value": [ + 959, + 4, + 3000 + ], + "weight": 50 + }, + "signal_rms": { + "value": 1.56099408047631e-05, + "weight": 50, + "tolerance": 1.56099408047631e-06 + } + } + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case19/case19_11.json b/sleep_assessment/cases/case19/case19_11.json new file mode 100644 index 0000000000000000000000000000000000000000..a09651a1a89ee3ef831627877eb6efad5455298d --- /dev/null +++ b/sleep_assessment/cases/case19/case19_11.json @@ -0,0 +1,44 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "instruction": "Please use the provided sleep file to complete preprocessing for sleep staging. First select EEG channels only. Then apply a 60Hz notch filter because this case treats all Sleep subset data sources as non-China datasets, apply a 0.5-40Hz bandpass filter(apply IIR filtering), set an average reference, and resample to 100Hz. Next split the processed signal into non-overlapping 30-second epochs and save the data as a .npy array with shape (epoch_num, channel_num, 3000). Discard any final segment shorter than 30 seconds. Save the file to the file_check directory with the filename SHHS1_01_case19_processed_data.npy." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved preprocessed NPY file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved NPY file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_processed_data.npy\"}", + "metrics": [ + { + "metric_id": "processed_sleep_epoch_npy_status", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "eeg_file_match", + "validator_config": { + "file_name": { + "value": "SHHS1_01_case19_processed_data.npy" + }, + "npy_shape": { + "value": [ + 1084, + 2, + 3000 + ], + "weight": 50 + }, + "signal_rms": { + "value": 2.3383364617041573e-05, + "weight": 50, + "tolerance": 2.3383364617041573e-06 + } + } + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case19/case19_12.json b/sleep_assessment/cases/case19/case19_12.json new file mode 100644 index 0000000000000000000000000000000000000000..9d73c882906611d72de4821fef4aba9b079077e4 --- /dev/null +++ b/sleep_assessment/cases/case19/case19_12.json @@ -0,0 +1,44 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "instruction": "Please use the provided sleep file to complete preprocessing for sleep staging. First select EEG channels only. Then apply a 60Hz notch filter because this case treats all Sleep subset data sources as non-China datasets, apply a 0.5-40Hz bandpass filter(apply IIR filtering), set an average reference, and resample to 100Hz. Next split the processed signal into non-overlapping 30-second epochs and save the data as a .npy array with shape (epoch_num, channel_num, 3000). Discard any final segment shorter than 30 seconds. Save the file to the file_check directory with the filename SHHS1_02_case19_processed_data.npy." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved preprocessed NPY file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved NPY file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_processed_data.npy\"}", + "metrics": [ + { + "metric_id": "processed_sleep_epoch_npy_status", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "eeg_file_match", + "validator_config": { + "file_name": { + "value": "SHHS1_02_case19_processed_data.npy" + }, + "npy_shape": { + "value": [ + 1079, + 2, + 3000 + ], + "weight": 50 + }, + "signal_rms": { + "value": 5.887706030978153e-05, + "weight": 50, + "tolerance": 5.887706030978153e-06 + } + } + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case19/case19_13.json b/sleep_assessment/cases/case19/case19_13.json new file mode 100644 index 0000000000000000000000000000000000000000..cbbba891099a10cefee033b738ba6843c1435701 --- /dev/null +++ b/sleep_assessment/cases/case19/case19_13.json @@ -0,0 +1,44 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "instruction": "Please use the provided sleep file to complete preprocessing for sleep staging. First select EEG channels only. Then apply a 60Hz notch filter because this case treats all Sleep subset data sources as non-China datasets, apply a 0.5-40Hz bandpass filter(apply IIR filtering), set an average reference, and resample to 100Hz. Next split the processed signal into non-overlapping 30-second epochs and save the data as a .npy array with shape (epoch_num, channel_num, 3000). Discard any final segment shorter than 30 seconds. Save the file to the file_check directory with the filename SHHS1_03_case19_processed_data.npy." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved preprocessed NPY file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved NPY file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_processed_data.npy\"}", + "metrics": [ + { + "metric_id": "processed_sleep_epoch_npy_status", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "eeg_file_match", + "validator_config": { + "file_name": { + "value": "SHHS1_03_case19_processed_data.npy" + }, + "npy_shape": { + "value": [ + 1049, + 2, + 3000 + ], + "weight": 50 + }, + "signal_rms": { + "value": 1.2323215794421976e-05, + "weight": 50, + "tolerance": 1.2323215794421976e-06 + } + } + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case19/case19_14.json b/sleep_assessment/cases/case19/case19_14.json new file mode 100644 index 0000000000000000000000000000000000000000..f9e65951f869b5e3d68b3dbf1adb5d6f11ac6b61 --- /dev/null +++ b/sleep_assessment/cases/case19/case19_14.json @@ -0,0 +1,44 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "instruction": "Please use the provided sleep file to complete preprocessing for sleep staging. First select EEG channels only. Then apply a 60Hz notch filter because this case treats all Sleep subset data sources as non-China datasets, apply a 0.5-40Hz bandpass filter(apply IIR filtering), set an average reference, and resample to 100Hz. Next split the processed signal into non-overlapping 30-second epochs and save the data as a .npy array with shape (epoch_num, channel_num, 3000). Discard any final segment shorter than 30 seconds. Save the file to the file_check directory with the filename SHHS1_04_case19_processed_data.npy." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved preprocessed NPY file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved NPY file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_processed_data.npy\"}", + "metrics": [ + { + "metric_id": "processed_sleep_epoch_npy_status", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "eeg_file_match", + "validator_config": { + "file_name": { + "value": "SHHS1_04_case19_processed_data.npy" + }, + "npy_shape": { + "value": [ + 875, + 2, + 3000 + ], + "weight": 50 + }, + "signal_rms": { + "value": 1.038779187795578e-05, + "weight": 50, + "tolerance": 1.0387791877955781e-06 + } + } + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case19/case19_15.json b/sleep_assessment/cases/case19/case19_15.json new file mode 100644 index 0000000000000000000000000000000000000000..f5a0c73a9ebcc66fffcc849be57cfc49503113f1 --- /dev/null +++ b/sleep_assessment/cases/case19/case19_15.json @@ -0,0 +1,44 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "instruction": "Please use the provided sleep file to complete preprocessing for sleep staging. First select EEG channels only. Then apply a 60Hz notch filter because this case treats all Sleep subset data sources as non-China datasets, apply a 0.5-40Hz bandpass filter(apply IIR filtering), set an average reference, and resample to 100Hz. Next split the processed signal into non-overlapping 30-second epochs and save the data as a .npy array with shape (epoch_num, channel_num, 3000). Discard any final segment shorter than 30 seconds. Save the file to the file_check directory with the filename SHHS1_05_case19_processed_data.npy." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved preprocessed NPY file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved NPY file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_processed_data.npy\"}", + "metrics": [ + { + "metric_id": "processed_sleep_epoch_npy_status", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "eeg_file_match", + "validator_config": { + "file_name": { + "value": "SHHS1_05_case19_processed_data.npy" + }, + "npy_shape": { + "value": [ + 1084, + 2, + 3000 + ], + "weight": 50 + }, + "signal_rms": { + "value": 9.001139407932791e-06, + "weight": 50, + "tolerance": 9.001139407932792e-07 + } + } + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case19/case19_16.json b/sleep_assessment/cases/case19/case19_16.json new file mode 100644 index 0000000000000000000000000000000000000000..72ed550dfdb48fe216ab9ad41f35922cf959e125 --- /dev/null +++ b/sleep_assessment/cases/case19/case19_16.json @@ -0,0 +1,44 @@ +{ + "meta_info": { + "case_id": "MASSSS3_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_01.edf", + "instruction": "Please use the provided sleep file to complete preprocessing for sleep staging. First select EEG channels only. Then apply a 60Hz notch filter because this case treats all Sleep subset data sources as non-China datasets, apply a 0.5-40Hz bandpass filter(apply IIR filtering), set an average reference, and resample to 100Hz. Next split the processed signal into non-overlapping 30-second epochs and save the data as a .npy array with shape (epoch_num, channel_num, 3000). Discard any final segment shorter than 30 seconds. Save the file to the file_check directory with the filename MASSSS3_01_case19_processed_data.npy." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved preprocessed NPY file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved NPY file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_processed_data.npy\"}", + "metrics": [ + { + "metric_id": "processed_sleep_epoch_npy_status", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "eeg_file_match", + "validator_config": { + "file_name": { + "value": "MASSSS3_01_case19_processed_data.npy" + }, + "npy_shape": { + "value": [ + 1008, + 20, + 3000 + ], + "weight": 50 + }, + "signal_rms": { + "value": 8.169859077706497e-06, + "weight": 50, + "tolerance": 8.169859077706497e-07 + } + } + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case19/case19_17.json b/sleep_assessment/cases/case19/case19_17.json new file mode 100644 index 0000000000000000000000000000000000000000..38da174c2c3e79d93638f615d133cce490665a51 --- /dev/null +++ b/sleep_assessment/cases/case19/case19_17.json @@ -0,0 +1,44 @@ +{ + "meta_info": { + "case_id": "MASSSS3_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_02.edf", + "instruction": "Please use the provided sleep file to complete preprocessing for sleep staging. First select EEG channels only. Then apply a 60Hz notch filter because this case treats all Sleep subset data sources as non-China datasets, apply a 0.5-40Hz bandpass filter(apply IIR filtering), set an average reference, and resample to 100Hz. Next split the processed signal into non-overlapping 30-second epochs and save the data as a .npy array with shape (epoch_num, channel_num, 3000). Discard any final segment shorter than 30 seconds. Save the file to the file_check directory with the filename MASSSS3_02_case19_processed_data.npy." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved preprocessed NPY file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved NPY file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_processed_data.npy\"}", + "metrics": [ + { + "metric_id": "processed_sleep_epoch_npy_status", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "eeg_file_match", + "validator_config": { + "file_name": { + "value": "MASSSS3_02_case19_processed_data.npy" + }, + "npy_shape": { + "value": [ + 1072, + 20, + 3000 + ], + "weight": 50 + }, + "signal_rms": { + "value": 9.701064503925927e-06, + "weight": 50, + "tolerance": 9.701064503925927e-07 + } + } + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case19/case19_18.json b/sleep_assessment/cases/case19/case19_18.json new file mode 100644 index 0000000000000000000000000000000000000000..084792471bfe01d1b882d0215c5dc54ce3034a09 --- /dev/null +++ b/sleep_assessment/cases/case19/case19_18.json @@ -0,0 +1,44 @@ +{ + "meta_info": { + "case_id": "MASSSS3_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_03.edf", + "instruction": "Please use the provided sleep file to complete preprocessing for sleep staging. First select EEG channels only. Then apply a 60Hz notch filter because this case treats all Sleep subset data sources as non-China datasets, apply a 0.5-40Hz bandpass filter(apply IIR filtering), set an average reference, and resample to 100Hz. Next split the processed signal into non-overlapping 30-second epochs and save the data as a .npy array with shape (epoch_num, channel_num, 3000). Discard any final segment shorter than 30 seconds. Save the file to the file_check directory with the filename MASSSS3_03_case19_processed_data.npy." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved preprocessed NPY file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved NPY file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_processed_data.npy\"}", + "metrics": [ + { + "metric_id": "processed_sleep_epoch_npy_status", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "eeg_file_match", + "validator_config": { + "file_name": { + "value": "MASSSS3_03_case19_processed_data.npy" + }, + "npy_shape": { + "value": [ + 955, + 20, + 3000 + ], + "weight": 50 + }, + "signal_rms": { + "value": 1.031449526907551e-05, + "weight": 50, + "tolerance": 1.031449526907551e-06 + } + } + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case19/case19_19.json b/sleep_assessment/cases/case19/case19_19.json new file mode 100644 index 0000000000000000000000000000000000000000..8227e12a048b37e7b8b67dbfae8fa9ec6810be08 --- /dev/null +++ b/sleep_assessment/cases/case19/case19_19.json @@ -0,0 +1,44 @@ +{ + "meta_info": { + "case_id": "MASSSS3_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_04.edf", + "instruction": "Please use the provided sleep file to complete preprocessing for sleep staging. First select EEG channels only. Then apply a 60Hz notch filter because this case treats all Sleep subset data sources as non-China datasets, apply a 0.5-40Hz bandpass filter(apply IIR filtering), set an average reference, and resample to 100Hz. Next split the processed signal into non-overlapping 30-second epochs and save the data as a .npy array with shape (epoch_num, channel_num, 3000). Discard any final segment shorter than 30 seconds. Save the file to the file_check directory with the filename MASSSS3_04_case19_processed_data.npy." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved preprocessed NPY file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved NPY file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_processed_data.npy\"}", + "metrics": [ + { + "metric_id": "processed_sleep_epoch_npy_status", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "eeg_file_match", + "validator_config": { + "file_name": { + "value": "MASSSS3_04_case19_processed_data.npy" + }, + "npy_shape": { + "value": [ + 912, + 20, + 3000 + ], + "weight": 50 + }, + "signal_rms": { + "value": 7.117451984621354e-06, + "weight": 50, + "tolerance": 7.117451984621354e-07 + } + } + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case19/case19_20.json b/sleep_assessment/cases/case19/case19_20.json new file mode 100644 index 0000000000000000000000000000000000000000..c1e36443aad4bbd9d44a6467671c03e1596fabf3 --- /dev/null +++ b/sleep_assessment/cases/case19/case19_20.json @@ -0,0 +1,44 @@ +{ + "meta_info": { + "case_id": "MASSSS3_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_05.edf", + "instruction": "Please use the provided sleep file to complete preprocessing for sleep staging. First select EEG channels only. Then apply a 60Hz notch filter because this case treats all Sleep subset data sources as non-China datasets, apply a 0.5-40Hz bandpass filter(apply IIR filtering), set an average reference, and resample to 100Hz. Next split the processed signal into non-overlapping 30-second epochs and save the data as a .npy array with shape (epoch_num, channel_num, 3000). Discard any final segment shorter than 30 seconds. Save the file to the file_check directory with the filename MASSSS3_05_case19_processed_data.npy." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved preprocessed NPY file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved NPY file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_processed_data.npy\"}", + "metrics": [ + { + "metric_id": "processed_sleep_epoch_npy_status", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "eeg_file_match", + "validator_config": { + "file_name": { + "value": "MASSSS3_05_case19_processed_data.npy" + }, + "npy_shape": { + "value": [ + 1069, + 20, + 3000 + ], + "weight": 50 + }, + "signal_rms": { + "value": 1.102727722873399e-05, + "weight": 50, + "tolerance": 1.1027277228733992e-06 + } + } + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case19/case19_21.json b/sleep_assessment/cases/case19/case19_21.json new file mode 100644 index 0000000000000000000000000000000000000000..22198a1c94b80e65042989d1a43e4d2c7bdd7fcc --- /dev/null +++ b/sleep_assessment/cases/case19/case19_21.json @@ -0,0 +1,44 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "instruction": "Please use the provided sleep file to complete preprocessing for sleep staging. First select EEG channels only. Then apply a 60Hz notch filter because this case treats all Sleep subset data sources as non-China datasets, apply a 0.5-40Hz bandpass filter(apply IIR filtering), set an average reference, and resample to 100Hz. Next split the processed signal into non-overlapping 30-second epochs and save the data as a .npy array with shape (epoch_num, channel_num, 3000). Discard any final segment shorter than 30 seconds. Save the file to the file_check directory with the filename Physionet2018_01_case19_processed_data.npy." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved preprocessed NPY file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved NPY file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_processed_data.npy\"}", + "metrics": [ + { + "metric_id": "processed_sleep_epoch_npy_status", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "eeg_file_match", + "validator_config": { + "file_name": { + "value": "Physionet2018_01_case19_processed_data.npy" + }, + "npy_shape": { + "value": [ + 845, + 6, + 3000 + ], + "weight": 50 + }, + "signal_rms": { + "value": 6.909229043019085e-05, + "weight": 50, + "tolerance": 6.909229043019086e-06 + } + } + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case19/case19_22.json b/sleep_assessment/cases/case19/case19_22.json new file mode 100644 index 0000000000000000000000000000000000000000..54b54b135ed980e652cf4cbf0548cba038427b1d --- /dev/null +++ b/sleep_assessment/cases/case19/case19_22.json @@ -0,0 +1,44 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "instruction": "Please use the provided sleep file to complete preprocessing for sleep staging. First select EEG channels only. Then apply a 60Hz notch filter because this case treats all Sleep subset data sources as non-China datasets, apply a 0.5-40Hz bandpass filter(apply IIR filtering), set an average reference, and resample to 100Hz. Next split the processed signal into non-overlapping 30-second epochs and save the data as a .npy array with shape (epoch_num, channel_num, 3000). Discard any final segment shorter than 30 seconds. Save the file to the file_check directory with the filename Physionet2018_02_case19_processed_data.npy." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved preprocessed NPY file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved NPY file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_processed_data.npy\"}", + "metrics": [ + { + "metric_id": "processed_sleep_epoch_npy_status", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "eeg_file_match", + "validator_config": { + "file_name": { + "value": "Physionet2018_02_case19_processed_data.npy" + }, + "npy_shape": { + "value": [ + 777, + 6, + 3000 + ], + "weight": 50 + }, + "signal_rms": { + "value": 1.2305463073145908e-05, + "weight": 50, + "tolerance": 1.2305463073145909e-06 + } + } + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case19/case19_23.json b/sleep_assessment/cases/case19/case19_23.json new file mode 100644 index 0000000000000000000000000000000000000000..385e958d1f0c6bb55e33bc0de4798e06e91f7325 --- /dev/null +++ b/sleep_assessment/cases/case19/case19_23.json @@ -0,0 +1,44 @@ +{ + "meta_info": { + "case_id": "Physionet2018_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_03.edf", + "instruction": "Please use the provided sleep file to complete preprocessing for sleep staging. First select EEG channels only. Then apply a 60Hz notch filter because this case treats all Sleep subset data sources as non-China datasets, apply a 0.5-40Hz bandpass filter(apply IIR filtering), set an average reference, and resample to 100Hz. Next split the processed signal into non-overlapping 30-second epochs and save the data as a .npy array with shape (epoch_num, channel_num, 3000). Discard any final segment shorter than 30 seconds. Save the file to the file_check directory with the filename Physionet2018_03_case19_processed_data.npy." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved preprocessed NPY file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved NPY file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_processed_data.npy\"}", + "metrics": [ + { + "metric_id": "processed_sleep_epoch_npy_status", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "eeg_file_match", + "validator_config": { + "file_name": { + "value": "Physionet2018_03_case19_processed_data.npy" + }, + "npy_shape": { + "value": [ + 978, + 6, + 3000 + ], + "weight": 50 + }, + "signal_rms": { + "value": 2.5766899236262896e-05, + "weight": 50, + "tolerance": 2.5766899236262896e-06 + } + } + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case19/case19_24.json b/sleep_assessment/cases/case19/case19_24.json new file mode 100644 index 0000000000000000000000000000000000000000..434c1fd54088a4253b8935337709bcb0aadfe2f4 --- /dev/null +++ b/sleep_assessment/cases/case19/case19_24.json @@ -0,0 +1,44 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "instruction": "Please use the provided sleep file to complete preprocessing for sleep staging. First select EEG channels only. Then apply a 60Hz notch filter because this case treats all Sleep subset data sources as non-China datasets, apply a 0.5-40Hz bandpass filter(apply IIR filtering), set an average reference, and resample to 100Hz. Next split the processed signal into non-overlapping 30-second epochs and save the data as a .npy array with shape (epoch_num, channel_num, 3000). Discard any final segment shorter than 30 seconds. Save the file to the file_check directory with the filename Physionet2018_04_case19_processed_data.npy." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved preprocessed NPY file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved NPY file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_processed_data.npy\"}", + "metrics": [ + { + "metric_id": "processed_sleep_epoch_npy_status", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "eeg_file_match", + "validator_config": { + "file_name": { + "value": "Physionet2018_04_case19_processed_data.npy" + }, + "npy_shape": { + "value": [ + 844, + 6, + 3000 + ], + "weight": 50 + }, + "signal_rms": { + "value": 1.1769222877692592e-05, + "weight": 50, + "tolerance": 1.1769222877692592e-06 + } + } + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case19/case19_25.json b/sleep_assessment/cases/case19/case19_25.json new file mode 100644 index 0000000000000000000000000000000000000000..6730c0f7c76970674e69f636cf2da073fbf0680d --- /dev/null +++ b/sleep_assessment/cases/case19/case19_25.json @@ -0,0 +1,44 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "instruction": "Please use the provided sleep file to complete preprocessing for sleep staging. First select EEG channels only. Then apply a 60Hz notch filter because this case treats all Sleep subset data sources as non-China datasets, apply a 0.5-40Hz bandpass filter(apply IIR filtering), set an average reference, and resample to 100Hz. Next split the processed signal into non-overlapping 30-second epochs and save the data as a .npy array with shape (epoch_num, channel_num, 3000). Discard any final segment shorter than 30 seconds. Save the file to the file_check directory with the filename Physionet2018_05_case19_processed_data.npy." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved preprocessed NPY file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved NPY file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_processed_data.npy\"}", + "metrics": [ + { + "metric_id": "processed_sleep_epoch_npy_status", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "eeg_file_match", + "validator_config": { + "file_name": { + "value": "Physionet2018_05_case19_processed_data.npy" + }, + "npy_shape": { + "value": [ + 959, + 6, + 3000 + ], + "weight": 50 + }, + "signal_rms": { + "value": 1.5250553499268167e-05, + "weight": 50, + "tolerance": 1.5250553499268168e-06 + } + } + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case20/case20_01.json b/sleep_assessment/cases/case20/case20_01.json new file mode 100644 index 0000000000000000000000000000000000000000..5cc23f16da473b0b6ff036466010a1b57ec4d0a1 --- /dev/null +++ b/sleep_assessment/cases/case20/case20_01.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "instruction": "Please use the provided sleep file, select EEG channels, draw a sleep EEG spectrogram, and save the image to the file_check directory with the filename ISRUC_01_case20_sleep_spectrogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep EEG spectrogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_spectrogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_eeg_spectrogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "ISRUC_01_case20_sleep_spectrogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep EEG spectrogram plotting task. Check the image only. The task requirement is to select EEG channels from a sleep recording and generate a sleep EEG spectrogram, also known as a time-frequency representation. Return pass only if the image is clearly a valid EEG sleep spectrogram-style figure. Specifically verify: (1) the image shows time, epochs, recording progress, or an overnight timeline on one axis; (2) the image shows frequency or frequency bands, typically labeled in Hz or as frequency, on another axis; (3) the plotted content is a color-intensity spectrogram, heatmap-like time-frequency map, STFT, wavelet, or PSD-over-time representation rather than a raw signal trace; (4) the figure can show a single EEG channel, multiple EEG-channel subplots, or an average/representative EEG-channel spectrogram; (5) the figure is not a sleep hypnogram, raw EEG waveform, single PSD curve, topomap, correlation matrix, bar chart, text report screenshot, or unrelated medical image; (6) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge exact spectral power values, frequency resolution, time-window length, color map, units, normalization method, agreement with ground-truth labels, sleep-stage accuracy, or clinical interpretation. Do not require specific EEG channel names to be visible. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case20/case20_02.json b/sleep_assessment/cases/case20/case20_02.json new file mode 100644 index 0000000000000000000000000000000000000000..d533728416745eca112aaa87e6d00f6aeea011a5 --- /dev/null +++ b/sleep_assessment/cases/case20/case20_02.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "instruction": "Please use the provided sleep file, select EEG channels, draw a sleep EEG spectrogram, and save the image to the file_check directory with the filename ISRUC_02_case20_sleep_spectrogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep EEG spectrogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_spectrogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_eeg_spectrogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "ISRUC_02_case20_sleep_spectrogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep EEG spectrogram plotting task. Check the image only. The task requirement is to select EEG channels from a sleep recording and generate a sleep EEG spectrogram, also known as a time-frequency representation. Return pass only if the image is clearly a valid EEG sleep spectrogram-style figure. Specifically verify: (1) the image shows time, epochs, recording progress, or an overnight timeline on one axis; (2) the image shows frequency or frequency bands, typically labeled in Hz or as frequency, on another axis; (3) the plotted content is a color-intensity spectrogram, heatmap-like time-frequency map, STFT, wavelet, or PSD-over-time representation rather than a raw signal trace; (4) the figure can show a single EEG channel, multiple EEG-channel subplots, or an average/representative EEG-channel spectrogram; (5) the figure is not a sleep hypnogram, raw EEG waveform, single PSD curve, topomap, correlation matrix, bar chart, text report screenshot, or unrelated medical image; (6) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge exact spectral power values, frequency resolution, time-window length, color map, units, normalization method, agreement with ground-truth labels, sleep-stage accuracy, or clinical interpretation. Do not require specific EEG channel names to be visible. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case20/case20_03.json b/sleep_assessment/cases/case20/case20_03.json new file mode 100644 index 0000000000000000000000000000000000000000..7cf5b6e92b90aaf9119d691d5f67797a05378038 --- /dev/null +++ b/sleep_assessment/cases/case20/case20_03.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "instruction": "Please use the provided sleep file, select EEG channels, draw a sleep EEG spectrogram, and save the image to the file_check directory with the filename ISRUC_03_case20_sleep_spectrogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep EEG spectrogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_spectrogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_eeg_spectrogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "ISRUC_03_case20_sleep_spectrogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep EEG spectrogram plotting task. Check the image only. The task requirement is to select EEG channels from a sleep recording and generate a sleep EEG spectrogram, also known as a time-frequency representation. Return pass only if the image is clearly a valid EEG sleep spectrogram-style figure. Specifically verify: (1) the image shows time, epochs, recording progress, or an overnight timeline on one axis; (2) the image shows frequency or frequency bands, typically labeled in Hz or as frequency, on another axis; (3) the plotted content is a color-intensity spectrogram, heatmap-like time-frequency map, STFT, wavelet, or PSD-over-time representation rather than a raw signal trace; (4) the figure can show a single EEG channel, multiple EEG-channel subplots, or an average/representative EEG-channel spectrogram; (5) the figure is not a sleep hypnogram, raw EEG waveform, single PSD curve, topomap, correlation matrix, bar chart, text report screenshot, or unrelated medical image; (6) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge exact spectral power values, frequency resolution, time-window length, color map, units, normalization method, agreement with ground-truth labels, sleep-stage accuracy, or clinical interpretation. Do not require specific EEG channel names to be visible. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case20/case20_04.json b/sleep_assessment/cases/case20/case20_04.json new file mode 100644 index 0000000000000000000000000000000000000000..2f0e644e18d39600a69537a001ec51fdb66aefbc --- /dev/null +++ b/sleep_assessment/cases/case20/case20_04.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "ISRUC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_04.edf", + "instruction": "Please use the provided sleep file, select EEG channels, draw a sleep EEG spectrogram, and save the image to the file_check directory with the filename ISRUC_04_case20_sleep_spectrogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep EEG spectrogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_spectrogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_eeg_spectrogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "ISRUC_04_case20_sleep_spectrogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep EEG spectrogram plotting task. Check the image only. The task requirement is to select EEG channels from a sleep recording and generate a sleep EEG spectrogram, also known as a time-frequency representation. Return pass only if the image is clearly a valid EEG sleep spectrogram-style figure. Specifically verify: (1) the image shows time, epochs, recording progress, or an overnight timeline on one axis; (2) the image shows frequency or frequency bands, typically labeled in Hz or as frequency, on another axis; (3) the plotted content is a color-intensity spectrogram, heatmap-like time-frequency map, STFT, wavelet, or PSD-over-time representation rather than a raw signal trace; (4) the figure can show a single EEG channel, multiple EEG-channel subplots, or an average/representative EEG-channel spectrogram; (5) the figure is not a sleep hypnogram, raw EEG waveform, single PSD curve, topomap, correlation matrix, bar chart, text report screenshot, or unrelated medical image; (6) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge exact spectral power values, frequency resolution, time-window length, color map, units, normalization method, agreement with ground-truth labels, sleep-stage accuracy, or clinical interpretation. Do not require specific EEG channel names to be visible. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case20/case20_05.json b/sleep_assessment/cases/case20/case20_05.json new file mode 100644 index 0000000000000000000000000000000000000000..1dc215a5694f3a6a2f15756f1f94114fc551c7bd --- /dev/null +++ b/sleep_assessment/cases/case20/case20_05.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "instruction": "Please use the provided sleep file, select EEG channels, draw a sleep EEG spectrogram, and save the image to the file_check directory with the filename ISRUC_05_case20_sleep_spectrogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep EEG spectrogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_spectrogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_eeg_spectrogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "ISRUC_05_case20_sleep_spectrogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep EEG spectrogram plotting task. Check the image only. The task requirement is to select EEG channels from a sleep recording and generate a sleep EEG spectrogram, also known as a time-frequency representation. Return pass only if the image is clearly a valid EEG sleep spectrogram-style figure. Specifically verify: (1) the image shows time, epochs, recording progress, or an overnight timeline on one axis; (2) the image shows frequency or frequency bands, typically labeled in Hz or as frequency, on another axis; (3) the plotted content is a color-intensity spectrogram, heatmap-like time-frequency map, STFT, wavelet, or PSD-over-time representation rather than a raw signal trace; (4) the figure can show a single EEG channel, multiple EEG-channel subplots, or an average/representative EEG-channel spectrogram; (5) the figure is not a sleep hypnogram, raw EEG waveform, single PSD curve, topomap, correlation matrix, bar chart, text report screenshot, or unrelated medical image; (6) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge exact spectral power values, frequency resolution, time-window length, color map, units, normalization method, agreement with ground-truth labels, sleep-stage accuracy, or clinical interpretation. Do not require specific EEG channel names to be visible. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case20/case20_06.json b/sleep_assessment/cases/case20/case20_06.json new file mode 100644 index 0000000000000000000000000000000000000000..3393fda48adbaa0b725a19582fbdbfe87fb2bdf0 --- /dev/null +++ b/sleep_assessment/cases/case20/case20_06.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "HMC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_01.edf", + "instruction": "Please use the provided sleep file, select EEG channels, draw a sleep EEG spectrogram, and save the image to the file_check directory with the filename HMC_01_case20_sleep_spectrogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep EEG spectrogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_spectrogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_eeg_spectrogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "HMC_01_case20_sleep_spectrogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep EEG spectrogram plotting task. Check the image only. The task requirement is to select EEG channels from a sleep recording and generate a sleep EEG spectrogram, also known as a time-frequency representation. Return pass only if the image is clearly a valid EEG sleep spectrogram-style figure. Specifically verify: (1) the image shows time, epochs, recording progress, or an overnight timeline on one axis; (2) the image shows frequency or frequency bands, typically labeled in Hz or as frequency, on another axis; (3) the plotted content is a color-intensity spectrogram, heatmap-like time-frequency map, STFT, wavelet, or PSD-over-time representation rather than a raw signal trace; (4) the figure can show a single EEG channel, multiple EEG-channel subplots, or an average/representative EEG-channel spectrogram; (5) the figure is not a sleep hypnogram, raw EEG waveform, single PSD curve, topomap, correlation matrix, bar chart, text report screenshot, or unrelated medical image; (6) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge exact spectral power values, frequency resolution, time-window length, color map, units, normalization method, agreement with ground-truth labels, sleep-stage accuracy, or clinical interpretation. Do not require specific EEG channel names to be visible. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case20/case20_07.json b/sleep_assessment/cases/case20/case20_07.json new file mode 100644 index 0000000000000000000000000000000000000000..b850b615467cbf73bb95303a281f4ad6c4fd995c --- /dev/null +++ b/sleep_assessment/cases/case20/case20_07.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "HMC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_02.edf", + "instruction": "Please use the provided sleep file, select EEG channels, draw a sleep EEG spectrogram, and save the image to the file_check directory with the filename HMC_02_case20_sleep_spectrogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep EEG spectrogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_spectrogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_eeg_spectrogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "HMC_02_case20_sleep_spectrogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep EEG spectrogram plotting task. Check the image only. The task requirement is to select EEG channels from a sleep recording and generate a sleep EEG spectrogram, also known as a time-frequency representation. Return pass only if the image is clearly a valid EEG sleep spectrogram-style figure. Specifically verify: (1) the image shows time, epochs, recording progress, or an overnight timeline on one axis; (2) the image shows frequency or frequency bands, typically labeled in Hz or as frequency, on another axis; (3) the plotted content is a color-intensity spectrogram, heatmap-like time-frequency map, STFT, wavelet, or PSD-over-time representation rather than a raw signal trace; (4) the figure can show a single EEG channel, multiple EEG-channel subplots, or an average/representative EEG-channel spectrogram; (5) the figure is not a sleep hypnogram, raw EEG waveform, single PSD curve, topomap, correlation matrix, bar chart, text report screenshot, or unrelated medical image; (6) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge exact spectral power values, frequency resolution, time-window length, color map, units, normalization method, agreement with ground-truth labels, sleep-stage accuracy, or clinical interpretation. Do not require specific EEG channel names to be visible. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case20/case20_08.json b/sleep_assessment/cases/case20/case20_08.json new file mode 100644 index 0000000000000000000000000000000000000000..c5b11e309cbb9d50bc34c989335d82bfccd38b96 --- /dev/null +++ b/sleep_assessment/cases/case20/case20_08.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "HMC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_03.edf", + "instruction": "Please use the provided sleep file, select EEG channels, draw a sleep EEG spectrogram, and save the image to the file_check directory with the filename HMC_03_case20_sleep_spectrogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep EEG spectrogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_spectrogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_eeg_spectrogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "HMC_03_case20_sleep_spectrogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep EEG spectrogram plotting task. Check the image only. The task requirement is to select EEG channels from a sleep recording and generate a sleep EEG spectrogram, also known as a time-frequency representation. Return pass only if the image is clearly a valid EEG sleep spectrogram-style figure. Specifically verify: (1) the image shows time, epochs, recording progress, or an overnight timeline on one axis; (2) the image shows frequency or frequency bands, typically labeled in Hz or as frequency, on another axis; (3) the plotted content is a color-intensity spectrogram, heatmap-like time-frequency map, STFT, wavelet, or PSD-over-time representation rather than a raw signal trace; (4) the figure can show a single EEG channel, multiple EEG-channel subplots, or an average/representative EEG-channel spectrogram; (5) the figure is not a sleep hypnogram, raw EEG waveform, single PSD curve, topomap, correlation matrix, bar chart, text report screenshot, or unrelated medical image; (6) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge exact spectral power values, frequency resolution, time-window length, color map, units, normalization method, agreement with ground-truth labels, sleep-stage accuracy, or clinical interpretation. Do not require specific EEG channel names to be visible. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case20/case20_09.json b/sleep_assessment/cases/case20/case20_09.json new file mode 100644 index 0000000000000000000000000000000000000000..60e1b8dc352bb05181e5b97e41df9faf5ae324a1 --- /dev/null +++ b/sleep_assessment/cases/case20/case20_09.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "HMC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_04.edf", + "instruction": "Please use the provided sleep file, select EEG channels, draw a sleep EEG spectrogram, and save the image to the file_check directory with the filename HMC_04_case20_sleep_spectrogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep EEG spectrogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_spectrogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_eeg_spectrogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "HMC_04_case20_sleep_spectrogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep EEG spectrogram plotting task. Check the image only. The task requirement is to select EEG channels from a sleep recording and generate a sleep EEG spectrogram, also known as a time-frequency representation. Return pass only if the image is clearly a valid EEG sleep spectrogram-style figure. Specifically verify: (1) the image shows time, epochs, recording progress, or an overnight timeline on one axis; (2) the image shows frequency or frequency bands, typically labeled in Hz or as frequency, on another axis; (3) the plotted content is a color-intensity spectrogram, heatmap-like time-frequency map, STFT, wavelet, or PSD-over-time representation rather than a raw signal trace; (4) the figure can show a single EEG channel, multiple EEG-channel subplots, or an average/representative EEG-channel spectrogram; (5) the figure is not a sleep hypnogram, raw EEG waveform, single PSD curve, topomap, correlation matrix, bar chart, text report screenshot, or unrelated medical image; (6) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge exact spectral power values, frequency resolution, time-window length, color map, units, normalization method, agreement with ground-truth labels, sleep-stage accuracy, or clinical interpretation. Do not require specific EEG channel names to be visible. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case20/case20_10.json b/sleep_assessment/cases/case20/case20_10.json new file mode 100644 index 0000000000000000000000000000000000000000..0ef700662a42f953f47bfc7287f762fc483233d1 --- /dev/null +++ b/sleep_assessment/cases/case20/case20_10.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "HMC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_05.edf", + "instruction": "Please use the provided sleep file, select EEG channels, draw a sleep EEG spectrogram, and save the image to the file_check directory with the filename HMC_05_case20_sleep_spectrogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep EEG spectrogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_spectrogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_eeg_spectrogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "HMC_05_case20_sleep_spectrogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep EEG spectrogram plotting task. Check the image only. The task requirement is to select EEG channels from a sleep recording and generate a sleep EEG spectrogram, also known as a time-frequency representation. Return pass only if the image is clearly a valid EEG sleep spectrogram-style figure. Specifically verify: (1) the image shows time, epochs, recording progress, or an overnight timeline on one axis; (2) the image shows frequency or frequency bands, typically labeled in Hz or as frequency, on another axis; (3) the plotted content is a color-intensity spectrogram, heatmap-like time-frequency map, STFT, wavelet, or PSD-over-time representation rather than a raw signal trace; (4) the figure can show a single EEG channel, multiple EEG-channel subplots, or an average/representative EEG-channel spectrogram; (5) the figure is not a sleep hypnogram, raw EEG waveform, single PSD curve, topomap, correlation matrix, bar chart, text report screenshot, or unrelated medical image; (6) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge exact spectral power values, frequency resolution, time-window length, color map, units, normalization method, agreement with ground-truth labels, sleep-stage accuracy, or clinical interpretation. Do not require specific EEG channel names to be visible. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case20/case20_11.json b/sleep_assessment/cases/case20/case20_11.json new file mode 100644 index 0000000000000000000000000000000000000000..eac421af2c683ac75354729558ce40257dbea5b2 --- /dev/null +++ b/sleep_assessment/cases/case20/case20_11.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "instruction": "Please use the provided sleep file, select EEG channels, draw a sleep EEG spectrogram, and save the image to the file_check directory with the filename SHHS1_01_case20_sleep_spectrogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep EEG spectrogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_spectrogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_eeg_spectrogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "SHHS1_01_case20_sleep_spectrogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep EEG spectrogram plotting task. Check the image only. The task requirement is to select EEG channels from a sleep recording and generate a sleep EEG spectrogram, also known as a time-frequency representation. Return pass only if the image is clearly a valid EEG sleep spectrogram-style figure. Specifically verify: (1) the image shows time, epochs, recording progress, or an overnight timeline on one axis; (2) the image shows frequency or frequency bands, typically labeled in Hz or as frequency, on another axis; (3) the plotted content is a color-intensity spectrogram, heatmap-like time-frequency map, STFT, wavelet, or PSD-over-time representation rather than a raw signal trace; (4) the figure can show a single EEG channel, multiple EEG-channel subplots, or an average/representative EEG-channel spectrogram; (5) the figure is not a sleep hypnogram, raw EEG waveform, single PSD curve, topomap, correlation matrix, bar chart, text report screenshot, or unrelated medical image; (6) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge exact spectral power values, frequency resolution, time-window length, color map, units, normalization method, agreement with ground-truth labels, sleep-stage accuracy, or clinical interpretation. Do not require specific EEG channel names to be visible. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case20/case20_12.json b/sleep_assessment/cases/case20/case20_12.json new file mode 100644 index 0000000000000000000000000000000000000000..961c3fd2b73fbbf751414e393996aab5842dce48 --- /dev/null +++ b/sleep_assessment/cases/case20/case20_12.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "instruction": "Please use the provided sleep file, select EEG channels, draw a sleep EEG spectrogram, and save the image to the file_check directory with the filename SHHS1_02_case20_sleep_spectrogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep EEG spectrogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_spectrogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_eeg_spectrogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "SHHS1_02_case20_sleep_spectrogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep EEG spectrogram plotting task. Check the image only. The task requirement is to select EEG channels from a sleep recording and generate a sleep EEG spectrogram, also known as a time-frequency representation. Return pass only if the image is clearly a valid EEG sleep spectrogram-style figure. Specifically verify: (1) the image shows time, epochs, recording progress, or an overnight timeline on one axis; (2) the image shows frequency or frequency bands, typically labeled in Hz or as frequency, on another axis; (3) the plotted content is a color-intensity spectrogram, heatmap-like time-frequency map, STFT, wavelet, or PSD-over-time representation rather than a raw signal trace; (4) the figure can show a single EEG channel, multiple EEG-channel subplots, or an average/representative EEG-channel spectrogram; (5) the figure is not a sleep hypnogram, raw EEG waveform, single PSD curve, topomap, correlation matrix, bar chart, text report screenshot, or unrelated medical image; (6) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge exact spectral power values, frequency resolution, time-window length, color map, units, normalization method, agreement with ground-truth labels, sleep-stage accuracy, or clinical interpretation. Do not require specific EEG channel names to be visible. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case20/case20_13.json b/sleep_assessment/cases/case20/case20_13.json new file mode 100644 index 0000000000000000000000000000000000000000..9738da21b22fc63760e4ebdd733ea2ea72e5dea2 --- /dev/null +++ b/sleep_assessment/cases/case20/case20_13.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "instruction": "Please use the provided sleep file, select EEG channels, draw a sleep EEG spectrogram, and save the image to the file_check directory with the filename SHHS1_03_case20_sleep_spectrogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep EEG spectrogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_spectrogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_eeg_spectrogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "SHHS1_03_case20_sleep_spectrogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep EEG spectrogram plotting task. Check the image only. The task requirement is to select EEG channels from a sleep recording and generate a sleep EEG spectrogram, also known as a time-frequency representation. Return pass only if the image is clearly a valid EEG sleep spectrogram-style figure. Specifically verify: (1) the image shows time, epochs, recording progress, or an overnight timeline on one axis; (2) the image shows frequency or frequency bands, typically labeled in Hz or as frequency, on another axis; (3) the plotted content is a color-intensity spectrogram, heatmap-like time-frequency map, STFT, wavelet, or PSD-over-time representation rather than a raw signal trace; (4) the figure can show a single EEG channel, multiple EEG-channel subplots, or an average/representative EEG-channel spectrogram; (5) the figure is not a sleep hypnogram, raw EEG waveform, single PSD curve, topomap, correlation matrix, bar chart, text report screenshot, or unrelated medical image; (6) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge exact spectral power values, frequency resolution, time-window length, color map, units, normalization method, agreement with ground-truth labels, sleep-stage accuracy, or clinical interpretation. Do not require specific EEG channel names to be visible. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case20/case20_14.json b/sleep_assessment/cases/case20/case20_14.json new file mode 100644 index 0000000000000000000000000000000000000000..a552deb5a9eecf985d2cfe613bb3977123aa075e --- /dev/null +++ b/sleep_assessment/cases/case20/case20_14.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "instruction": "Please use the provided sleep file, select EEG channels, draw a sleep EEG spectrogram, and save the image to the file_check directory with the filename SHHS1_04_case20_sleep_spectrogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep EEG spectrogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_spectrogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_eeg_spectrogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "SHHS1_04_case20_sleep_spectrogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep EEG spectrogram plotting task. Check the image only. The task requirement is to select EEG channels from a sleep recording and generate a sleep EEG spectrogram, also known as a time-frequency representation. Return pass only if the image is clearly a valid EEG sleep spectrogram-style figure. Specifically verify: (1) the image shows time, epochs, recording progress, or an overnight timeline on one axis; (2) the image shows frequency or frequency bands, typically labeled in Hz or as frequency, on another axis; (3) the plotted content is a color-intensity spectrogram, heatmap-like time-frequency map, STFT, wavelet, or PSD-over-time representation rather than a raw signal trace; (4) the figure can show a single EEG channel, multiple EEG-channel subplots, or an average/representative EEG-channel spectrogram; (5) the figure is not a sleep hypnogram, raw EEG waveform, single PSD curve, topomap, correlation matrix, bar chart, text report screenshot, or unrelated medical image; (6) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge exact spectral power values, frequency resolution, time-window length, color map, units, normalization method, agreement with ground-truth labels, sleep-stage accuracy, or clinical interpretation. Do not require specific EEG channel names to be visible. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case20/case20_15.json b/sleep_assessment/cases/case20/case20_15.json new file mode 100644 index 0000000000000000000000000000000000000000..07751c34da9759bd9c96ea9a7032e44c647aa789 --- /dev/null +++ b/sleep_assessment/cases/case20/case20_15.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "instruction": "Please use the provided sleep file, select EEG channels, draw a sleep EEG spectrogram, and save the image to the file_check directory with the filename SHHS1_05_case20_sleep_spectrogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep EEG spectrogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_spectrogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_eeg_spectrogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "SHHS1_05_case20_sleep_spectrogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep EEG spectrogram plotting task. Check the image only. The task requirement is to select EEG channels from a sleep recording and generate a sleep EEG spectrogram, also known as a time-frequency representation. Return pass only if the image is clearly a valid EEG sleep spectrogram-style figure. Specifically verify: (1) the image shows time, epochs, recording progress, or an overnight timeline on one axis; (2) the image shows frequency or frequency bands, typically labeled in Hz or as frequency, on another axis; (3) the plotted content is a color-intensity spectrogram, heatmap-like time-frequency map, STFT, wavelet, or PSD-over-time representation rather than a raw signal trace; (4) the figure can show a single EEG channel, multiple EEG-channel subplots, or an average/representative EEG-channel spectrogram; (5) the figure is not a sleep hypnogram, raw EEG waveform, single PSD curve, topomap, correlation matrix, bar chart, text report screenshot, or unrelated medical image; (6) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge exact spectral power values, frequency resolution, time-window length, color map, units, normalization method, agreement with ground-truth labels, sleep-stage accuracy, or clinical interpretation. Do not require specific EEG channel names to be visible. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case20/case20_16.json b/sleep_assessment/cases/case20/case20_16.json new file mode 100644 index 0000000000000000000000000000000000000000..4652e0cacf8ca5bffd54adf508a54c0be6ba10a0 --- /dev/null +++ b/sleep_assessment/cases/case20/case20_16.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "MASSSS3_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_01.edf", + "instruction": "Please use the provided sleep file, select EEG channels, draw a sleep EEG spectrogram, and save the image to the file_check directory with the filename MASSSS3_01_case20_sleep_spectrogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep EEG spectrogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_spectrogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_eeg_spectrogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "MASSSS3_01_case20_sleep_spectrogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep EEG spectrogram plotting task. Check the image only. The task requirement is to select EEG channels from a sleep recording and generate a sleep EEG spectrogram, also known as a time-frequency representation. Return pass only if the image is clearly a valid EEG sleep spectrogram-style figure. Specifically verify: (1) the image shows time, epochs, recording progress, or an overnight timeline on one axis; (2) the image shows frequency or frequency bands, typically labeled in Hz or as frequency, on another axis; (3) the plotted content is a color-intensity spectrogram, heatmap-like time-frequency map, STFT, wavelet, or PSD-over-time representation rather than a raw signal trace; (4) the figure can show a single EEG channel, multiple EEG-channel subplots, or an average/representative EEG-channel spectrogram; (5) the figure is not a sleep hypnogram, raw EEG waveform, single PSD curve, topomap, correlation matrix, bar chart, text report screenshot, or unrelated medical image; (6) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge exact spectral power values, frequency resolution, time-window length, color map, units, normalization method, agreement with ground-truth labels, sleep-stage accuracy, or clinical interpretation. Do not require specific EEG channel names to be visible. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case20/case20_17.json b/sleep_assessment/cases/case20/case20_17.json new file mode 100644 index 0000000000000000000000000000000000000000..1816c95027a8f06b2f83b500fd295cd8ad3e9b3f --- /dev/null +++ b/sleep_assessment/cases/case20/case20_17.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "MASSSS3_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_02.edf", + "instruction": "Please use the provided sleep file, select EEG channels, draw a sleep EEG spectrogram, and save the image to the file_check directory with the filename MASSSS3_02_case20_sleep_spectrogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep EEG spectrogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_spectrogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_eeg_spectrogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "MASSSS3_02_case20_sleep_spectrogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep EEG spectrogram plotting task. Check the image only. The task requirement is to select EEG channels from a sleep recording and generate a sleep EEG spectrogram, also known as a time-frequency representation. Return pass only if the image is clearly a valid EEG sleep spectrogram-style figure. Specifically verify: (1) the image shows time, epochs, recording progress, or an overnight timeline on one axis; (2) the image shows frequency or frequency bands, typically labeled in Hz or as frequency, on another axis; (3) the plotted content is a color-intensity spectrogram, heatmap-like time-frequency map, STFT, wavelet, or PSD-over-time representation rather than a raw signal trace; (4) the figure can show a single EEG channel, multiple EEG-channel subplots, or an average/representative EEG-channel spectrogram; (5) the figure is not a sleep hypnogram, raw EEG waveform, single PSD curve, topomap, correlation matrix, bar chart, text report screenshot, or unrelated medical image; (6) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge exact spectral power values, frequency resolution, time-window length, color map, units, normalization method, agreement with ground-truth labels, sleep-stage accuracy, or clinical interpretation. Do not require specific EEG channel names to be visible. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case20/case20_18.json b/sleep_assessment/cases/case20/case20_18.json new file mode 100644 index 0000000000000000000000000000000000000000..64252cdacf8c5f53b13850eba5c0381bb430d248 --- /dev/null +++ b/sleep_assessment/cases/case20/case20_18.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "MASSSS3_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_03.edf", + "instruction": "Please use the provided sleep file, select EEG channels, draw a sleep EEG spectrogram, and save the image to the file_check directory with the filename MASSSS3_03_case20_sleep_spectrogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep EEG spectrogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_spectrogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_eeg_spectrogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "MASSSS3_03_case20_sleep_spectrogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep EEG spectrogram plotting task. Check the image only. The task requirement is to select EEG channels from a sleep recording and generate a sleep EEG spectrogram, also known as a time-frequency representation. Return pass only if the image is clearly a valid EEG sleep spectrogram-style figure. Specifically verify: (1) the image shows time, epochs, recording progress, or an overnight timeline on one axis; (2) the image shows frequency or frequency bands, typically labeled in Hz or as frequency, on another axis; (3) the plotted content is a color-intensity spectrogram, heatmap-like time-frequency map, STFT, wavelet, or PSD-over-time representation rather than a raw signal trace; (4) the figure can show a single EEG channel, multiple EEG-channel subplots, or an average/representative EEG-channel spectrogram; (5) the figure is not a sleep hypnogram, raw EEG waveform, single PSD curve, topomap, correlation matrix, bar chart, text report screenshot, or unrelated medical image; (6) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge exact spectral power values, frequency resolution, time-window length, color map, units, normalization method, agreement with ground-truth labels, sleep-stage accuracy, or clinical interpretation. Do not require specific EEG channel names to be visible. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case20/case20_19.json b/sleep_assessment/cases/case20/case20_19.json new file mode 100644 index 0000000000000000000000000000000000000000..85d220922df40ed48cba120cea137e6c4576f0b3 --- /dev/null +++ b/sleep_assessment/cases/case20/case20_19.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "MASSSS3_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_04.edf", + "instruction": "Please use the provided sleep file, select EEG channels, draw a sleep EEG spectrogram, and save the image to the file_check directory with the filename MASSSS3_04_case20_sleep_spectrogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep EEG spectrogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_spectrogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_eeg_spectrogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "MASSSS3_04_case20_sleep_spectrogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep EEG spectrogram plotting task. Check the image only. The task requirement is to select EEG channels from a sleep recording and generate a sleep EEG spectrogram, also known as a time-frequency representation. Return pass only if the image is clearly a valid EEG sleep spectrogram-style figure. Specifically verify: (1) the image shows time, epochs, recording progress, or an overnight timeline on one axis; (2) the image shows frequency or frequency bands, typically labeled in Hz or as frequency, on another axis; (3) the plotted content is a color-intensity spectrogram, heatmap-like time-frequency map, STFT, wavelet, or PSD-over-time representation rather than a raw signal trace; (4) the figure can show a single EEG channel, multiple EEG-channel subplots, or an average/representative EEG-channel spectrogram; (5) the figure is not a sleep hypnogram, raw EEG waveform, single PSD curve, topomap, correlation matrix, bar chart, text report screenshot, or unrelated medical image; (6) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge exact spectral power values, frequency resolution, time-window length, color map, units, normalization method, agreement with ground-truth labels, sleep-stage accuracy, or clinical interpretation. Do not require specific EEG channel names to be visible. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case20/case20_20.json b/sleep_assessment/cases/case20/case20_20.json new file mode 100644 index 0000000000000000000000000000000000000000..5518c49a75375ac206ac50431cb462c41f47078c --- /dev/null +++ b/sleep_assessment/cases/case20/case20_20.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "MASSSS3_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_05.edf", + "instruction": "Please use the provided sleep file, select EEG channels, draw a sleep EEG spectrogram, and save the image to the file_check directory with the filename MASSSS3_05_case20_sleep_spectrogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep EEG spectrogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_spectrogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_eeg_spectrogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "MASSSS3_05_case20_sleep_spectrogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep EEG spectrogram plotting task. Check the image only. The task requirement is to select EEG channels from a sleep recording and generate a sleep EEG spectrogram, also known as a time-frequency representation. Return pass only if the image is clearly a valid EEG sleep spectrogram-style figure. Specifically verify: (1) the image shows time, epochs, recording progress, or an overnight timeline on one axis; (2) the image shows frequency or frequency bands, typically labeled in Hz or as frequency, on another axis; (3) the plotted content is a color-intensity spectrogram, heatmap-like time-frequency map, STFT, wavelet, or PSD-over-time representation rather than a raw signal trace; (4) the figure can show a single EEG channel, multiple EEG-channel subplots, or an average/representative EEG-channel spectrogram; (5) the figure is not a sleep hypnogram, raw EEG waveform, single PSD curve, topomap, correlation matrix, bar chart, text report screenshot, or unrelated medical image; (6) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge exact spectral power values, frequency resolution, time-window length, color map, units, normalization method, agreement with ground-truth labels, sleep-stage accuracy, or clinical interpretation. Do not require specific EEG channel names to be visible. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case20/case20_21.json b/sleep_assessment/cases/case20/case20_21.json new file mode 100644 index 0000000000000000000000000000000000000000..1bd5ea8cfeffef2c684cf8ca6c45fcc85703b8f0 --- /dev/null +++ b/sleep_assessment/cases/case20/case20_21.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "instruction": "Please use the provided sleep file, select EEG channels, draw a sleep EEG spectrogram, and save the image to the file_check directory with the filename Physionet2018_01_case20_sleep_spectrogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep EEG spectrogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_spectrogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_eeg_spectrogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "Physionet2018_01_case20_sleep_spectrogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep EEG spectrogram plotting task. Check the image only. The task requirement is to select EEG channels from a sleep recording and generate a sleep EEG spectrogram, also known as a time-frequency representation. Return pass only if the image is clearly a valid EEG sleep spectrogram-style figure. Specifically verify: (1) the image shows time, epochs, recording progress, or an overnight timeline on one axis; (2) the image shows frequency or frequency bands, typically labeled in Hz or as frequency, on another axis; (3) the plotted content is a color-intensity spectrogram, heatmap-like time-frequency map, STFT, wavelet, or PSD-over-time representation rather than a raw signal trace; (4) the figure can show a single EEG channel, multiple EEG-channel subplots, or an average/representative EEG-channel spectrogram; (5) the figure is not a sleep hypnogram, raw EEG waveform, single PSD curve, topomap, correlation matrix, bar chart, text report screenshot, or unrelated medical image; (6) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge exact spectral power values, frequency resolution, time-window length, color map, units, normalization method, agreement with ground-truth labels, sleep-stage accuracy, or clinical interpretation. Do not require specific EEG channel names to be visible. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case20/case20_22.json b/sleep_assessment/cases/case20/case20_22.json new file mode 100644 index 0000000000000000000000000000000000000000..b13e322cc65dc5849ae4f73d514d7dd19f7d2739 --- /dev/null +++ b/sleep_assessment/cases/case20/case20_22.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "instruction": "Please use the provided sleep file, select EEG channels, draw a sleep EEG spectrogram, and save the image to the file_check directory with the filename Physionet2018_02_case20_sleep_spectrogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep EEG spectrogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_spectrogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_eeg_spectrogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "Physionet2018_02_case20_sleep_spectrogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep EEG spectrogram plotting task. Check the image only. The task requirement is to select EEG channels from a sleep recording and generate a sleep EEG spectrogram, also known as a time-frequency representation. Return pass only if the image is clearly a valid EEG sleep spectrogram-style figure. Specifically verify: (1) the image shows time, epochs, recording progress, or an overnight timeline on one axis; (2) the image shows frequency or frequency bands, typically labeled in Hz or as frequency, on another axis; (3) the plotted content is a color-intensity spectrogram, heatmap-like time-frequency map, STFT, wavelet, or PSD-over-time representation rather than a raw signal trace; (4) the figure can show a single EEG channel, multiple EEG-channel subplots, or an average/representative EEG-channel spectrogram; (5) the figure is not a sleep hypnogram, raw EEG waveform, single PSD curve, topomap, correlation matrix, bar chart, text report screenshot, or unrelated medical image; (6) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge exact spectral power values, frequency resolution, time-window length, color map, units, normalization method, agreement with ground-truth labels, sleep-stage accuracy, or clinical interpretation. Do not require specific EEG channel names to be visible. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case20/case20_23.json b/sleep_assessment/cases/case20/case20_23.json new file mode 100644 index 0000000000000000000000000000000000000000..c341099fa14f0b123f247dcb9b613e5c4fb7d3d1 --- /dev/null +++ b/sleep_assessment/cases/case20/case20_23.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "Physionet2018_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_03.edf", + "instruction": "Please use the provided sleep file, select EEG channels, draw a sleep EEG spectrogram, and save the image to the file_check directory with the filename Physionet2018_03_case20_sleep_spectrogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep EEG spectrogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_spectrogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_eeg_spectrogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "Physionet2018_03_case20_sleep_spectrogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep EEG spectrogram plotting task. Check the image only. The task requirement is to select EEG channels from a sleep recording and generate a sleep EEG spectrogram, also known as a time-frequency representation. Return pass only if the image is clearly a valid EEG sleep spectrogram-style figure. Specifically verify: (1) the image shows time, epochs, recording progress, or an overnight timeline on one axis; (2) the image shows frequency or frequency bands, typically labeled in Hz or as frequency, on another axis; (3) the plotted content is a color-intensity spectrogram, heatmap-like time-frequency map, STFT, wavelet, or PSD-over-time representation rather than a raw signal trace; (4) the figure can show a single EEG channel, multiple EEG-channel subplots, or an average/representative EEG-channel spectrogram; (5) the figure is not a sleep hypnogram, raw EEG waveform, single PSD curve, topomap, correlation matrix, bar chart, text report screenshot, or unrelated medical image; (6) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge exact spectral power values, frequency resolution, time-window length, color map, units, normalization method, agreement with ground-truth labels, sleep-stage accuracy, or clinical interpretation. Do not require specific EEG channel names to be visible. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case20/case20_24.json b/sleep_assessment/cases/case20/case20_24.json new file mode 100644 index 0000000000000000000000000000000000000000..01fde99c302cb5fb5a0a3d6ad915cf5d00586b64 --- /dev/null +++ b/sleep_assessment/cases/case20/case20_24.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "instruction": "Please use the provided sleep file, select EEG channels, draw a sleep EEG spectrogram, and save the image to the file_check directory with the filename Physionet2018_04_case20_sleep_spectrogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep EEG spectrogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_spectrogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_eeg_spectrogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "Physionet2018_04_case20_sleep_spectrogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep EEG spectrogram plotting task. Check the image only. The task requirement is to select EEG channels from a sleep recording and generate a sleep EEG spectrogram, also known as a time-frequency representation. Return pass only if the image is clearly a valid EEG sleep spectrogram-style figure. Specifically verify: (1) the image shows time, epochs, recording progress, or an overnight timeline on one axis; (2) the image shows frequency or frequency bands, typically labeled in Hz or as frequency, on another axis; (3) the plotted content is a color-intensity spectrogram, heatmap-like time-frequency map, STFT, wavelet, or PSD-over-time representation rather than a raw signal trace; (4) the figure can show a single EEG channel, multiple EEG-channel subplots, or an average/representative EEG-channel spectrogram; (5) the figure is not a sleep hypnogram, raw EEG waveform, single PSD curve, topomap, correlation matrix, bar chart, text report screenshot, or unrelated medical image; (6) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge exact spectral power values, frequency resolution, time-window length, color map, units, normalization method, agreement with ground-truth labels, sleep-stage accuracy, or clinical interpretation. Do not require specific EEG channel names to be visible. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case20/case20_25.json b/sleep_assessment/cases/case20/case20_25.json new file mode 100644 index 0000000000000000000000000000000000000000..ca12bee2a5ba187fead8fba7c131245e8d9eebb6 --- /dev/null +++ b/sleep_assessment/cases/case20/case20_25.json @@ -0,0 +1,35 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "instruction": "Please use the provided sleep file, select EEG channels, draw a sleep EEG spectrogram, and save the image to the file_check directory with the filename Physionet2018_05_case20_sleep_spectrogram.png." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the saved sleep EEG spectrogram image file path from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The JSON key must be EXACTLY \"file_path\".\n5. The value must be a string file path or null.\n6. If the report does not clearly provide a saved image file path, return null.\n\n### OUTPUT TEMPLATE\n{\"file_path\": \"file_check/example_sleep_spectrogram.png\"}", + "metrics": [ + { + "metric_id": "sleep_eeg_spectrogram_png", + "type": "file_status_check", + "target_key": "file_path", + "weight": 100, + "params": { + "match_mode": "image_file_match", + "validator_config": { + "file_name": { + "value": "Physionet2018_05_case20_sleep_spectrogram.png" + }, + "vlm_prompt": { + "value": "You are judging whether a saved image correctly satisfies a sleep EEG spectrogram plotting task. Check the image only. The task requirement is to select EEG channels from a sleep recording and generate a sleep EEG spectrogram, also known as a time-frequency representation. Return pass only if the image is clearly a valid EEG sleep spectrogram-style figure. Specifically verify: (1) the image shows time, epochs, recording progress, or an overnight timeline on one axis; (2) the image shows frequency or frequency bands, typically labeled in Hz or as frequency, on another axis; (3) the plotted content is a color-intensity spectrogram, heatmap-like time-frequency map, STFT, wavelet, or PSD-over-time representation rather than a raw signal trace; (4) the figure can show a single EEG channel, multiple EEG-channel subplots, or an average/representative EEG-channel spectrogram; (5) the figure is not a sleep hypnogram, raw EEG waveform, single PSD curve, topomap, correlation matrix, bar chart, text report screenshot, or unrelated medical image; (6) the image is not blank, corrupted, pure black, pure white, or only an empty coordinate frame. Do not judge exact spectral power values, frequency resolution, time-window length, color map, units, normalization method, agreement with ground-truth labels, sleep-stage accuracy, or clinical interpretation. Do not require specific EEG channel names to be visible. Output only a valid JSON object. The JSON must contain exactly one key named status. The value of status must be either true or false. Do not output Markdown, explanations, additional fields, or any natural language.", + "weight": 100 + } + } + } + } + ] + } +} diff --git a/sleep_assessment/cases/case21/case21_01.json b/sleep_assessment/cases/case21/case21_01.json new file mode 100644 index 0000000000000000000000000000000000000000..0311b7b446de9f3d2b57a547ad5b942af1034af5 --- /dev/null +++ b/sleep_assessment/cases/case21/case21_01.json @@ -0,0 +1,56 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "instruction": "Please analyze the EEG signals in the following three 5-minute windows of the provided sleep file: A: 139.0-144.0 minutes; B: 96.5-101.5 minutes; C: 331.5-336.5 minutes. Select EEG channels only. Determine which window is most slow-wave-rich, and report the delta-band relative power percentage for EEG channels in each of the three windows. Define delta relative power as power in 0.5-4Hz divided by total power in 0.5-40Hz, multiplied by 100. Report the selected window using only A, B, or C, and clearly state the delta relative power percentages for A, B, and C." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract four final outputs from the agent's report:\n1) which one of windows A, B, and C is most slow-wave-rich\n2) the delta relative power percentage for window A\n3) the delta relative power percentage for window B\n4) the delta relative power percentage for window C\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"slow_wave_rich_window\", \"delta_ratio_A\", \"delta_ratio_B\", and \"delta_ratio_C\".\n5. \"slow_wave_rich_window\" must be exactly one of: \"A\", \"B\", \"C\", or null.\n6. Each delta ratio value must be a FLOAT percentage without a percent sign.\n7. If a report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n8. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"slow_wave_rich_window\": <\"A\"|\"B\"|\"C\"|null>, \"delta_ratio_A\": , \"delta_ratio_B\": , \"delta_ratio_C\": }", + "metrics": [ + { + "metric_id": "slow_wave_rich_window_accuracy", + "type": "categorical_check", + "target_key": "slow_wave_rich_window", + "weight": 40, + "params": { + "gt_value": "A" + } + }, + { + "metric_id": "delta_ratio_A_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_A", + "weight": 20, + "params": { + "gt_value": 87.5906727468016, + "tolerance": 4.379533637340081 + } + }, + { + "metric_id": "delta_ratio_B_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_B", + "weight": 20, + "params": { + "gt_value": 26.623459084838053, + "tolerance": 1.3311729542419029 + } + }, + { + "metric_id": "delta_ratio_C_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_C", + "weight": 20, + "params": { + "gt_value": 49.97306894714169, + "tolerance": 2.4986534473570847 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case21/case21_02.json b/sleep_assessment/cases/case21/case21_02.json new file mode 100644 index 0000000000000000000000000000000000000000..8cadbaefdc755d3942712c429396bd42208600ca --- /dev/null +++ b/sleep_assessment/cases/case21/case21_02.json @@ -0,0 +1,56 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "instruction": "Please analyze the EEG signals in the following three 5-minute windows of the provided sleep file: A: 64.0-69.0 minutes; B: 184.0-189.0 minutes; C: 354.5-359.5 minutes. Select EEG channels only. Determine which window is most slow-wave-rich, and report the delta-band relative power percentage for EEG channels in each of the three windows. Define delta relative power as power in 0.5-4Hz divided by total power in 0.5-40Hz, multiplied by 100. Report the selected window using only A, B, or C, and clearly state the delta relative power percentages for A, B, and C." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract four final outputs from the agent's report:\n1) which one of windows A, B, and C is most slow-wave-rich\n2) the delta relative power percentage for window A\n3) the delta relative power percentage for window B\n4) the delta relative power percentage for window C\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"slow_wave_rich_window\", \"delta_ratio_A\", \"delta_ratio_B\", and \"delta_ratio_C\".\n5. \"slow_wave_rich_window\" must be exactly one of: \"A\", \"B\", \"C\", or null.\n6. Each delta ratio value must be a FLOAT percentage without a percent sign.\n7. If a report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n8. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"slow_wave_rich_window\": <\"A\"|\"B\"|\"C\"|null>, \"delta_ratio_A\": , \"delta_ratio_B\": , \"delta_ratio_C\": }", + "metrics": [ + { + "metric_id": "slow_wave_rich_window_accuracy", + "type": "categorical_check", + "target_key": "slow_wave_rich_window", + "weight": 40, + "params": { + "gt_value": "B" + } + }, + { + "metric_id": "delta_ratio_A_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_A", + "weight": 20, + "params": { + "gt_value": 19.16429604391837, + "tolerance": 0.9582148021959185 + } + }, + { + "metric_id": "delta_ratio_B_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_B", + "weight": 20, + "params": { + "gt_value": 83.8454977168564, + "tolerance": 4.19227488584282 + } + }, + { + "metric_id": "delta_ratio_C_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_C", + "weight": 20, + "params": { + "gt_value": 54.972775817497435, + "tolerance": 2.748638790874872 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case21/case21_03.json b/sleep_assessment/cases/case21/case21_03.json new file mode 100644 index 0000000000000000000000000000000000000000..38041aa354f38bec56220c4ad8272d768bf71ec3 --- /dev/null +++ b/sleep_assessment/cases/case21/case21_03.json @@ -0,0 +1,56 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "instruction": "Please analyze the EEG signals in the following three 5-minute windows of the provided sleep file: A: 250.5-255.5 minutes; B: 5.5-10.5 minutes; C: 145.5-150.5 minutes. Select EEG channels only. Determine which window is most slow-wave-rich, and report the delta-band relative power percentage for EEG channels in each of the three windows. Define delta relative power as power in 0.5-4Hz divided by total power in 0.5-40Hz, multiplied by 100. Report the selected window using only A, B, or C, and clearly state the delta relative power percentages for A, B, and C." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract four final outputs from the agent's report:\n1) which one of windows A, B, and C is most slow-wave-rich\n2) the delta relative power percentage for window A\n3) the delta relative power percentage for window B\n4) the delta relative power percentage for window C\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"slow_wave_rich_window\", \"delta_ratio_A\", \"delta_ratio_B\", and \"delta_ratio_C\".\n5. \"slow_wave_rich_window\" must be exactly one of: \"A\", \"B\", \"C\", or null.\n6. Each delta ratio value must be a FLOAT percentage without a percent sign.\n7. If a report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n8. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"slow_wave_rich_window\": <\"A\"|\"B\"|\"C\"|null>, \"delta_ratio_A\": , \"delta_ratio_B\": , \"delta_ratio_C\": }", + "metrics": [ + { + "metric_id": "slow_wave_rich_window_accuracy", + "type": "categorical_check", + "target_key": "slow_wave_rich_window", + "weight": 40, + "params": { + "gt_value": "C" + } + }, + { + "metric_id": "delta_ratio_A_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_A", + "weight": 20, + "params": { + "gt_value": 31.34329516095618, + "tolerance": 1.5671647580478092 + } + }, + { + "metric_id": "delta_ratio_B_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_B", + "weight": 20, + "params": { + "gt_value": 44.09978377869893, + "tolerance": 2.2049891889349467 + } + }, + { + "metric_id": "delta_ratio_C_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_C", + "weight": 20, + "params": { + "gt_value": 85.60521495626206, + "tolerance": 4.280260747813103 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case21/case21_04.json b/sleep_assessment/cases/case21/case21_04.json new file mode 100644 index 0000000000000000000000000000000000000000..589d850eec7318252adfc7d02a39edbb34d8bddf --- /dev/null +++ b/sleep_assessment/cases/case21/case21_04.json @@ -0,0 +1,56 @@ +{ + "meta_info": { + "case_id": "ISRUC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_04.edf", + "instruction": "Please analyze the EEG signals in the following three 5-minute windows of the provided sleep file: A: 152.0-157.0 minutes; B: 196.0-201.0 minutes; C: 377.5-382.5 minutes. Select EEG channels only. Determine which window is most slow-wave-rich, and report the delta-band relative power percentage for EEG channels in each of the three windows. Define delta relative power as power in 0.5-4Hz divided by total power in 0.5-40Hz, multiplied by 100. Report the selected window using only A, B, or C, and clearly state the delta relative power percentages for A, B, and C." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract four final outputs from the agent's report:\n1) which one of windows A, B, and C is most slow-wave-rich\n2) the delta relative power percentage for window A\n3) the delta relative power percentage for window B\n4) the delta relative power percentage for window C\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"slow_wave_rich_window\", \"delta_ratio_A\", \"delta_ratio_B\", and \"delta_ratio_C\".\n5. \"slow_wave_rich_window\" must be exactly one of: \"A\", \"B\", \"C\", or null.\n6. Each delta ratio value must be a FLOAT percentage without a percent sign.\n7. If a report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n8. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"slow_wave_rich_window\": <\"A\"|\"B\"|\"C\"|null>, \"delta_ratio_A\": , \"delta_ratio_B\": , \"delta_ratio_C\": }", + "metrics": [ + { + "metric_id": "slow_wave_rich_window_accuracy", + "type": "categorical_check", + "target_key": "slow_wave_rich_window", + "weight": 40, + "params": { + "gt_value": "A" + } + }, + { + "metric_id": "delta_ratio_A_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_A", + "weight": 20, + "params": { + "gt_value": 91.59016371604208, + "tolerance": 4.579508185802104 + } + }, + { + "metric_id": "delta_ratio_B_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_B", + "weight": 20, + "params": { + "gt_value": 54.75510737858522, + "tolerance": 2.7377553689292613 + } + }, + { + "metric_id": "delta_ratio_C_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_C", + "weight": 20, + "params": { + "gt_value": 63.232329113094984, + "tolerance": 3.1616164556547495 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case21/case21_05.json b/sleep_assessment/cases/case21/case21_05.json new file mode 100644 index 0000000000000000000000000000000000000000..0318946f7838813482d595a0b7d3c57bba5fd033 --- /dev/null +++ b/sleep_assessment/cases/case21/case21_05.json @@ -0,0 +1,56 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "instruction": "Please analyze the EEG signals in the following three 5-minute windows of the provided sleep file: A: 101.5-106.5 minutes; B: 194.0-199.0 minutes; C: 115.0-120.0 minutes. Select EEG channels only. Determine which window is most slow-wave-rich, and report the delta-band relative power percentage for EEG channels in each of the three windows. Define delta relative power as power in 0.5-4Hz divided by total power in 0.5-40Hz, multiplied by 100. Report the selected window using only A, B, or C, and clearly state the delta relative power percentages for A, B, and C." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract four final outputs from the agent's report:\n1) which one of windows A, B, and C is most slow-wave-rich\n2) the delta relative power percentage for window A\n3) the delta relative power percentage for window B\n4) the delta relative power percentage for window C\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"slow_wave_rich_window\", \"delta_ratio_A\", \"delta_ratio_B\", and \"delta_ratio_C\".\n5. \"slow_wave_rich_window\" must be exactly one of: \"A\", \"B\", \"C\", or null.\n6. Each delta ratio value must be a FLOAT percentage without a percent sign.\n7. If a report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n8. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"slow_wave_rich_window\": <\"A\"|\"B\"|\"C\"|null>, \"delta_ratio_A\": , \"delta_ratio_B\": , \"delta_ratio_C\": }", + "metrics": [ + { + "metric_id": "slow_wave_rich_window_accuracy", + "type": "categorical_check", + "target_key": "slow_wave_rich_window", + "weight": 40, + "params": { + "gt_value": "B" + } + }, + { + "metric_id": "delta_ratio_A_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_A", + "weight": 20, + "params": { + "gt_value": 35.24534175407361, + "tolerance": 1.7622670877036803 + } + }, + { + "metric_id": "delta_ratio_B_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_B", + "weight": 20, + "params": { + "gt_value": 88.31902669703837, + "tolerance": 4.415951334851918 + } + }, + { + "metric_id": "delta_ratio_C_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_C", + "weight": 20, + "params": { + "gt_value": 53.50000902531341, + "tolerance": 2.675000451265671 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case21/case21_06.json b/sleep_assessment/cases/case21/case21_06.json new file mode 100644 index 0000000000000000000000000000000000000000..8becda12836956457db0ec95380e6d6fc3404a0a --- /dev/null +++ b/sleep_assessment/cases/case21/case21_06.json @@ -0,0 +1,56 @@ +{ + "meta_info": { + "case_id": "HMC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_01.edf", + "instruction": "Please analyze the EEG signals in the following three 5-minute windows of the provided sleep file: A: 91.0-96.0 minutes; B: 150.5-155.5 minutes; C: 54.5-59.5 minutes. Select EEG channels only. Determine which window is most slow-wave-rich, and report the delta-band relative power percentage for EEG channels in each of the three windows. Define delta relative power as power in 0.5-4Hz divided by total power in 0.5-40Hz, multiplied by 100. Report the selected window using only A, B, or C, and clearly state the delta relative power percentages for A, B, and C." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract four final outputs from the agent's report:\n1) which one of windows A, B, and C is most slow-wave-rich\n2) the delta relative power percentage for window A\n3) the delta relative power percentage for window B\n4) the delta relative power percentage for window C\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"slow_wave_rich_window\", \"delta_ratio_A\", \"delta_ratio_B\", and \"delta_ratio_C\".\n5. \"slow_wave_rich_window\" must be exactly one of: \"A\", \"B\", \"C\", or null.\n6. Each delta ratio value must be a FLOAT percentage without a percent sign.\n7. If a report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n8. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"slow_wave_rich_window\": <\"A\"|\"B\"|\"C\"|null>, \"delta_ratio_A\": , \"delta_ratio_B\": , \"delta_ratio_C\": }", + "metrics": [ + { + "metric_id": "slow_wave_rich_window_accuracy", + "type": "categorical_check", + "target_key": "slow_wave_rich_window", + "weight": 40, + "params": { + "gt_value": "C" + } + }, + { + "metric_id": "delta_ratio_A_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_A", + "weight": 20, + "params": { + "gt_value": 58.487758791987346, + "tolerance": 2.9243879395993675 + } + }, + { + "metric_id": "delta_ratio_B_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_B", + "weight": 20, + "params": { + "gt_value": 66.87223181181125, + "tolerance": 3.3436115905905623 + } + }, + { + "metric_id": "delta_ratio_C_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_C", + "weight": 20, + "params": { + "gt_value": 76.79686764596676, + "tolerance": 3.839843382298338 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case21/case21_07.json b/sleep_assessment/cases/case21/case21_07.json new file mode 100644 index 0000000000000000000000000000000000000000..e97155d14504ef42165b9520ef32a6f9e2fd3a0d --- /dev/null +++ b/sleep_assessment/cases/case21/case21_07.json @@ -0,0 +1,56 @@ +{ + "meta_info": { + "case_id": "HMC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_02.edf", + "instruction": "Please analyze the EEG signals in the following three 5-minute windows of the provided sleep file: A: 157.0-162.0 minutes; B: 359.0-364.0 minutes; C: 319.0-324.0 minutes. Select EEG channels only. Determine which window is most slow-wave-rich, and report the delta-band relative power percentage for EEG channels in each of the three windows. Define delta relative power as power in 0.5-4Hz divided by total power in 0.5-40Hz, multiplied by 100. Report the selected window using only A, B, or C, and clearly state the delta relative power percentages for A, B, and C." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract four final outputs from the agent's report:\n1) which one of windows A, B, and C is most slow-wave-rich\n2) the delta relative power percentage for window A\n3) the delta relative power percentage for window B\n4) the delta relative power percentage for window C\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"slow_wave_rich_window\", \"delta_ratio_A\", \"delta_ratio_B\", and \"delta_ratio_C\".\n5. \"slow_wave_rich_window\" must be exactly one of: \"A\", \"B\", \"C\", or null.\n6. Each delta ratio value must be a FLOAT percentage without a percent sign.\n7. If a report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n8. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"slow_wave_rich_window\": <\"A\"|\"B\"|\"C\"|null>, \"delta_ratio_A\": , \"delta_ratio_B\": , \"delta_ratio_C\": }", + "metrics": [ + { + "metric_id": "slow_wave_rich_window_accuracy", + "type": "categorical_check", + "target_key": "slow_wave_rich_window", + "weight": 40, + "params": { + "gt_value": "A" + } + }, + { + "metric_id": "delta_ratio_A_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_A", + "weight": 20, + "params": { + "gt_value": 90.89976165674851, + "tolerance": 4.544988082837426 + } + }, + { + "metric_id": "delta_ratio_B_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_B", + "weight": 20, + "params": { + "gt_value": 61.618960822576305, + "tolerance": 3.0809480411288153 + } + }, + { + "metric_id": "delta_ratio_C_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_C", + "weight": 20, + "params": { + "gt_value": 62.95006957930629, + "tolerance": 3.1475034789653145 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case21/case21_08.json b/sleep_assessment/cases/case21/case21_08.json new file mode 100644 index 0000000000000000000000000000000000000000..0976b3004e55bcf652eb83b912ff27b27d95e921 --- /dev/null +++ b/sleep_assessment/cases/case21/case21_08.json @@ -0,0 +1,56 @@ +{ + "meta_info": { + "case_id": "HMC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_03.edf", + "instruction": "Please analyze the EEG signals in the following three 5-minute windows of the provided sleep file: A: 60.5-65.5 minutes; B: 232.0-237.0 minutes; C: 440.0-445.0 minutes. Select EEG channels only. Determine which window is most slow-wave-rich, and report the delta-band relative power percentage for EEG channels in each of the three windows. Define delta relative power as power in 0.5-4Hz divided by total power in 0.5-40Hz, multiplied by 100. Report the selected window using only A, B, or C, and clearly state the delta relative power percentages for A, B, and C." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract four final outputs from the agent's report:\n1) which one of windows A, B, and C is most slow-wave-rich\n2) the delta relative power percentage for window A\n3) the delta relative power percentage for window B\n4) the delta relative power percentage for window C\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"slow_wave_rich_window\", \"delta_ratio_A\", \"delta_ratio_B\", and \"delta_ratio_C\".\n5. \"slow_wave_rich_window\" must be exactly one of: \"A\", \"B\", \"C\", or null.\n6. Each delta ratio value must be a FLOAT percentage without a percent sign.\n7. If a report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n8. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"slow_wave_rich_window\": <\"A\"|\"B\"|\"C\"|null>, \"delta_ratio_A\": , \"delta_ratio_B\": , \"delta_ratio_C\": }", + "metrics": [ + { + "metric_id": "slow_wave_rich_window_accuracy", + "type": "categorical_check", + "target_key": "slow_wave_rich_window", + "weight": 40, + "params": { + "gt_value": "B" + } + }, + { + "metric_id": "delta_ratio_A_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_A", + "weight": 20, + "params": { + "gt_value": 51.82226765014284, + "tolerance": 2.5911133825071424 + } + }, + { + "metric_id": "delta_ratio_B_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_B", + "weight": 20, + "params": { + "gt_value": 96.16343018525266, + "tolerance": 4.808171509262634 + } + }, + { + "metric_id": "delta_ratio_C_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_C", + "weight": 20, + "params": { + "gt_value": 63.403341361508346, + "tolerance": 3.1701670680754175 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case21/case21_09.json b/sleep_assessment/cases/case21/case21_09.json new file mode 100644 index 0000000000000000000000000000000000000000..94a8542be7f2f0685a4720045651c5077222ac05 --- /dev/null +++ b/sleep_assessment/cases/case21/case21_09.json @@ -0,0 +1,56 @@ +{ + "meta_info": { + "case_id": "HMC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_04.edf", + "instruction": "Please analyze the EEG signals in the following three 5-minute windows of the provided sleep file: A: 496.5-501.5 minutes; B: 379.5-384.5 minutes; C: 43.0-48.0 minutes. Select EEG channels only. Determine which window is most slow-wave-rich, and report the delta-band relative power percentage for EEG channels in each of the three windows. Define delta relative power as power in 0.5-4Hz divided by total power in 0.5-40Hz, multiplied by 100. Report the selected window using only A, B, or C, and clearly state the delta relative power percentages for A, B, and C." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract four final outputs from the agent's report:\n1) which one of windows A, B, and C is most slow-wave-rich\n2) the delta relative power percentage for window A\n3) the delta relative power percentage for window B\n4) the delta relative power percentage for window C\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"slow_wave_rich_window\", \"delta_ratio_A\", \"delta_ratio_B\", and \"delta_ratio_C\".\n5. \"slow_wave_rich_window\" must be exactly one of: \"A\", \"B\", \"C\", or null.\n6. Each delta ratio value must be a FLOAT percentage without a percent sign.\n7. If a report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n8. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"slow_wave_rich_window\": <\"A\"|\"B\"|\"C\"|null>, \"delta_ratio_A\": , \"delta_ratio_B\": , \"delta_ratio_C\": }", + "metrics": [ + { + "metric_id": "slow_wave_rich_window_accuracy", + "type": "categorical_check", + "target_key": "slow_wave_rich_window", + "weight": 40, + "params": { + "gt_value": "C" + } + }, + { + "metric_id": "delta_ratio_A_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_A", + "weight": 20, + "params": { + "gt_value": 46.901744669264325, + "tolerance": 2.3450872334632162 + } + }, + { + "metric_id": "delta_ratio_B_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_B", + "weight": 20, + "params": { + "gt_value": 60.26390585909406, + "tolerance": 3.0131952929547032 + } + }, + { + "metric_id": "delta_ratio_C_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_C", + "weight": 20, + "params": { + "gt_value": 84.39028196005106, + "tolerance": 4.219514098002553 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case21/case21_10.json b/sleep_assessment/cases/case21/case21_10.json new file mode 100644 index 0000000000000000000000000000000000000000..617af70ade9a63e2409f88d7aa3271375051afe8 --- /dev/null +++ b/sleep_assessment/cases/case21/case21_10.json @@ -0,0 +1,56 @@ +{ + "meta_info": { + "case_id": "HMC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_05.edf", + "instruction": "Please analyze the EEG signals in the following three 5-minute windows of the provided sleep file: A: 348.5-353.5 minutes; B: 138.5-143.5 minutes; C: 42.5-47.5 minutes. Select EEG channels only. Determine which window is most slow-wave-rich, and report the delta-band relative power percentage for EEG channels in each of the three windows. Define delta relative power as power in 0.5-4Hz divided by total power in 0.5-40Hz, multiplied by 100. Report the selected window using only A, B, or C, and clearly state the delta relative power percentages for A, B, and C." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract four final outputs from the agent's report:\n1) which one of windows A, B, and C is most slow-wave-rich\n2) the delta relative power percentage for window A\n3) the delta relative power percentage for window B\n4) the delta relative power percentage for window C\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"slow_wave_rich_window\", \"delta_ratio_A\", \"delta_ratio_B\", and \"delta_ratio_C\".\n5. \"slow_wave_rich_window\" must be exactly one of: \"A\", \"B\", \"C\", or null.\n6. Each delta ratio value must be a FLOAT percentage without a percent sign.\n7. If a report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n8. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"slow_wave_rich_window\": <\"A\"|\"B\"|\"C\"|null>, \"delta_ratio_A\": , \"delta_ratio_B\": , \"delta_ratio_C\": }", + "metrics": [ + { + "metric_id": "slow_wave_rich_window_accuracy", + "type": "categorical_check", + "target_key": "slow_wave_rich_window", + "weight": 40, + "params": { + "gt_value": "A" + } + }, + { + "metric_id": "delta_ratio_A_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_A", + "weight": 20, + "params": { + "gt_value": 89.2399238121076, + "tolerance": 4.46199619060538 + } + }, + { + "metric_id": "delta_ratio_B_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_B", + "weight": 20, + "params": { + "gt_value": 51.65480174696761, + "tolerance": 2.5827400873483803 + } + }, + { + "metric_id": "delta_ratio_C_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_C", + "weight": 20, + "params": { + "gt_value": 65.51899845038473, + "tolerance": 3.275949922519237 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case21/case21_11.json b/sleep_assessment/cases/case21/case21_11.json new file mode 100644 index 0000000000000000000000000000000000000000..4b2ac0e54c255237b60c3f3d8d4bce2d89245adf --- /dev/null +++ b/sleep_assessment/cases/case21/case21_11.json @@ -0,0 +1,56 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "instruction": "Please analyze the EEG signals in the following three 5-minute windows of the provided sleep file: A: 469.5-474.5 minutes; B: 53.0-58.0 minutes; C: 128.0-133.0 minutes. Select EEG channels only. Determine which window is most slow-wave-rich, and report the delta-band relative power percentage for EEG channels in each of the three windows. Define delta relative power as power in 0.5-4Hz divided by total power in 0.5-40Hz, multiplied by 100. Report the selected window using only A, B, or C, and clearly state the delta relative power percentages for A, B, and C." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract four final outputs from the agent's report:\n1) which one of windows A, B, and C is most slow-wave-rich\n2) the delta relative power percentage for window A\n3) the delta relative power percentage for window B\n4) the delta relative power percentage for window C\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"slow_wave_rich_window\", \"delta_ratio_A\", \"delta_ratio_B\", and \"delta_ratio_C\".\n5. \"slow_wave_rich_window\" must be exactly one of: \"A\", \"B\", \"C\", or null.\n6. Each delta ratio value must be a FLOAT percentage without a percent sign.\n7. If a report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n8. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"slow_wave_rich_window\": <\"A\"|\"B\"|\"C\"|null>, \"delta_ratio_A\": , \"delta_ratio_B\": , \"delta_ratio_C\": }", + "metrics": [ + { + "metric_id": "slow_wave_rich_window_accuracy", + "type": "categorical_check", + "target_key": "slow_wave_rich_window", + "weight": 40, + "params": { + "gt_value": "B" + } + }, + { + "metric_id": "delta_ratio_A_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_A", + "weight": 20, + "params": { + "gt_value": 15.689854260455554, + "tolerance": 0.7844927130227778 + } + }, + { + "metric_id": "delta_ratio_B_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_B", + "weight": 20, + "params": { + "gt_value": 93.13518686115162, + "tolerance": 4.6567593430575815 + } + }, + { + "metric_id": "delta_ratio_C_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_C", + "weight": 20, + "params": { + "gt_value": 51.75967892808275, + "tolerance": 2.5879839464041376 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case21/case21_12.json b/sleep_assessment/cases/case21/case21_12.json new file mode 100644 index 0000000000000000000000000000000000000000..7a20f2278e47cb1a721ad29e5df8ef738d3f2fee --- /dev/null +++ b/sleep_assessment/cases/case21/case21_12.json @@ -0,0 +1,56 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "instruction": "Please analyze the EEG signals in the following three 5-minute windows of the provided sleep file: A: 522.0-527.0 minutes; B: 43.0-48.0 minutes; C: 22.0-27.0 minutes. Select EEG channels only. Determine which window is most slow-wave-rich, and report the delta-band relative power percentage for EEG channels in each of the three windows. Define delta relative power as power in 0.5-4Hz divided by total power in 0.5-40Hz, multiplied by 100. Report the selected window using only A, B, or C, and clearly state the delta relative power percentages for A, B, and C." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract four final outputs from the agent's report:\n1) which one of windows A, B, and C is most slow-wave-rich\n2) the delta relative power percentage for window A\n3) the delta relative power percentage for window B\n4) the delta relative power percentage for window C\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"slow_wave_rich_window\", \"delta_ratio_A\", \"delta_ratio_B\", and \"delta_ratio_C\".\n5. \"slow_wave_rich_window\" must be exactly one of: \"A\", \"B\", \"C\", or null.\n6. Each delta ratio value must be a FLOAT percentage without a percent sign.\n7. If a report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n8. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"slow_wave_rich_window\": <\"A\"|\"B\"|\"C\"|null>, \"delta_ratio_A\": , \"delta_ratio_B\": , \"delta_ratio_C\": }", + "metrics": [ + { + "metric_id": "slow_wave_rich_window_accuracy", + "type": "categorical_check", + "target_key": "slow_wave_rich_window", + "weight": 40, + "params": { + "gt_value": "C" + } + }, + { + "metric_id": "delta_ratio_A_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_A", + "weight": 20, + "params": { + "gt_value": 1.2385372169456097, + "tolerance": 0.06192686084728049 + } + }, + { + "metric_id": "delta_ratio_B_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_B", + "weight": 20, + "params": { + "gt_value": 41.615293543358725, + "tolerance": 2.0807646771679362 + } + }, + { + "metric_id": "delta_ratio_C_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_C", + "weight": 20, + "params": { + "gt_value": 77.04710423208743, + "tolerance": 3.852355211604372 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case21/case21_13.json b/sleep_assessment/cases/case21/case21_13.json new file mode 100644 index 0000000000000000000000000000000000000000..d0d2e7467ee9387abf7bec16e7c26bab2c3854da --- /dev/null +++ b/sleep_assessment/cases/case21/case21_13.json @@ -0,0 +1,56 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "instruction": "Please analyze the EEG signals in the following three 5-minute windows of the provided sleep file: A: 100.5-105.5 minutes; B: 377.0-382.0 minutes; C: 413.0-418.0 minutes. Select EEG channels only. Determine which window is most slow-wave-rich, and report the delta-band relative power percentage for EEG channels in each of the three windows. Define delta relative power as power in 0.5-4Hz divided by total power in 0.5-40Hz, multiplied by 100. Report the selected window using only A, B, or C, and clearly state the delta relative power percentages for A, B, and C." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract four final outputs from the agent's report:\n1) which one of windows A, B, and C is most slow-wave-rich\n2) the delta relative power percentage for window A\n3) the delta relative power percentage for window B\n4) the delta relative power percentage for window C\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"slow_wave_rich_window\", \"delta_ratio_A\", \"delta_ratio_B\", and \"delta_ratio_C\".\n5. \"slow_wave_rich_window\" must be exactly one of: \"A\", \"B\", \"C\", or null.\n6. Each delta ratio value must be a FLOAT percentage without a percent sign.\n7. If a report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n8. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"slow_wave_rich_window\": <\"A\"|\"B\"|\"C\"|null>, \"delta_ratio_A\": , \"delta_ratio_B\": , \"delta_ratio_C\": }", + "metrics": [ + { + "metric_id": "slow_wave_rich_window_accuracy", + "type": "categorical_check", + "target_key": "slow_wave_rich_window", + "weight": 40, + "params": { + "gt_value": "A" + } + }, + { + "metric_id": "delta_ratio_A_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_A", + "weight": 20, + "params": { + "gt_value": 90.32094560477164, + "tolerance": 4.516047280238582 + } + }, + { + "metric_id": "delta_ratio_B_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_B", + "weight": 20, + "params": { + "gt_value": 11.840750907955568, + "tolerance": 0.5920375453977784 + } + }, + { + "metric_id": "delta_ratio_C_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_C", + "weight": 20, + "params": { + "gt_value": 43.09124214961489, + "tolerance": 2.1545621074807446 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case21/case21_14.json b/sleep_assessment/cases/case21/case21_14.json new file mode 100644 index 0000000000000000000000000000000000000000..fa25efa58e8d6b8636ff92f6ce51db664354f043 --- /dev/null +++ b/sleep_assessment/cases/case21/case21_14.json @@ -0,0 +1,56 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "instruction": "Please analyze the EEG signals in the following three 5-minute windows of the provided sleep file: A: 418.5-423.5 minutes; B: 135.5-140.5 minutes; C: 150.0-155.0 minutes. Select EEG channels only. Determine which window is most slow-wave-rich, and report the delta-band relative power percentage for EEG channels in each of the three windows. Define delta relative power as power in 0.5-4Hz divided by total power in 0.5-40Hz, multiplied by 100. Report the selected window using only A, B, or C, and clearly state the delta relative power percentages for A, B, and C." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract four final outputs from the agent's report:\n1) which one of windows A, B, and C is most slow-wave-rich\n2) the delta relative power percentage for window A\n3) the delta relative power percentage for window B\n4) the delta relative power percentage for window C\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"slow_wave_rich_window\", \"delta_ratio_A\", \"delta_ratio_B\", and \"delta_ratio_C\".\n5. \"slow_wave_rich_window\" must be exactly one of: \"A\", \"B\", \"C\", or null.\n6. Each delta ratio value must be a FLOAT percentage without a percent sign.\n7. If a report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n8. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"slow_wave_rich_window\": <\"A\"|\"B\"|\"C\"|null>, \"delta_ratio_A\": , \"delta_ratio_B\": , \"delta_ratio_C\": }", + "metrics": [ + { + "metric_id": "slow_wave_rich_window_accuracy", + "type": "categorical_check", + "target_key": "slow_wave_rich_window", + "weight": 40, + "params": { + "gt_value": "B" + } + }, + { + "metric_id": "delta_ratio_A_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_A", + "weight": 20, + "params": { + "gt_value": 12.27055346030883, + "tolerance": 0.6135276730154415 + } + }, + { + "metric_id": "delta_ratio_B_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_B", + "weight": 20, + "params": { + "gt_value": 92.56927964330413, + "tolerance": 4.628463982165207 + } + }, + { + "metric_id": "delta_ratio_C_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_C", + "weight": 20, + "params": { + "gt_value": 33.346197460620544, + "tolerance": 1.6673098730310274 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case21/case21_15.json b/sleep_assessment/cases/case21/case21_15.json new file mode 100644 index 0000000000000000000000000000000000000000..2f8746f881159ae4f838faf0cdb33352d759e8c8 --- /dev/null +++ b/sleep_assessment/cases/case21/case21_15.json @@ -0,0 +1,56 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "instruction": "Please analyze the EEG signals in the following three 5-minute windows of the provided sleep file: A: 489.0-494.0 minutes; B: 140.0-145.0 minutes; C: 105.0-110.0 minutes. Select EEG channels only. Determine which window is most slow-wave-rich, and report the delta-band relative power percentage for EEG channels in each of the three windows. Define delta relative power as power in 0.5-4Hz divided by total power in 0.5-40Hz, multiplied by 100. Report the selected window using only A, B, or C, and clearly state the delta relative power percentages for A, B, and C." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract four final outputs from the agent's report:\n1) which one of windows A, B, and C is most slow-wave-rich\n2) the delta relative power percentage for window A\n3) the delta relative power percentage for window B\n4) the delta relative power percentage for window C\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"slow_wave_rich_window\", \"delta_ratio_A\", \"delta_ratio_B\", and \"delta_ratio_C\".\n5. \"slow_wave_rich_window\" must be exactly one of: \"A\", \"B\", \"C\", or null.\n6. Each delta ratio value must be a FLOAT percentage without a percent sign.\n7. If a report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n8. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"slow_wave_rich_window\": <\"A\"|\"B\"|\"C\"|null>, \"delta_ratio_A\": , \"delta_ratio_B\": , \"delta_ratio_C\": }", + "metrics": [ + { + "metric_id": "slow_wave_rich_window_accuracy", + "type": "categorical_check", + "target_key": "slow_wave_rich_window", + "weight": 40, + "params": { + "gt_value": "C" + } + }, + { + "metric_id": "delta_ratio_A_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_A", + "weight": 20, + "params": { + "gt_value": 27.673189351111194, + "tolerance": 1.3836594675555598 + } + }, + { + "metric_id": "delta_ratio_B_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_B", + "weight": 20, + "params": { + "gt_value": 43.46777410870381, + "tolerance": 2.1733887054351904 + } + }, + { + "metric_id": "delta_ratio_C_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_C", + "weight": 20, + "params": { + "gt_value": 86.80179483765484, + "tolerance": 4.340089741882742 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case21/case21_16.json b/sleep_assessment/cases/case21/case21_16.json new file mode 100644 index 0000000000000000000000000000000000000000..ac48752f459dddf6fcd38037156f4967bffc74ed --- /dev/null +++ b/sleep_assessment/cases/case21/case21_16.json @@ -0,0 +1,56 @@ +{ + "meta_info": { + "case_id": "MASSSS3_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_01.edf", + "instruction": "Please analyze the EEG signals in the following three 5-minute windows of the provided sleep file: A: 132.5-137.5 minutes; B: 489.5-494.5 minutes; C: 162.0-167.0 minutes. Select EEG channels only. Determine which window is most slow-wave-rich, and report the delta-band relative power percentage for EEG channels in each of the three windows. Define delta relative power as power in 0.5-4Hz divided by total power in 0.5-40Hz, multiplied by 100. Report the selected window using only A, B, or C, and clearly state the delta relative power percentages for A, B, and C." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract four final outputs from the agent's report:\n1) which one of windows A, B, and C is most slow-wave-rich\n2) the delta relative power percentage for window A\n3) the delta relative power percentage for window B\n4) the delta relative power percentage for window C\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"slow_wave_rich_window\", \"delta_ratio_A\", \"delta_ratio_B\", and \"delta_ratio_C\".\n5. \"slow_wave_rich_window\" must be exactly one of: \"A\", \"B\", \"C\", or null.\n6. Each delta ratio value must be a FLOAT percentage without a percent sign.\n7. If a report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n8. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"slow_wave_rich_window\": <\"A\"|\"B\"|\"C\"|null>, \"delta_ratio_A\": , \"delta_ratio_B\": , \"delta_ratio_C\": }", + "metrics": [ + { + "metric_id": "slow_wave_rich_window_accuracy", + "type": "categorical_check", + "target_key": "slow_wave_rich_window", + "weight": 40, + "params": { + "gt_value": "A" + } + }, + { + "metric_id": "delta_ratio_A_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_A", + "weight": 20, + "params": { + "gt_value": 93.27637253133445, + "tolerance": 4.663818626566722 + } + }, + { + "metric_id": "delta_ratio_B_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_B", + "weight": 20, + "params": { + "gt_value": 61.45997426567755, + "tolerance": 3.072998713283878 + } + }, + { + "metric_id": "delta_ratio_C_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_C", + "weight": 20, + "params": { + "gt_value": 64.66326311853447, + "tolerance": 3.2331631559267238 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case21/case21_17.json b/sleep_assessment/cases/case21/case21_17.json new file mode 100644 index 0000000000000000000000000000000000000000..6c3c7e50b99470c2b89db614a3ec8e7d3e93c8ac --- /dev/null +++ b/sleep_assessment/cases/case21/case21_17.json @@ -0,0 +1,56 @@ +{ + "meta_info": { + "case_id": "MASSSS3_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_02.edf", + "instruction": "Please analyze the EEG signals in the following three 5-minute windows of the provided sleep file: A: 6.0-11.0 minutes; B: 47.5-52.5 minutes; C: 64.0-69.0 minutes. Select EEG channels only. Determine which window is most slow-wave-rich, and report the delta-band relative power percentage for EEG channels in each of the three windows. Define delta relative power as power in 0.5-4Hz divided by total power in 0.5-40Hz, multiplied by 100. Report the selected window using only A, B, or C, and clearly state the delta relative power percentages for A, B, and C." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract four final outputs from the agent's report:\n1) which one of windows A, B, and C is most slow-wave-rich\n2) the delta relative power percentage for window A\n3) the delta relative power percentage for window B\n4) the delta relative power percentage for window C\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"slow_wave_rich_window\", \"delta_ratio_A\", \"delta_ratio_B\", and \"delta_ratio_C\".\n5. \"slow_wave_rich_window\" must be exactly one of: \"A\", \"B\", \"C\", or null.\n6. Each delta ratio value must be a FLOAT percentage without a percent sign.\n7. If a report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n8. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"slow_wave_rich_window\": <\"A\"|\"B\"|\"C\"|null>, \"delta_ratio_A\": , \"delta_ratio_B\": , \"delta_ratio_C\": }", + "metrics": [ + { + "metric_id": "slow_wave_rich_window_accuracy", + "type": "categorical_check", + "target_key": "slow_wave_rich_window", + "weight": 40, + "params": { + "gt_value": "B" + } + }, + { + "metric_id": "delta_ratio_A_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_A", + "weight": 20, + "params": { + "gt_value": 23.19850896800388, + "tolerance": 1.159925448400194 + } + }, + { + "metric_id": "delta_ratio_B_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_B", + "weight": 20, + "params": { + "gt_value": 92.34760687018822, + "tolerance": 4.617380343509411 + } + }, + { + "metric_id": "delta_ratio_C_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_C", + "weight": 20, + "params": { + "gt_value": 54.11837997230321, + "tolerance": 2.705918998615161 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case21/case21_18.json b/sleep_assessment/cases/case21/case21_18.json new file mode 100644 index 0000000000000000000000000000000000000000..89ba4e996f55c039b0095bfdf41943c875cbb35f --- /dev/null +++ b/sleep_assessment/cases/case21/case21_18.json @@ -0,0 +1,56 @@ +{ + "meta_info": { + "case_id": "MASSSS3_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_03.edf", + "instruction": "Please analyze the EEG signals in the following three 5-minute windows of the provided sleep file: A: 384.5-389.5 minutes; B: 180.0-185.0 minutes; C: 78.0-83.0 minutes. Select EEG channels only. Determine which window is most slow-wave-rich, and report the delta-band relative power percentage for EEG channels in each of the three windows. Define delta relative power as power in 0.5-4Hz divided by total power in 0.5-40Hz, multiplied by 100. Report the selected window using only A, B, or C, and clearly state the delta relative power percentages for A, B, and C." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract four final outputs from the agent's report:\n1) which one of windows A, B, and C is most slow-wave-rich\n2) the delta relative power percentage for window A\n3) the delta relative power percentage for window B\n4) the delta relative power percentage for window C\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"slow_wave_rich_window\", \"delta_ratio_A\", \"delta_ratio_B\", and \"delta_ratio_C\".\n5. \"slow_wave_rich_window\" must be exactly one of: \"A\", \"B\", \"C\", or null.\n6. Each delta ratio value must be a FLOAT percentage without a percent sign.\n7. If a report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n8. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"slow_wave_rich_window\": <\"A\"|\"B\"|\"C\"|null>, \"delta_ratio_A\": , \"delta_ratio_B\": , \"delta_ratio_C\": }", + "metrics": [ + { + "metric_id": "slow_wave_rich_window_accuracy", + "type": "categorical_check", + "target_key": "slow_wave_rich_window", + "weight": 40, + "params": { + "gt_value": "C" + } + }, + { + "metric_id": "delta_ratio_A_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_A", + "weight": 20, + "params": { + "gt_value": 58.17958854040366, + "tolerance": 2.908979427020183 + } + }, + { + "metric_id": "delta_ratio_B_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_B", + "weight": 20, + "params": { + "gt_value": 61.46403534512942, + "tolerance": 3.073201767256471 + } + }, + { + "metric_id": "delta_ratio_C_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_C", + "weight": 20, + "params": { + "gt_value": 88.29564335221875, + "tolerance": 4.414782167610937 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case21/case21_19.json b/sleep_assessment/cases/case21/case21_19.json new file mode 100644 index 0000000000000000000000000000000000000000..8fef3f3c4d4c480f17c15b7815224268f6fc2293 --- /dev/null +++ b/sleep_assessment/cases/case21/case21_19.json @@ -0,0 +1,56 @@ +{ + "meta_info": { + "case_id": "MASSSS3_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_04.edf", + "instruction": "Please analyze the EEG signals in the following three 5-minute windows of the provided sleep file: A: 54.5-59.5 minutes; B: 391.5-396.5 minutes; C: 2.0-7.0 minutes. Select EEG channels only. Determine which window is most slow-wave-rich, and report the delta-band relative power percentage for EEG channels in each of the three windows. Define delta relative power as power in 0.5-4Hz divided by total power in 0.5-40Hz, multiplied by 100. Report the selected window using only A, B, or C, and clearly state the delta relative power percentages for A, B, and C." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract four final outputs from the agent's report:\n1) which one of windows A, B, and C is most slow-wave-rich\n2) the delta relative power percentage for window A\n3) the delta relative power percentage for window B\n4) the delta relative power percentage for window C\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"slow_wave_rich_window\", \"delta_ratio_A\", \"delta_ratio_B\", and \"delta_ratio_C\".\n5. \"slow_wave_rich_window\" must be exactly one of: \"A\", \"B\", \"C\", or null.\n6. Each delta ratio value must be a FLOAT percentage without a percent sign.\n7. If a report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n8. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"slow_wave_rich_window\": <\"A\"|\"B\"|\"C\"|null>, \"delta_ratio_A\": , \"delta_ratio_B\": , \"delta_ratio_C\": }", + "metrics": [ + { + "metric_id": "slow_wave_rich_window_accuracy", + "type": "categorical_check", + "target_key": "slow_wave_rich_window", + "weight": 40, + "params": { + "gt_value": "A" + } + }, + { + "metric_id": "delta_ratio_A_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_A", + "weight": 20, + "params": { + "gt_value": 90.73300852862864, + "tolerance": 4.536650426431432 + } + }, + { + "metric_id": "delta_ratio_B_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_B", + "weight": 20, + "params": { + "gt_value": 55.998864756870795, + "tolerance": 2.79994323784354 + } + }, + { + "metric_id": "delta_ratio_C_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_C", + "weight": 20, + "params": { + "gt_value": 57.39987466171825, + "tolerance": 2.869993733085913 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case21/case21_20.json b/sleep_assessment/cases/case21/case21_20.json new file mode 100644 index 0000000000000000000000000000000000000000..28707fa57eab46bd9779279b52674cf893b3273c --- /dev/null +++ b/sleep_assessment/cases/case21/case21_20.json @@ -0,0 +1,56 @@ +{ + "meta_info": { + "case_id": "MASSSS3_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_05.edf", + "instruction": "Please analyze the EEG signals in the following three 5-minute windows of the provided sleep file: A: 473.0-478.0 minutes; B: 76.0-81.0 minutes; C: 335.0-340.0 minutes. Select EEG channels only. Determine which window is most slow-wave-rich, and report the delta-band relative power percentage for EEG channels in each of the three windows. Define delta relative power as power in 0.5-4Hz divided by total power in 0.5-40Hz, multiplied by 100. Report the selected window using only A, B, or C, and clearly state the delta relative power percentages for A, B, and C." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract four final outputs from the agent's report:\n1) which one of windows A, B, and C is most slow-wave-rich\n2) the delta relative power percentage for window A\n3) the delta relative power percentage for window B\n4) the delta relative power percentage for window C\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"slow_wave_rich_window\", \"delta_ratio_A\", \"delta_ratio_B\", and \"delta_ratio_C\".\n5. \"slow_wave_rich_window\" must be exactly one of: \"A\", \"B\", \"C\", or null.\n6. Each delta ratio value must be a FLOAT percentage without a percent sign.\n7. If a report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n8. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"slow_wave_rich_window\": <\"A\"|\"B\"|\"C\"|null>, \"delta_ratio_A\": , \"delta_ratio_B\": , \"delta_ratio_C\": }", + "metrics": [ + { + "metric_id": "slow_wave_rich_window_accuracy", + "type": "categorical_check", + "target_key": "slow_wave_rich_window", + "weight": 40, + "params": { + "gt_value": "B" + } + }, + { + "metric_id": "delta_ratio_A_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_A", + "weight": 20, + "params": { + "gt_value": 55.67484078891339, + "tolerance": 2.78374203944567 + } + }, + { + "metric_id": "delta_ratio_B_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_B", + "weight": 20, + "params": { + "gt_value": 95.34960832725586, + "tolerance": 4.767480416362793 + } + }, + { + "metric_id": "delta_ratio_C_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_C", + "weight": 20, + "params": { + "gt_value": 66.126918150972, + "tolerance": 3.3063459075486 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case21/case21_21.json b/sleep_assessment/cases/case21/case21_21.json new file mode 100644 index 0000000000000000000000000000000000000000..fa6830b7a252d4638e17e42e8041dbc1586593b0 --- /dev/null +++ b/sleep_assessment/cases/case21/case21_21.json @@ -0,0 +1,56 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "instruction": "Please analyze the EEG signals in the following three 5-minute windows of the provided sleep file: A: 291.5-296.5 minutes; B: 25.0-30.0 minutes; C: 52.5-57.5 minutes. Select EEG channels only. Determine which window is most slow-wave-rich, and report the delta-band relative power percentage for EEG channels in each of the three windows. Define delta relative power as power in 0.5-4Hz divided by total power in 0.5-40Hz, multiplied by 100. Report the selected window using only A, B, or C, and clearly state the delta relative power percentages for A, B, and C." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract four final outputs from the agent's report:\n1) which one of windows A, B, and C is most slow-wave-rich\n2) the delta relative power percentage for window A\n3) the delta relative power percentage for window B\n4) the delta relative power percentage for window C\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"slow_wave_rich_window\", \"delta_ratio_A\", \"delta_ratio_B\", and \"delta_ratio_C\".\n5. \"slow_wave_rich_window\" must be exactly one of: \"A\", \"B\", \"C\", or null.\n6. Each delta ratio value must be a FLOAT percentage without a percent sign.\n7. If a report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n8. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"slow_wave_rich_window\": <\"A\"|\"B\"|\"C\"|null>, \"delta_ratio_A\": , \"delta_ratio_B\": , \"delta_ratio_C\": }", + "metrics": [ + { + "metric_id": "slow_wave_rich_window_accuracy", + "type": "categorical_check", + "target_key": "slow_wave_rich_window", + "weight": 40, + "params": { + "gt_value": "C" + } + }, + { + "metric_id": "delta_ratio_A_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_A", + "weight": 20, + "params": { + "gt_value": 59.31677355978399, + "tolerance": 2.9658386779891996 + } + }, + { + "metric_id": "delta_ratio_B_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_B", + "weight": 20, + "params": { + "gt_value": 81.05282741279751, + "tolerance": 4.052641370639876 + } + }, + { + "metric_id": "delta_ratio_C_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_C", + "weight": 20, + "params": { + "gt_value": 86.88394324697192, + "tolerance": 4.344197162348596 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case21/case21_22.json b/sleep_assessment/cases/case21/case21_22.json new file mode 100644 index 0000000000000000000000000000000000000000..2392dd4b3a78b481943b57493cffe7a053fbb357 --- /dev/null +++ b/sleep_assessment/cases/case21/case21_22.json @@ -0,0 +1,56 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "instruction": "Please analyze the EEG signals in the following three 5-minute windows of the provided sleep file: A: 315.0-320.0 minutes; B: 296.0-301.0 minutes; C: 224.5-229.5 minutes. Select EEG channels only. Determine which window is most slow-wave-rich, and report the delta-band relative power percentage for EEG channels in each of the three windows. Define delta relative power as power in 0.5-4Hz divided by total power in 0.5-40Hz, multiplied by 100. Report the selected window using only A, B, or C, and clearly state the delta relative power percentages for A, B, and C." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract four final outputs from the agent's report:\n1) which one of windows A, B, and C is most slow-wave-rich\n2) the delta relative power percentage for window A\n3) the delta relative power percentage for window B\n4) the delta relative power percentage for window C\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"slow_wave_rich_window\", \"delta_ratio_A\", \"delta_ratio_B\", and \"delta_ratio_C\".\n5. \"slow_wave_rich_window\" must be exactly one of: \"A\", \"B\", \"C\", or null.\n6. Each delta ratio value must be a FLOAT percentage without a percent sign.\n7. If a report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n8. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"slow_wave_rich_window\": <\"A\"|\"B\"|\"C\"|null>, \"delta_ratio_A\": , \"delta_ratio_B\": , \"delta_ratio_C\": }", + "metrics": [ + { + "metric_id": "slow_wave_rich_window_accuracy", + "type": "categorical_check", + "target_key": "slow_wave_rich_window", + "weight": 40, + "params": { + "gt_value": "A" + } + }, + { + "metric_id": "delta_ratio_A_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_A", + "weight": 20, + "params": { + "gt_value": 78.04769500659657, + "tolerance": 3.9023847503298286 + } + }, + { + "metric_id": "delta_ratio_B_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_B", + "weight": 20, + "params": { + "gt_value": 75.74397205594231, + "tolerance": 3.7871986027971154 + } + }, + { + "metric_id": "delta_ratio_C_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_C", + "weight": 20, + "params": { + "gt_value": 27.43508758259991, + "tolerance": 1.3717543791299955 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case21/case21_23.json b/sleep_assessment/cases/case21/case21_23.json new file mode 100644 index 0000000000000000000000000000000000000000..ddb5b7430a1d2735f25da5823d82775dac6f8320 --- /dev/null +++ b/sleep_assessment/cases/case21/case21_23.json @@ -0,0 +1,56 @@ +{ + "meta_info": { + "case_id": "Physionet2018_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_03.edf", + "instruction": "Please analyze the EEG signals in the following three 5-minute windows of the provided sleep file: A: 430.5-435.5 minutes; B: 65.0-70.0 minutes; C: 12.0-17.0 minutes. Select EEG channels only. Determine which window is most slow-wave-rich, and report the delta-band relative power percentage for EEG channels in each of the three windows. Define delta relative power as power in 0.5-4Hz divided by total power in 0.5-40Hz, multiplied by 100. Report the selected window using only A, B, or C, and clearly state the delta relative power percentages for A, B, and C." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract four final outputs from the agent's report:\n1) which one of windows A, B, and C is most slow-wave-rich\n2) the delta relative power percentage for window A\n3) the delta relative power percentage for window B\n4) the delta relative power percentage for window C\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"slow_wave_rich_window\", \"delta_ratio_A\", \"delta_ratio_B\", and \"delta_ratio_C\".\n5. \"slow_wave_rich_window\" must be exactly one of: \"A\", \"B\", \"C\", or null.\n6. Each delta ratio value must be a FLOAT percentage without a percent sign.\n7. If a report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n8. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"slow_wave_rich_window\": <\"A\"|\"B\"|\"C\"|null>, \"delta_ratio_A\": , \"delta_ratio_B\": , \"delta_ratio_C\": }", + "metrics": [ + { + "metric_id": "slow_wave_rich_window_accuracy", + "type": "categorical_check", + "target_key": "slow_wave_rich_window", + "weight": 40, + "params": { + "gt_value": "B" + } + }, + { + "metric_id": "delta_ratio_A_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_A", + "weight": 20, + "params": { + "gt_value": 59.252773178980036, + "tolerance": 2.962638658949002 + } + }, + { + "metric_id": "delta_ratio_B_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_B", + "weight": 20, + "params": { + "gt_value": 92.71052166121406, + "tolerance": 4.635526083060703 + } + }, + { + "metric_id": "delta_ratio_C_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_C", + "weight": 20, + "params": { + "gt_value": 60.27504120405306, + "tolerance": 3.0137520602026533 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case21/case21_24.json b/sleep_assessment/cases/case21/case21_24.json new file mode 100644 index 0000000000000000000000000000000000000000..a7478ca1405866fb784b24cd4859a84a442c14a0 --- /dev/null +++ b/sleep_assessment/cases/case21/case21_24.json @@ -0,0 +1,56 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "instruction": "Please analyze the EEG signals in the following three 5-minute windows of the provided sleep file: A: 281.5-286.5 minutes; B: 291.5-296.5 minutes; C: 137.0-142.0 minutes. Select EEG channels only. Determine which window is most slow-wave-rich, and report the delta-band relative power percentage for EEG channels in each of the three windows. Define delta relative power as power in 0.5-4Hz divided by total power in 0.5-40Hz, multiplied by 100. Report the selected window using only A, B, or C, and clearly state the delta relative power percentages for A, B, and C." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract four final outputs from the agent's report:\n1) which one of windows A, B, and C is most slow-wave-rich\n2) the delta relative power percentage for window A\n3) the delta relative power percentage for window B\n4) the delta relative power percentage for window C\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"slow_wave_rich_window\", \"delta_ratio_A\", \"delta_ratio_B\", and \"delta_ratio_C\".\n5. \"slow_wave_rich_window\" must be exactly one of: \"A\", \"B\", \"C\", or null.\n6. Each delta ratio value must be a FLOAT percentage without a percent sign.\n7. If a report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n8. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"slow_wave_rich_window\": <\"A\"|\"B\"|\"C\"|null>, \"delta_ratio_A\": , \"delta_ratio_B\": , \"delta_ratio_C\": }", + "metrics": [ + { + "metric_id": "slow_wave_rich_window_accuracy", + "type": "categorical_check", + "target_key": "slow_wave_rich_window", + "weight": 40, + "params": { + "gt_value": "C" + } + }, + { + "metric_id": "delta_ratio_A_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_A", + "weight": 20, + "params": { + "gt_value": 32.48030419404315, + "tolerance": 1.6240152097021576 + } + }, + { + "metric_id": "delta_ratio_B_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_B", + "weight": 20, + "params": { + "gt_value": 67.38808770844375, + "tolerance": 3.369404385422188 + } + }, + { + "metric_id": "delta_ratio_C_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_C", + "weight": 20, + "params": { + "gt_value": 89.21449458576754, + "tolerance": 4.460724729288377 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case21/case21_25.json b/sleep_assessment/cases/case21/case21_25.json new file mode 100644 index 0000000000000000000000000000000000000000..945329a352f1908dcbe7346bacd4b82460992ecd --- /dev/null +++ b/sleep_assessment/cases/case21/case21_25.json @@ -0,0 +1,56 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "instruction": "Please analyze the EEG signals in the following three 5-minute windows of the provided sleep file: A: 365.5-370.5 minutes; B: 260.0-265.0 minutes; C: 273.5-278.5 minutes. Select EEG channels only. Determine which window is most slow-wave-rich, and report the delta-band relative power percentage for EEG channels in each of the three windows. Define delta relative power as power in 0.5-4Hz divided by total power in 0.5-40Hz, multiplied by 100. Report the selected window using only A, B, or C, and clearly state the delta relative power percentages for A, B, and C." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract four final outputs from the agent's report:\n1) which one of windows A, B, and C is most slow-wave-rich\n2) the delta relative power percentage for window A\n3) the delta relative power percentage for window B\n4) the delta relative power percentage for window C\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"slow_wave_rich_window\", \"delta_ratio_A\", \"delta_ratio_B\", and \"delta_ratio_C\".\n5. \"slow_wave_rich_window\" must be exactly one of: \"A\", \"B\", \"C\", or null.\n6. Each delta ratio value must be a FLOAT percentage without a percent sign.\n7. If a report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n8. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"slow_wave_rich_window\": <\"A\"|\"B\"|\"C\"|null>, \"delta_ratio_A\": , \"delta_ratio_B\": , \"delta_ratio_C\": }", + "metrics": [ + { + "metric_id": "slow_wave_rich_window_accuracy", + "type": "categorical_check", + "target_key": "slow_wave_rich_window", + "weight": 40, + "params": { + "gt_value": "A" + } + }, + { + "metric_id": "delta_ratio_A_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_A", + "weight": 20, + "params": { + "gt_value": 87.95827716935813, + "tolerance": 4.3979138584679065 + } + }, + { + "metric_id": "delta_ratio_B_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_B", + "weight": 20, + "params": { + "gt_value": 45.63446686581834, + "tolerance": 2.281723343290917 + } + }, + { + "metric_id": "delta_ratio_C_accuracy", + "type": "numeric_check", + "target_key": "delta_ratio_C", + "weight": 20, + "params": { + "gt_value": 67.24085984957885, + "tolerance": 3.362042992478943 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case22/case22_01.json b/sleep_assessment/cases/case22/case22_01.json new file mode 100644 index 0000000000000000000000000000000000000000..aa2676f2795ab907b97914de980a3cc2e658cae4 --- /dev/null +++ b/sleep_assessment/cases/case22/case22_01.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "instruction": "Please analyze the provided sleep file in the time window from minute 189.5 to minute 192.5. For this ISRUC file, X1, X3, and X4 are EMG channels. Select the chin EMG/EMG channel(s), apply a 10-80Hz bandpass filter appropriate for the 200Hz sampling rate, and compute the mean chin EMG RMS value in this event window. Also compute the whole-night EMG RMS mean using the same EMG channel(s) and filter. Compare the event-window RMS with the whole-night RMS, determine whether the event window follows a REM-like pattern, and give a brief analysis. Clearly report the event-window EMG RMS, the whole-night EMG RMS, and the final REM-like or non-REM-like judgment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the event-window chin EMG RMS value and preserve the complete report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"event_emg_rms\" and \"full_text\".\n5. \"event_emg_rms\" must be a FLOAT or null. Do not include units.\n6. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n7. If the event-window EMG RMS value is missing or unclear, set \"event_emg_rms\" to null.\n\n### OUTPUT TEMPLATE\n{\"event_emg_rms\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "event_chin_emg_rms_accuracy", + "type": "numeric_check", + "target_key": "event_emg_rms", + "weight": 40, + "params": { + "gt_value": 1.9257770939298657e-07, + "tolerance": 1.9257770939298657e-08 + } + }, + { + "metric_id": "rem_like_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 60, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_01.edf. The analyzed event window is 189.5-192.5 minutes. The current window is a true REM segment. The EMG channels are X1, X3, X4. The bandpass filter is 10-80Hz. The ground-truth event-window EMG RMS is 1.92577709393e-07; the whole-night EMG RMS is 1.63135693279e-06; the event/whole-night RMS ratio is 0.118048. The REM-like threshold is event RMS <= 0.8 * whole-night RMS. The correct final REM-like conclusion is: REM-like. The correct response must state that the event-window EMG RMS is significantly lower than the whole-night EMG RMS and must judge the window as REM or REM-like. Mark the response as correct only if it uses the RMS comparison consistently and gives the correct REM-like versus non-REM-like conclusion. If the response gives the opposite conclusion, ignores the RMS comparison, claims REM-like for a true Non-REM segment, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case22/case22_02.json b/sleep_assessment/cases/case22/case22_02.json new file mode 100644 index 0000000000000000000000000000000000000000..d27f393c777ada3939b3033d2e3168805eb870bf --- /dev/null +++ b/sleep_assessment/cases/case22/case22_02.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "instruction": "Please analyze the provided sleep file in the time window from minute 462.0 to minute 465.0. For this ISRUC file, X1, X3, and X4 are EMG channels. Select the chin EMG/EMG channel(s), apply a 10-80Hz bandpass filter appropriate for the 200Hz sampling rate, and compute the mean chin EMG RMS value in this event window. Also compute the whole-night EMG RMS mean using the same EMG channel(s) and filter. Compare the event-window RMS with the whole-night RMS, determine whether the event window follows a REM-like pattern, and give a brief analysis. Clearly report the event-window EMG RMS, the whole-night EMG RMS, and the final REM-like or non-REM-like judgment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the event-window chin EMG RMS value and preserve the complete report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"event_emg_rms\" and \"full_text\".\n5. \"event_emg_rms\" must be a FLOAT or null. Do not include units.\n6. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n7. If the event-window EMG RMS value is missing or unclear, set \"event_emg_rms\" to null.\n\n### OUTPUT TEMPLATE\n{\"event_emg_rms\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "event_chin_emg_rms_accuracy", + "type": "numeric_check", + "target_key": "event_emg_rms", + "weight": 40, + "params": { + "gt_value": 6.807735444943595e-08, + "tolerance": 6.807735444943596e-09 + } + }, + { + "metric_id": "rem_like_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 60, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_02.edf. The analyzed event window is 462.0-465.0 minutes. The current window is a true REM segment. The EMG channels are X1, X3, X4. The bandpass filter is 10-80Hz. The ground-truth event-window EMG RMS is 6.80773544494e-08; the whole-night EMG RMS is 1.04680457673e-06; the event/whole-night RMS ratio is 0.065033. The REM-like threshold is event RMS <= 0.8 * whole-night RMS. The correct final REM-like conclusion is: REM-like. The correct response must state that the event-window EMG RMS is significantly lower than the whole-night EMG RMS and must judge the window as REM or REM-like. Mark the response as correct only if it uses the RMS comparison consistently and gives the correct REM-like versus non-REM-like conclusion. If the response gives the opposite conclusion, ignores the RMS comparison, claims REM-like for a true Non-REM segment, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case22/case22_03.json b/sleep_assessment/cases/case22/case22_03.json new file mode 100644 index 0000000000000000000000000000000000000000..699d88c675a8b6efea4b96a4f2ae0128dad2725c --- /dev/null +++ b/sleep_assessment/cases/case22/case22_03.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "instruction": "Please analyze the provided sleep file in the time window from minute 175.0 to minute 178.0. For this ISRUC file, X1, X3, and X4 are EMG channels. Select the chin EMG/EMG channel(s), apply a 10-80Hz bandpass filter appropriate for the 200Hz sampling rate, and compute the mean chin EMG RMS value in this event window. Also compute the whole-night EMG RMS mean using the same EMG channel(s) and filter. Compare the event-window RMS with the whole-night RMS, determine whether the event window follows a REM-like pattern, and give a brief analysis. Clearly report the event-window EMG RMS, the whole-night EMG RMS, and the final REM-like or non-REM-like judgment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the event-window chin EMG RMS value and preserve the complete report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"event_emg_rms\" and \"full_text\".\n5. \"event_emg_rms\" must be a FLOAT or null. Do not include units.\n6. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n7. If the event-window EMG RMS value is missing or unclear, set \"event_emg_rms\" to null.\n\n### OUTPUT TEMPLATE\n{\"event_emg_rms\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "event_chin_emg_rms_accuracy", + "type": "numeric_check", + "target_key": "event_emg_rms", + "weight": 40, + "params": { + "gt_value": 1.63447757802915e-07, + "tolerance": 1.6344775780291502e-08 + } + }, + { + "metric_id": "rem_like_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 60, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_03.edf. The analyzed event window is 175.0-178.0 minutes. The current window is a true REM segment. The EMG channels are X1, X3, X4. The bandpass filter is 10-80Hz. The ground-truth event-window EMG RMS is 1.63447757803e-07; the whole-night EMG RMS is 9.73054813996e-07; the event/whole-night RMS ratio is 0.167974. The REM-like threshold is event RMS <= 0.8 * whole-night RMS. The correct final REM-like conclusion is: REM-like. The correct response must state that the event-window EMG RMS is significantly lower than the whole-night EMG RMS and must judge the window as REM or REM-like. Mark the response as correct only if it uses the RMS comparison consistently and gives the correct REM-like versus non-REM-like conclusion. If the response gives the opposite conclusion, ignores the RMS comparison, claims REM-like for a true Non-REM segment, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case22/case22_04.json b/sleep_assessment/cases/case22/case22_04.json new file mode 100644 index 0000000000000000000000000000000000000000..1eaf2e246042ff868acc3006f8b41d4c17ebf12c --- /dev/null +++ b/sleep_assessment/cases/case22/case22_04.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "ISRUC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_04.edf", + "instruction": "Please analyze the provided sleep file in the time window from minute 253.0 to minute 256.0. For this ISRUC file, X1, X3, and X4 are EMG channels. Select the chin EMG/EMG channel(s), apply a 10-80Hz bandpass filter appropriate for the 200Hz sampling rate, and compute the mean chin EMG RMS value in this event window. Also compute the whole-night EMG RMS mean using the same EMG channel(s) and filter. Compare the event-window RMS with the whole-night RMS, determine whether the event window follows a REM-like pattern, and give a brief analysis. Clearly report the event-window EMG RMS, the whole-night EMG RMS, and the final REM-like or non-REM-like judgment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the event-window chin EMG RMS value and preserve the complete report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"event_emg_rms\" and \"full_text\".\n5. \"event_emg_rms\" must be a FLOAT or null. Do not include units.\n6. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n7. If the event-window EMG RMS value is missing or unclear, set \"event_emg_rms\" to null.\n\n### OUTPUT TEMPLATE\n{\"event_emg_rms\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "event_chin_emg_rms_accuracy", + "type": "numeric_check", + "target_key": "event_emg_rms", + "weight": 40, + "params": { + "gt_value": 1.7294296493802682e-06, + "tolerance": 1.7294296493802683e-07 + } + }, + { + "metric_id": "rem_like_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 60, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_04.edf. The analyzed event window is 253.0-256.0 minutes. The current window is a true Non-REM segment. The EMG channels are X1, X3, X4. The bandpass filter is 10-80Hz. The ground-truth event-window EMG RMS is 1.72942964938e-06; the whole-night EMG RMS is 4.6295173256e-07; the event/whole-night RMS ratio is 3.735659. The REM-like threshold is event RMS <= 0.8 * whole-night RMS. The correct final REM-like conclusion is: not REM-like. The correct response must judge the window as non-REM-like or not consistent with a REM-like pattern. Mark the response as correct only if it uses the RMS comparison consistently and gives the correct REM-like versus non-REM-like conclusion. If the response gives the opposite conclusion, ignores the RMS comparison, claims REM-like for a true Non-REM segment, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case22/case22_05.json b/sleep_assessment/cases/case22/case22_05.json new file mode 100644 index 0000000000000000000000000000000000000000..e74016ad0b48eaf13a62845e0733ecb1b96a0ebf --- /dev/null +++ b/sleep_assessment/cases/case22/case22_05.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "instruction": "Please analyze the provided sleep file in the time window from minute 143.0 to minute 146.0. For this ISRUC file, X1, X3, and X4 are EMG channels. Select the chin EMG/EMG channel(s), apply a 10-80Hz bandpass filter appropriate for the 200Hz sampling rate, and compute the mean chin EMG RMS value in this event window. Also compute the whole-night EMG RMS mean using the same EMG channel(s) and filter. Compare the event-window RMS with the whole-night RMS, determine whether the event window follows a REM-like pattern, and give a brief analysis. Clearly report the event-window EMG RMS, the whole-night EMG RMS, and the final REM-like or non-REM-like judgment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the event-window chin EMG RMS value and preserve the complete report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"event_emg_rms\" and \"full_text\".\n5. \"event_emg_rms\" must be a FLOAT or null. Do not include units.\n6. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n7. If the event-window EMG RMS value is missing or unclear, set \"event_emg_rms\" to null.\n\n### OUTPUT TEMPLATE\n{\"event_emg_rms\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "event_chin_emg_rms_accuracy", + "type": "numeric_check", + "target_key": "event_emg_rms", + "weight": 40, + "params": { + "gt_value": 1.2709966405964798e-06, + "tolerance": 1.27099664059648e-07 + } + }, + { + "metric_id": "rem_like_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 60, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_05.edf. The analyzed event window is 143.0-146.0 minutes. The current window is a true Non-REM segment. The EMG channels are X1, X3, X4. The bandpass filter is 10-80Hz. The ground-truth event-window EMG RMS is 1.2709966406e-06; the whole-night EMG RMS is 1.26925715064e-06; the event/whole-night RMS ratio is 1.001370. The REM-like threshold is event RMS <= 0.8 * whole-night RMS. The correct final REM-like conclusion is: not REM-like. The correct response must judge the window as non-REM-like or not consistent with a REM-like pattern. Mark the response as correct only if it uses the RMS comparison consistently and gives the correct REM-like versus non-REM-like conclusion. If the response gives the opposite conclusion, ignores the RMS comparison, claims REM-like for a true Non-REM segment, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case22/case22_06.json b/sleep_assessment/cases/case22/case22_06.json new file mode 100644 index 0000000000000000000000000000000000000000..d13ba34e42379d3e712756851b1694becc9dab3a --- /dev/null +++ b/sleep_assessment/cases/case22/case22_06.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "HMC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_01.edf", + "instruction": "Please analyze the provided sleep file in the time window from minute 365.5 to minute 368.5. Use the following EMG channel(s): EMG chin. Select the chin EMG/EMG channel(s), apply a 10-100Hz bandpass filter appropriate for the 256Hz sampling rate, and compute the mean chin EMG RMS value in this event window. Also compute the whole-night EMG RMS mean using the same EMG channel(s) and filter. Compare the event-window RMS with the whole-night RMS, determine whether the event window follows a REM-like pattern, and give a brief analysis. Clearly report the event-window EMG RMS, the whole-night EMG RMS, and the final REM-like or non-REM-like judgment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the event-window chin EMG RMS value and preserve the complete report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"event_emg_rms\" and \"full_text\".\n5. \"event_emg_rms\" must be a FLOAT or null. Do not include units.\n6. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n7. If the event-window EMG RMS value is missing or unclear, set \"event_emg_rms\" to null.\n\n### OUTPUT TEMPLATE\n{\"event_emg_rms\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "event_chin_emg_rms_accuracy", + "type": "numeric_check", + "target_key": "event_emg_rms", + "weight": 40, + "params": { + "gt_value": 1.2125094878058998e-06, + "tolerance": 1.2125094878059e-07 + } + }, + { + "metric_id": "rem_like_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 60, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is HMC_01.edf. The analyzed event window is 365.5-368.5 minutes. The current window is a true REM segment. The EMG channels are EMG chin. The bandpass filter is 10-100Hz. The ground-truth event-window EMG RMS is 1.21250948781e-06; the whole-night EMG RMS is 2.89342732424e-06; the event/whole-night RMS ratio is 0.419056. The REM-like threshold is event RMS <= 0.8 * whole-night RMS. The correct final REM-like conclusion is: REM-like. The correct response must state that the event-window EMG RMS is significantly lower than the whole-night EMG RMS and must judge the window as REM or REM-like. Mark the response as correct only if it uses the RMS comparison consistently and gives the correct REM-like versus non-REM-like conclusion. If the response gives the opposite conclusion, ignores the RMS comparison, claims REM-like for a true Non-REM segment, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case22/case22_07.json b/sleep_assessment/cases/case22/case22_07.json new file mode 100644 index 0000000000000000000000000000000000000000..48d6117742a5707487c76ab62037dbea71d8c5dc --- /dev/null +++ b/sleep_assessment/cases/case22/case22_07.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "HMC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_02.edf", + "instruction": "Please analyze the provided sleep file in the time window from minute 296.5 to minute 299.5. Use the following EMG channel(s): EMG chin. Select the chin EMG/EMG channel(s), apply a 10-100Hz bandpass filter appropriate for the 256Hz sampling rate, and compute the mean chin EMG RMS value in this event window. Also compute the whole-night EMG RMS mean using the same EMG channel(s) and filter. Compare the event-window RMS with the whole-night RMS, determine whether the event window follows a REM-like pattern, and give a brief analysis. Clearly report the event-window EMG RMS, the whole-night EMG RMS, and the final REM-like or non-REM-like judgment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the event-window chin EMG RMS value and preserve the complete report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"event_emg_rms\" and \"full_text\".\n5. \"event_emg_rms\" must be a FLOAT or null. Do not include units.\n6. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n7. If the event-window EMG RMS value is missing or unclear, set \"event_emg_rms\" to null.\n\n### OUTPUT TEMPLATE\n{\"event_emg_rms\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "event_chin_emg_rms_accuracy", + "type": "numeric_check", + "target_key": "event_emg_rms", + "weight": 40, + "params": { + "gt_value": 1.2927010134758717e-06, + "tolerance": 1.2927010134758718e-07 + } + }, + { + "metric_id": "rem_like_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 60, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is HMC_02.edf. The analyzed event window is 296.5-299.5 minutes. The current window is a true REM segment. The EMG channels are EMG chin. The bandpass filter is 10-100Hz. The ground-truth event-window EMG RMS is 1.29270101348e-06; the whole-night EMG RMS is 3.35097519953e-06; the event/whole-night RMS ratio is 0.385769. The REM-like threshold is event RMS <= 0.8 * whole-night RMS. The correct final REM-like conclusion is: REM-like. The correct response must state that the event-window EMG RMS is significantly lower than the whole-night EMG RMS and must judge the window as REM or REM-like. Mark the response as correct only if it uses the RMS comparison consistently and gives the correct REM-like versus non-REM-like conclusion. If the response gives the opposite conclusion, ignores the RMS comparison, claims REM-like for a true Non-REM segment, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case22/case22_08.json b/sleep_assessment/cases/case22/case22_08.json new file mode 100644 index 0000000000000000000000000000000000000000..b746104cf3804148094f9ece957d70d5de528b71 --- /dev/null +++ b/sleep_assessment/cases/case22/case22_08.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "HMC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_03.edf", + "instruction": "Please analyze the provided sleep file in the time window from minute 292.5 to minute 295.5. Use the following EMG channel(s): EMG chin. Select the chin EMG/EMG channel(s), apply a 10-100Hz bandpass filter appropriate for the 256Hz sampling rate, and compute the mean chin EMG RMS value in this event window. Also compute the whole-night EMG RMS mean using the same EMG channel(s) and filter. Compare the event-window RMS with the whole-night RMS, determine whether the event window follows a REM-like pattern, and give a brief analysis. Clearly report the event-window EMG RMS, the whole-night EMG RMS, and the final REM-like or non-REM-like judgment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the event-window chin EMG RMS value and preserve the complete report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"event_emg_rms\" and \"full_text\".\n5. \"event_emg_rms\" must be a FLOAT or null. Do not include units.\n6. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n7. If the event-window EMG RMS value is missing or unclear, set \"event_emg_rms\" to null.\n\n### OUTPUT TEMPLATE\n{\"event_emg_rms\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "event_chin_emg_rms_accuracy", + "type": "numeric_check", + "target_key": "event_emg_rms", + "weight": 40, + "params": { + "gt_value": 1.45519900392743e-06, + "tolerance": 1.4551990039274303e-07 + } + }, + { + "metric_id": "rem_like_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 60, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is HMC_03.edf. The analyzed event window is 292.5-295.5 minutes. The current window is a true REM segment. The EMG channels are EMG chin. The bandpass filter is 10-100Hz. The ground-truth event-window EMG RMS is 1.45519900393e-06; the whole-night EMG RMS is 4.77630487829e-06; the event/whole-night RMS ratio is 0.304670. The REM-like threshold is event RMS <= 0.8 * whole-night RMS. The correct final REM-like conclusion is: REM-like. The correct response must state that the event-window EMG RMS is significantly lower than the whole-night EMG RMS and must judge the window as REM or REM-like. Mark the response as correct only if it uses the RMS comparison consistently and gives the correct REM-like versus non-REM-like conclusion. If the response gives the opposite conclusion, ignores the RMS comparison, claims REM-like for a true Non-REM segment, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case22/case22_09.json b/sleep_assessment/cases/case22/case22_09.json new file mode 100644 index 0000000000000000000000000000000000000000..aab3868625627a88e9e8a688c54a2a3767b737f0 --- /dev/null +++ b/sleep_assessment/cases/case22/case22_09.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "HMC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_04.edf", + "instruction": "Please analyze the provided sleep file in the time window from minute 214.5 to minute 217.5. Use the following EMG channel(s): EMG chin. Select the chin EMG/EMG channel(s), apply a 10-100Hz bandpass filter appropriate for the 256Hz sampling rate, and compute the mean chin EMG RMS value in this event window. Also compute the whole-night EMG RMS mean using the same EMG channel(s) and filter. Compare the event-window RMS with the whole-night RMS, determine whether the event window follows a REM-like pattern, and give a brief analysis. Clearly report the event-window EMG RMS, the whole-night EMG RMS, and the final REM-like or non-REM-like judgment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the event-window chin EMG RMS value and preserve the complete report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"event_emg_rms\" and \"full_text\".\n5. \"event_emg_rms\" must be a FLOAT or null. Do not include units.\n6. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n7. If the event-window EMG RMS value is missing or unclear, set \"event_emg_rms\" to null.\n\n### OUTPUT TEMPLATE\n{\"event_emg_rms\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "event_chin_emg_rms_accuracy", + "type": "numeric_check", + "target_key": "event_emg_rms", + "weight": 40, + "params": { + "gt_value": 5.8366780036124085e-06, + "tolerance": 5.836678003612409e-07 + } + }, + { + "metric_id": "rem_like_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 60, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is HMC_04.edf. The analyzed event window is 214.5-217.5 minutes. The current window is a true Non-REM segment. The EMG channels are EMG chin. The bandpass filter is 10-100Hz. The ground-truth event-window EMG RMS is 5.83667800361e-06; the whole-night EMG RMS is 2.75387668133e-06; the event/whole-night RMS ratio is 2.119441. The REM-like threshold is event RMS <= 0.8 * whole-night RMS. The correct final REM-like conclusion is: not REM-like. The correct response must judge the window as non-REM-like or not consistent with a REM-like pattern. Mark the response as correct only if it uses the RMS comparison consistently and gives the correct REM-like versus non-REM-like conclusion. If the response gives the opposite conclusion, ignores the RMS comparison, claims REM-like for a true Non-REM segment, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case22/case22_10.json b/sleep_assessment/cases/case22/case22_10.json new file mode 100644 index 0000000000000000000000000000000000000000..e4ea0098997c68468a0f16313a974b1a27242a90 --- /dev/null +++ b/sleep_assessment/cases/case22/case22_10.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "HMC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_05.edf", + "instruction": "Please analyze the provided sleep file in the time window from minute 22.5 to minute 25.5. Use the following EMG channel(s): EMG chin. Select the chin EMG/EMG channel(s), apply a 10-100Hz bandpass filter appropriate for the 256Hz sampling rate, and compute the mean chin EMG RMS value in this event window. Also compute the whole-night EMG RMS mean using the same EMG channel(s) and filter. Compare the event-window RMS with the whole-night RMS, determine whether the event window follows a REM-like pattern, and give a brief analysis. Clearly report the event-window EMG RMS, the whole-night EMG RMS, and the final REM-like or non-REM-like judgment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the event-window chin EMG RMS value and preserve the complete report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"event_emg_rms\" and \"full_text\".\n5. \"event_emg_rms\" must be a FLOAT or null. Do not include units.\n6. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n7. If the event-window EMG RMS value is missing or unclear, set \"event_emg_rms\" to null.\n\n### OUTPUT TEMPLATE\n{\"event_emg_rms\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "event_chin_emg_rms_accuracy", + "type": "numeric_check", + "target_key": "event_emg_rms", + "weight": 40, + "params": { + "gt_value": 1.3648126561724187e-05, + "tolerance": 1.3648126561724189e-06 + } + }, + { + "metric_id": "rem_like_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 60, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is HMC_05.edf. The analyzed event window is 22.5-25.5 minutes. The current window is a true Non-REM segment. The EMG channels are EMG chin. The bandpass filter is 10-100Hz. The ground-truth event-window EMG RMS is 1.36481265617e-05; the whole-night EMG RMS is 8.68256865068e-06; the event/whole-night RMS ratio is 1.571900. The REM-like threshold is event RMS <= 0.8 * whole-night RMS. The correct final REM-like conclusion is: not REM-like. The correct response must judge the window as non-REM-like or not consistent with a REM-like pattern. Mark the response as correct only if it uses the RMS comparison consistently and gives the correct REM-like versus non-REM-like conclusion. If the response gives the opposite conclusion, ignores the RMS comparison, claims REM-like for a true Non-REM segment, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case22/case22_11.json b/sleep_assessment/cases/case22/case22_11.json new file mode 100644 index 0000000000000000000000000000000000000000..4c7f24c521180d7ae17ec8ef2f97e9a588cf4561 --- /dev/null +++ b/sleep_assessment/cases/case22/case22_11.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "instruction": "Please analyze the provided sleep file in the time window from minute 241.5 to minute 244.5. Use the following EMG channel(s): EMG. Select the chin EMG/EMG channel(s), apply a 10-50Hz bandpass filter appropriate for the 125Hz sampling rate, and compute the mean chin EMG RMS value in this event window. Also compute the whole-night EMG RMS mean using the same EMG channel(s) and filter. Compare the event-window RMS with the whole-night RMS, determine whether the event window follows a REM-like pattern, and give a brief analysis. Clearly report the event-window EMG RMS, the whole-night EMG RMS, and the final REM-like or non-REM-like judgment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the event-window chin EMG RMS value and preserve the complete report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"event_emg_rms\" and \"full_text\".\n5. \"event_emg_rms\" must be a FLOAT or null. Do not include units.\n6. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n7. If the event-window EMG RMS value is missing or unclear, set \"event_emg_rms\" to null.\n\n### OUTPUT TEMPLATE\n{\"event_emg_rms\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "event_chin_emg_rms_accuracy", + "type": "numeric_check", + "target_key": "event_emg_rms", + "weight": 40, + "params": { + "gt_value": 1.158347861593798e-06, + "tolerance": 1.1583478615937979e-07 + } + }, + { + "metric_id": "rem_like_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 60, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_01.edf. The analyzed event window is 241.5-244.5 minutes. The current window is a true REM segment. The EMG channels are EMG. The bandpass filter is 10-50Hz. The ground-truth event-window EMG RMS is 1.15834786159e-06; the whole-night EMG RMS is 8.79580293825e-06; the event/whole-night RMS ratio is 0.131693. The REM-like threshold is event RMS <= 0.8 * whole-night RMS. The correct final REM-like conclusion is: REM-like. The correct response must state that the event-window EMG RMS is significantly lower than the whole-night EMG RMS and must judge the window as REM or REM-like. Mark the response as correct only if it uses the RMS comparison consistently and gives the correct REM-like versus non-REM-like conclusion. If the response gives the opposite conclusion, ignores the RMS comparison, claims REM-like for a true Non-REM segment, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case22/case22_12.json b/sleep_assessment/cases/case22/case22_12.json new file mode 100644 index 0000000000000000000000000000000000000000..1f0e104fed826a04be47dddb2d47ecbd7090d08b --- /dev/null +++ b/sleep_assessment/cases/case22/case22_12.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "instruction": "Please analyze the provided sleep file in the time window from minute 152.0 to minute 155.0. Use the following EMG channel(s): EMG. Select the chin EMG/EMG channel(s), apply a 10-50Hz bandpass filter appropriate for the 125Hz sampling rate, and compute the mean chin EMG RMS value in this event window. Also compute the whole-night EMG RMS mean using the same EMG channel(s) and filter. Compare the event-window RMS with the whole-night RMS, determine whether the event window follows a REM-like pattern, and give a brief analysis. Clearly report the event-window EMG RMS, the whole-night EMG RMS, and the final REM-like or non-REM-like judgment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the event-window chin EMG RMS value and preserve the complete report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"event_emg_rms\" and \"full_text\".\n5. \"event_emg_rms\" must be a FLOAT or null. Do not include units.\n6. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n7. If the event-window EMG RMS value is missing or unclear, set \"event_emg_rms\" to null.\n\n### OUTPUT TEMPLATE\n{\"event_emg_rms\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "event_chin_emg_rms_accuracy", + "type": "numeric_check", + "target_key": "event_emg_rms", + "weight": 40, + "params": { + "gt_value": 1.3714052883906417e-06, + "tolerance": 1.3714052883906417e-07 + } + }, + { + "metric_id": "rem_like_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 60, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_02.edf. The analyzed event window is 152.0-155.0 minutes. The current window is a true REM segment. The EMG channels are EMG. The bandpass filter is 10-50Hz. The ground-truth event-window EMG RMS is 1.37140528839e-06; the whole-night EMG RMS is 8.80517358953e-06; the event/whole-night RMS ratio is 0.155750. The REM-like threshold is event RMS <= 0.8 * whole-night RMS. The correct final REM-like conclusion is: REM-like. The correct response must state that the event-window EMG RMS is significantly lower than the whole-night EMG RMS and must judge the window as REM or REM-like. Mark the response as correct only if it uses the RMS comparison consistently and gives the correct REM-like versus non-REM-like conclusion. If the response gives the opposite conclusion, ignores the RMS comparison, claims REM-like for a true Non-REM segment, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case22/case22_13.json b/sleep_assessment/cases/case22/case22_13.json new file mode 100644 index 0000000000000000000000000000000000000000..8f838715e9428b59b1c682f1527013ed16c30614 --- /dev/null +++ b/sleep_assessment/cases/case22/case22_13.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "instruction": "Please analyze the provided sleep file in the time window from minute 448.0 to minute 451.0. Use the following EMG channel(s): EMG. Select the chin EMG/EMG channel(s), apply a 10-50Hz bandpass filter appropriate for the 125Hz sampling rate, and compute the mean chin EMG RMS value in this event window. Also compute the whole-night EMG RMS mean using the same EMG channel(s) and filter. Compare the event-window RMS with the whole-night RMS, determine whether the event window follows a REM-like pattern, and give a brief analysis. Clearly report the event-window EMG RMS, the whole-night EMG RMS, and the final REM-like or non-REM-like judgment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the event-window chin EMG RMS value and preserve the complete report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"event_emg_rms\" and \"full_text\".\n5. \"event_emg_rms\" must be a FLOAT or null. Do not include units.\n6. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n7. If the event-window EMG RMS value is missing or unclear, set \"event_emg_rms\" to null.\n\n### OUTPUT TEMPLATE\n{\"event_emg_rms\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "event_chin_emg_rms_accuracy", + "type": "numeric_check", + "target_key": "event_emg_rms", + "weight": 40, + "params": { + "gt_value": 1.047611254669687e-06, + "tolerance": 1.047611254669687e-07 + } + }, + { + "metric_id": "rem_like_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 60, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_03.edf. The analyzed event window is 448.0-451.0 minutes. The current window is a true REM segment. The EMG channels are EMG. The bandpass filter is 10-50Hz. The ground-truth event-window EMG RMS is 1.04761125467e-06; the whole-night EMG RMS is 6.41498034226e-06; the event/whole-night RMS ratio is 0.163307. The REM-like threshold is event RMS <= 0.8 * whole-night RMS. The correct final REM-like conclusion is: REM-like. The correct response must state that the event-window EMG RMS is significantly lower than the whole-night EMG RMS and must judge the window as REM or REM-like. Mark the response as correct only if it uses the RMS comparison consistently and gives the correct REM-like versus non-REM-like conclusion. If the response gives the opposite conclusion, ignores the RMS comparison, claims REM-like for a true Non-REM segment, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case22/case22_14.json b/sleep_assessment/cases/case22/case22_14.json new file mode 100644 index 0000000000000000000000000000000000000000..acd25789c1a2e00e26013e95d8ac69bca386e693 --- /dev/null +++ b/sleep_assessment/cases/case22/case22_14.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "instruction": "Please analyze the provided sleep file in the time window from minute 41.0 to minute 44.0. Use the following EMG channel(s): EMG. Select the chin EMG/EMG channel(s), apply a 10-50Hz bandpass filter appropriate for the 125Hz sampling rate, and compute the mean chin EMG RMS value in this event window. Also compute the whole-night EMG RMS mean using the same EMG channel(s) and filter. Compare the event-window RMS with the whole-night RMS, determine whether the event window follows a REM-like pattern, and give a brief analysis. Clearly report the event-window EMG RMS, the whole-night EMG RMS, and the final REM-like or non-REM-like judgment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the event-window chin EMG RMS value and preserve the complete report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"event_emg_rms\" and \"full_text\".\n5. \"event_emg_rms\" must be a FLOAT or null. Do not include units.\n6. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n7. If the event-window EMG RMS value is missing or unclear, set \"event_emg_rms\" to null.\n\n### OUTPUT TEMPLATE\n{\"event_emg_rms\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "event_chin_emg_rms_accuracy", + "type": "numeric_check", + "target_key": "event_emg_rms", + "weight": 40, + "params": { + "gt_value": 5.529400258158198e-06, + "tolerance": 5.529400258158197e-07 + } + }, + { + "metric_id": "rem_like_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 60, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_04.edf. The analyzed event window is 41.0-44.0 minutes. The current window is a true Non-REM segment. The EMG channels are EMG. The bandpass filter is 10-50Hz. The ground-truth event-window EMG RMS is 5.52940025816e-06; the whole-night EMG RMS is 4.92841082717e-06; the event/whole-night RMS ratio is 1.121944. The REM-like threshold is event RMS <= 0.8 * whole-night RMS. The correct final REM-like conclusion is: not REM-like. The correct response must judge the window as non-REM-like or not consistent with a REM-like pattern. Mark the response as correct only if it uses the RMS comparison consistently and gives the correct REM-like versus non-REM-like conclusion. If the response gives the opposite conclusion, ignores the RMS comparison, claims REM-like for a true Non-REM segment, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case22/case22_15.json b/sleep_assessment/cases/case22/case22_15.json new file mode 100644 index 0000000000000000000000000000000000000000..7d3626adf4d3732d37fded8086a3d91bb251c196 --- /dev/null +++ b/sleep_assessment/cases/case22/case22_15.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "instruction": "Please analyze the provided sleep file in the time window from minute 84.5 to minute 87.5. Use the following EMG channel(s): EMG. Select the chin EMG/EMG channel(s), apply a 10-50Hz bandpass filter appropriate for the 125Hz sampling rate, and compute the mean chin EMG RMS value in this event window. Also compute the whole-night EMG RMS mean using the same EMG channel(s) and filter. Compare the event-window RMS with the whole-night RMS, determine whether the event window follows a REM-like pattern, and give a brief analysis. Clearly report the event-window EMG RMS, the whole-night EMG RMS, and the final REM-like or non-REM-like judgment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the event-window chin EMG RMS value and preserve the complete report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"event_emg_rms\" and \"full_text\".\n5. \"event_emg_rms\" must be a FLOAT or null. Do not include units.\n6. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n7. If the event-window EMG RMS value is missing or unclear, set \"event_emg_rms\" to null.\n\n### OUTPUT TEMPLATE\n{\"event_emg_rms\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "event_chin_emg_rms_accuracy", + "type": "numeric_check", + "target_key": "event_emg_rms", + "weight": 40, + "params": { + "gt_value": 5.1888570663235356e-06, + "tolerance": 5.188857066323536e-07 + } + }, + { + "metric_id": "rem_like_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 60, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_05.edf. The analyzed event window is 84.5-87.5 minutes. The current window is a true Non-REM segment. The EMG channels are EMG. The bandpass filter is 10-50Hz. The ground-truth event-window EMG RMS is 5.18885706632e-06; the whole-night EMG RMS is 5.07710627434e-06; the event/whole-night RMS ratio is 1.022011. The REM-like threshold is event RMS <= 0.8 * whole-night RMS. The correct final REM-like conclusion is: not REM-like. The correct response must judge the window as non-REM-like or not consistent with a REM-like pattern. Mark the response as correct only if it uses the RMS comparison consistently and gives the correct REM-like versus non-REM-like conclusion. If the response gives the opposite conclusion, ignores the RMS comparison, claims REM-like for a true Non-REM segment, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case22/case22_16.json b/sleep_assessment/cases/case22/case22_16.json new file mode 100644 index 0000000000000000000000000000000000000000..b91f0b51dc71473e1ebe90428bde1c71b3757bc6 --- /dev/null +++ b/sleep_assessment/cases/case22/case22_16.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "MASSSS3_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_01.edf", + "instruction": "Please analyze the provided sleep file in the time window from minute 158.5 to minute 161.5. Use the following EMG channel(s): EMG Chin1, EMG Chin2, EMG Chin3. Select the chin EMG/EMG channel(s), apply a 10-100Hz bandpass filter appropriate for the 256Hz sampling rate, and compute the mean chin EMG RMS value in this event window. Also compute the whole-night EMG RMS mean using the same EMG channel(s) and filter. Compare the event-window RMS with the whole-night RMS, determine whether the event window follows a REM-like pattern, and give a brief analysis. Clearly report the event-window EMG RMS, the whole-night EMG RMS, and the final REM-like or non-REM-like judgment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the event-window chin EMG RMS value and preserve the complete report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"event_emg_rms\" and \"full_text\".\n5. \"event_emg_rms\" must be a FLOAT or null. Do not include units.\n6. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n7. If the event-window EMG RMS value is missing or unclear, set \"event_emg_rms\" to null.\n\n### OUTPUT TEMPLATE\n{\"event_emg_rms\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "event_chin_emg_rms_accuracy", + "type": "numeric_check", + "target_key": "event_emg_rms", + "weight": 40, + "params": { + "gt_value": 1.3221747030390449e-06, + "tolerance": 1.322174703039045e-07 + } + }, + { + "metric_id": "rem_like_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 60, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is MASSSS3_01.edf. The analyzed event window is 158.5-161.5 minutes. The current window is a true REM segment. The EMG channels are EMG Chin1, EMG Chin2, EMG Chin3. The bandpass filter is 10-100Hz. The ground-truth event-window EMG RMS is 1.32217470304e-06; the whole-night EMG RMS is 8.00297894656e-06; the event/whole-night RMS ratio is 0.165210. The REM-like threshold is event RMS <= 0.8 * whole-night RMS. The correct final REM-like conclusion is: REM-like. The correct response must state that the event-window EMG RMS is significantly lower than the whole-night EMG RMS and must judge the window as REM or REM-like. Mark the response as correct only if it uses the RMS comparison consistently and gives the correct REM-like versus non-REM-like conclusion. If the response gives the opposite conclusion, ignores the RMS comparison, claims REM-like for a true Non-REM segment, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case22/case22_17.json b/sleep_assessment/cases/case22/case22_17.json new file mode 100644 index 0000000000000000000000000000000000000000..8865f58c6ff90f9982ab35e456a6e134dc962a0c --- /dev/null +++ b/sleep_assessment/cases/case22/case22_17.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "MASSSS3_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_02.edf", + "instruction": "Please analyze the provided sleep file in the time window from minute 74.5 to minute 77.5. Use the following EMG channel(s): EMG Chin1, EMG Chin2, EMG Chin3. Select the chin EMG/EMG channel(s), apply a 10-100Hz bandpass filter appropriate for the 256Hz sampling rate, and compute the mean chin EMG RMS value in this event window. Also compute the whole-night EMG RMS mean using the same EMG channel(s) and filter. Compare the event-window RMS with the whole-night RMS, determine whether the event window follows a REM-like pattern, and give a brief analysis. Clearly report the event-window EMG RMS, the whole-night EMG RMS, and the final REM-like or non-REM-like judgment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the event-window chin EMG RMS value and preserve the complete report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"event_emg_rms\" and \"full_text\".\n5. \"event_emg_rms\" must be a FLOAT or null. Do not include units.\n6. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n7. If the event-window EMG RMS value is missing or unclear, set \"event_emg_rms\" to null.\n\n### OUTPUT TEMPLATE\n{\"event_emg_rms\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "event_chin_emg_rms_accuracy", + "type": "numeric_check", + "target_key": "event_emg_rms", + "weight": 40, + "params": { + "gt_value": 8.410101287302534e-07, + "tolerance": 8.410101287302535e-08 + } + }, + { + "metric_id": "rem_like_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 60, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is MASSSS3_02.edf. The analyzed event window is 74.5-77.5 minutes. The current window is a true REM segment. The EMG channels are EMG Chin1, EMG Chin2, EMG Chin3. The bandpass filter is 10-100Hz. The ground-truth event-window EMG RMS is 8.4101012873e-07; the whole-night EMG RMS is 3.85566057862e-06; the event/whole-night RMS ratio is 0.218123. The REM-like threshold is event RMS <= 0.8 * whole-night RMS. The correct final REM-like conclusion is: REM-like. The correct response must state that the event-window EMG RMS is significantly lower than the whole-night EMG RMS and must judge the window as REM or REM-like. Mark the response as correct only if it uses the RMS comparison consistently and gives the correct REM-like versus non-REM-like conclusion. If the response gives the opposite conclusion, ignores the RMS comparison, claims REM-like for a true Non-REM segment, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case22/case22_18.json b/sleep_assessment/cases/case22/case22_18.json new file mode 100644 index 0000000000000000000000000000000000000000..7cab5324599dc75968f82ee51fdc062aec8a9c3c --- /dev/null +++ b/sleep_assessment/cases/case22/case22_18.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "MASSSS3_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_03.edf", + "instruction": "Please analyze the provided sleep file in the time window from minute 356.5 to minute 359.5. Use the following EMG channel(s): EMG Chin1, EMG Chin2, EMG Chin3. Select the chin EMG/EMG channel(s), apply a 10-100Hz bandpass filter appropriate for the 256Hz sampling rate, and compute the mean chin EMG RMS value in this event window. Also compute the whole-night EMG RMS mean using the same EMG channel(s) and filter. Compare the event-window RMS with the whole-night RMS, determine whether the event window follows a REM-like pattern, and give a brief analysis. Clearly report the event-window EMG RMS, the whole-night EMG RMS, and the final REM-like or non-REM-like judgment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the event-window chin EMG RMS value and preserve the complete report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"event_emg_rms\" and \"full_text\".\n5. \"event_emg_rms\" must be a FLOAT or null. Do not include units.\n6. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n7. If the event-window EMG RMS value is missing or unclear, set \"event_emg_rms\" to null.\n\n### OUTPUT TEMPLATE\n{\"event_emg_rms\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "event_chin_emg_rms_accuracy", + "type": "numeric_check", + "target_key": "event_emg_rms", + "weight": 40, + "params": { + "gt_value": 8.634979564874514e-07, + "tolerance": 8.634979564874515e-08 + } + }, + { + "metric_id": "rem_like_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 60, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is MASSSS3_03.edf. The analyzed event window is 356.5-359.5 minutes. The current window is a true REM segment. The EMG channels are EMG Chin1, EMG Chin2, EMG Chin3. The bandpass filter is 10-100Hz. The ground-truth event-window EMG RMS is 8.63497956487e-07; the whole-night EMG RMS is 4.78397922626e-06; the event/whole-night RMS ratio is 0.180498. The REM-like threshold is event RMS <= 0.8 * whole-night RMS. The correct final REM-like conclusion is: REM-like. The correct response must state that the event-window EMG RMS is significantly lower than the whole-night EMG RMS and must judge the window as REM or REM-like. Mark the response as correct only if it uses the RMS comparison consistently and gives the correct REM-like versus non-REM-like conclusion. If the response gives the opposite conclusion, ignores the RMS comparison, claims REM-like for a true Non-REM segment, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case22/case22_19.json b/sleep_assessment/cases/case22/case22_19.json new file mode 100644 index 0000000000000000000000000000000000000000..a409cad407b77fe3bc22b7c82335fd5db0954cb5 --- /dev/null +++ b/sleep_assessment/cases/case22/case22_19.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "MASSSS3_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_04.edf", + "instruction": "Please analyze the provided sleep file in the time window from minute 42.0 to minute 45.0. Use the following EMG channel(s): EMG Chin1, EMG Chin2, EMG Chin3. Select the chin EMG/EMG channel(s), apply a 10-100Hz bandpass filter appropriate for the 256Hz sampling rate, and compute the mean chin EMG RMS value in this event window. Also compute the whole-night EMG RMS mean using the same EMG channel(s) and filter. Compare the event-window RMS with the whole-night RMS, determine whether the event window follows a REM-like pattern, and give a brief analysis. Clearly report the event-window EMG RMS, the whole-night EMG RMS, and the final REM-like or non-REM-like judgment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the event-window chin EMG RMS value and preserve the complete report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"event_emg_rms\" and \"full_text\".\n5. \"event_emg_rms\" must be a FLOAT or null. Do not include units.\n6. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n7. If the event-window EMG RMS value is missing or unclear, set \"event_emg_rms\" to null.\n\n### OUTPUT TEMPLATE\n{\"event_emg_rms\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "event_chin_emg_rms_accuracy", + "type": "numeric_check", + "target_key": "event_emg_rms", + "weight": 40, + "params": { + "gt_value": 1.8436884876285053e-05, + "tolerance": 1.8436884876285053e-06 + } + }, + { + "metric_id": "rem_like_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 60, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is MASSSS3_04.edf. The analyzed event window is 42.0-45.0 minutes. The current window is a true Non-REM segment. The EMG channels are EMG Chin1, EMG Chin2, EMG Chin3. The bandpass filter is 10-100Hz. The ground-truth event-window EMG RMS is 1.84368848763e-05; the whole-night EMG RMS is 6.91086656454e-06; the event/whole-night RMS ratio is 2.667811. The REM-like threshold is event RMS <= 0.8 * whole-night RMS. The correct final REM-like conclusion is: not REM-like. The correct response must judge the window as non-REM-like or not consistent with a REM-like pattern. Mark the response as correct only if it uses the RMS comparison consistently and gives the correct REM-like versus non-REM-like conclusion. If the response gives the opposite conclusion, ignores the RMS comparison, claims REM-like for a true Non-REM segment, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case22/case22_20.json b/sleep_assessment/cases/case22/case22_20.json new file mode 100644 index 0000000000000000000000000000000000000000..651588865cd5d45adc94280d20f2fb8445e55567 --- /dev/null +++ b/sleep_assessment/cases/case22/case22_20.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "MASSSS3_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_05.edf", + "instruction": "Please analyze the provided sleep file in the time window from minute 93.5 to minute 96.5. Use the following EMG channel(s): EMG Chin1, EMG Chin2, EMG Chin3. Select the chin EMG/EMG channel(s), apply a 10-100Hz bandpass filter appropriate for the 256Hz sampling rate, and compute the mean chin EMG RMS value in this event window. Also compute the whole-night EMG RMS mean using the same EMG channel(s) and filter. Compare the event-window RMS with the whole-night RMS, determine whether the event window follows a REM-like pattern, and give a brief analysis. Clearly report the event-window EMG RMS, the whole-night EMG RMS, and the final REM-like or non-REM-like judgment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the event-window chin EMG RMS value and preserve the complete report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"event_emg_rms\" and \"full_text\".\n5. \"event_emg_rms\" must be a FLOAT or null. Do not include units.\n6. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n7. If the event-window EMG RMS value is missing or unclear, set \"event_emg_rms\" to null.\n\n### OUTPUT TEMPLATE\n{\"event_emg_rms\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "event_chin_emg_rms_accuracy", + "type": "numeric_check", + "target_key": "event_emg_rms", + "weight": 40, + "params": { + "gt_value": 2.226726128501951e-05, + "tolerance": 2.226726128501951e-06 + } + }, + { + "metric_id": "rem_like_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 60, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is MASSSS3_05.edf. The analyzed event window is 93.5-96.5 minutes. The current window is a true Non-REM segment. The EMG channels are EMG Chin1, EMG Chin2, EMG Chin3. The bandpass filter is 10-100Hz. The ground-truth event-window EMG RMS is 2.2267261285e-05; the whole-night EMG RMS is 7.47323196746e-06; the event/whole-night RMS ratio is 2.979603. The REM-like threshold is event RMS <= 0.8 * whole-night RMS. The correct final REM-like conclusion is: not REM-like. The correct response must judge the window as non-REM-like or not consistent with a REM-like pattern. Mark the response as correct only if it uses the RMS comparison consistently and gives the correct REM-like versus non-REM-like conclusion. If the response gives the opposite conclusion, ignores the RMS comparison, claims REM-like for a true Non-REM segment, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case22/case22_21.json b/sleep_assessment/cases/case22/case22_21.json new file mode 100644 index 0000000000000000000000000000000000000000..254776b7c2f8eb4c04385fa889aa84a5f72a8f81 --- /dev/null +++ b/sleep_assessment/cases/case22/case22_21.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "instruction": "Please analyze the provided sleep file in the time window from minute 239.5 to minute 242.5. Use the following EMG channel(s): Chin1-Chin2. Select the chin EMG/EMG channel(s), apply a 10-80Hz bandpass filter appropriate for the 200Hz sampling rate, and compute the mean chin EMG RMS value in this event window. Also compute the whole-night EMG RMS mean using the same EMG channel(s) and filter. Compare the event-window RMS with the whole-night RMS, determine whether the event window follows a REM-like pattern, and give a brief analysis. Clearly report the event-window EMG RMS, the whole-night EMG RMS, and the final REM-like or non-REM-like judgment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the event-window chin EMG RMS value and preserve the complete report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"event_emg_rms\" and \"full_text\".\n5. \"event_emg_rms\" must be a FLOAT or null. Do not include units.\n6. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n7. If the event-window EMG RMS value is missing or unclear, set \"event_emg_rms\" to null.\n\n### OUTPUT TEMPLATE\n{\"event_emg_rms\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "event_chin_emg_rms_accuracy", + "type": "numeric_check", + "target_key": "event_emg_rms", + "weight": 40, + "params": { + "gt_value": 6.9547744842769485e-06, + "tolerance": 6.954774484276949e-07 + } + }, + { + "metric_id": "rem_like_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 60, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is Physionet2018_01.edf. The analyzed event window is 239.5-242.5 minutes. The current window is a true REM segment. The EMG channels are Chin1-Chin2. The bandpass filter is 10-80Hz. The ground-truth event-window EMG RMS is 6.95477448428e-06; the whole-night EMG RMS is 1.97842728729e-05; the event/whole-night RMS ratio is 0.351530. The REM-like threshold is event RMS <= 0.8 * whole-night RMS. The correct final REM-like conclusion is: REM-like. The correct response must state that the event-window EMG RMS is significantly lower than the whole-night EMG RMS and must judge the window as REM or REM-like. Mark the response as correct only if it uses the RMS comparison consistently and gives the correct REM-like versus non-REM-like conclusion. If the response gives the opposite conclusion, ignores the RMS comparison, claims REM-like for a true Non-REM segment, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case22/case22_22.json b/sleep_assessment/cases/case22/case22_22.json new file mode 100644 index 0000000000000000000000000000000000000000..22a899d5e9185b46e9cde3de13322daf2a39e91e --- /dev/null +++ b/sleep_assessment/cases/case22/case22_22.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "instruction": "Please analyze the provided sleep file in the time window from minute 138.0 to minute 141.0. Use the following EMG channel(s): Chin1-Chin2. Select the chin EMG/EMG channel(s), apply a 10-80Hz bandpass filter appropriate for the 200Hz sampling rate, and compute the mean chin EMG RMS value in this event window. Also compute the whole-night EMG RMS mean using the same EMG channel(s) and filter. Compare the event-window RMS with the whole-night RMS, determine whether the event window follows a REM-like pattern, and give a brief analysis. Clearly report the event-window EMG RMS, the whole-night EMG RMS, and the final REM-like or non-REM-like judgment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the event-window chin EMG RMS value and preserve the complete report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"event_emg_rms\" and \"full_text\".\n5. \"event_emg_rms\" must be a FLOAT or null. Do not include units.\n6. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n7. If the event-window EMG RMS value is missing or unclear, set \"event_emg_rms\" to null.\n\n### OUTPUT TEMPLATE\n{\"event_emg_rms\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "event_chin_emg_rms_accuracy", + "type": "numeric_check", + "target_key": "event_emg_rms", + "weight": 40, + "params": { + "gt_value": 9.421387291713087e-06, + "tolerance": 9.421387291713087e-07 + } + }, + { + "metric_id": "rem_like_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 60, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is Physionet2018_02.edf. The analyzed event window is 138.0-141.0 minutes. The current window is a true Non-REM segment. The EMG channels are Chin1-Chin2. The bandpass filter is 10-80Hz. The ground-truth event-window EMG RMS is 9.42138729171e-06; the whole-night EMG RMS is 1.66585452873e-05; the event/whole-night RMS ratio is 0.565559. The REM-like threshold is event RMS <= 0.8 * whole-night RMS. The correct final REM-like conclusion is: not REM-like. The correct response must judge the window as non-REM-like or not consistent with a REM-like pattern. Mark the response as correct only if it uses the RMS comparison consistently and gives the correct REM-like versus non-REM-like conclusion. If the response gives the opposite conclusion, ignores the RMS comparison, claims REM-like for a true Non-REM segment, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case22/case22_23.json b/sleep_assessment/cases/case22/case22_23.json new file mode 100644 index 0000000000000000000000000000000000000000..1a508ff5470e81f539507a4a6d837b2bde6df9a1 --- /dev/null +++ b/sleep_assessment/cases/case22/case22_23.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_03.edf", + "instruction": "Please analyze the provided sleep file in the time window from minute 261.0 to minute 264.0. Use the following EMG channel(s): Chin1-Chin2. Select the chin EMG/EMG channel(s), apply a 10-80Hz bandpass filter appropriate for the 200Hz sampling rate, and compute the mean chin EMG RMS value in this event window. Also compute the whole-night EMG RMS mean using the same EMG channel(s) and filter. Compare the event-window RMS with the whole-night RMS, determine whether the event window follows a REM-like pattern, and give a brief analysis. Clearly report the event-window EMG RMS, the whole-night EMG RMS, and the final REM-like or non-REM-like judgment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the event-window chin EMG RMS value and preserve the complete report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"event_emg_rms\" and \"full_text\".\n5. \"event_emg_rms\" must be a FLOAT or null. Do not include units.\n6. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n7. If the event-window EMG RMS value is missing or unclear, set \"event_emg_rms\" to null.\n\n### OUTPUT TEMPLATE\n{\"event_emg_rms\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "event_chin_emg_rms_accuracy", + "type": "numeric_check", + "target_key": "event_emg_rms", + "weight": 40, + "params": { + "gt_value": 1.1150316310254854e-05, + "tolerance": 1.1150316310254855e-06 + } + }, + { + "metric_id": "rem_like_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 60, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is Physionet2018_03.edf. The analyzed event window is 261.0-264.0 minutes. The current window is a true Non-REM segment. The EMG channels are Chin1-Chin2. The bandpass filter is 10-80Hz. The ground-truth event-window EMG RMS is 1.11503163103e-05; the whole-night EMG RMS is 2.05199197901e-05; the event/whole-night RMS ratio is 0.543390. The REM-like threshold is event RMS <= 0.8 * whole-night RMS. The correct final REM-like conclusion is: not REM-like. The correct response must judge the window as non-REM-like or not consistent with a REM-like pattern. Mark the response as correct only if it uses the RMS comparison consistently and gives the correct REM-like versus non-REM-like conclusion. If the response gives the opposite conclusion, ignores the RMS comparison, claims REM-like for a true Non-REM segment, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case22/case22_24.json b/sleep_assessment/cases/case22/case22_24.json new file mode 100644 index 0000000000000000000000000000000000000000..392e4cdb0538110902cdd6c7434902e1a8d9d7bc --- /dev/null +++ b/sleep_assessment/cases/case22/case22_24.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "instruction": "Please analyze the provided sleep file in the time window from minute 66.5 to minute 69.5. Use the following EMG channel(s): Chin1-Chin2. Select the chin EMG/EMG channel(s), apply a 10-80Hz bandpass filter appropriate for the 200Hz sampling rate, and compute the mean chin EMG RMS value in this event window. Also compute the whole-night EMG RMS mean using the same EMG channel(s) and filter. Compare the event-window RMS with the whole-night RMS, determine whether the event window follows a REM-like pattern, and give a brief analysis. Clearly report the event-window EMG RMS, the whole-night EMG RMS, and the final REM-like or non-REM-like judgment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the event-window chin EMG RMS value and preserve the complete report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"event_emg_rms\" and \"full_text\".\n5. \"event_emg_rms\" must be a FLOAT or null. Do not include units.\n6. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n7. If the event-window EMG RMS value is missing or unclear, set \"event_emg_rms\" to null.\n\n### OUTPUT TEMPLATE\n{\"event_emg_rms\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "event_chin_emg_rms_accuracy", + "type": "numeric_check", + "target_key": "event_emg_rms", + "weight": 40, + "params": { + "gt_value": 6.68070263010799e-06, + "tolerance": 6.680702630107991e-07 + } + }, + { + "metric_id": "rem_like_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 60, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is Physionet2018_04.edf. The analyzed event window is 66.5-69.5 minutes. The current window is a true REM segment. The EMG channels are Chin1-Chin2. The bandpass filter is 10-80Hz. The ground-truth event-window EMG RMS is 6.68070263011e-06; the whole-night EMG RMS is 7.96777951305e-05; the event/whole-night RMS ratio is 0.083846. The REM-like threshold is event RMS <= 0.8 * whole-night RMS. The correct final REM-like conclusion is: REM-like. The correct response must state that the event-window EMG RMS is significantly lower than the whole-night EMG RMS and must judge the window as REM or REM-like. Mark the response as correct only if it uses the RMS comparison consistently and gives the correct REM-like versus non-REM-like conclusion. If the response gives the opposite conclusion, ignores the RMS comparison, claims REM-like for a true Non-REM segment, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case22/case22_25.json b/sleep_assessment/cases/case22/case22_25.json new file mode 100644 index 0000000000000000000000000000000000000000..eaa2004cbb748dfe057809525978d8c5f679e523 --- /dev/null +++ b/sleep_assessment/cases/case22/case22_25.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "instruction": "Please analyze the provided sleep file in the time window from minute 371.5 to minute 374.5. Use the following EMG channel(s): Chin1-Chin2. Select the chin EMG/EMG channel(s), apply a 10-80Hz bandpass filter appropriate for the 200Hz sampling rate, and compute the mean chin EMG RMS value in this event window. Also compute the whole-night EMG RMS mean using the same EMG channel(s) and filter. Compare the event-window RMS with the whole-night RMS, determine whether the event window follows a REM-like pattern, and give a brief analysis. Clearly report the event-window EMG RMS, the whole-night EMG RMS, and the final REM-like or non-REM-like judgment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the event-window chin EMG RMS value and preserve the complete report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"event_emg_rms\" and \"full_text\".\n5. \"event_emg_rms\" must be a FLOAT or null. Do not include units.\n6. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n7. If the event-window EMG RMS value is missing or unclear, set \"event_emg_rms\" to null.\n\n### OUTPUT TEMPLATE\n{\"event_emg_rms\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "event_chin_emg_rms_accuracy", + "type": "numeric_check", + "target_key": "event_emg_rms", + "weight": 40, + "params": { + "gt_value": 8.298027416980718e-07, + "tolerance": 8.298027416980719e-08 + } + }, + { + "metric_id": "rem_like_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 60, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is Physionet2018_05.edf. The analyzed event window is 371.5-374.5 minutes. The current window is a true REM segment. The EMG channels are Chin1-Chin2. The bandpass filter is 10-80Hz. The ground-truth event-window EMG RMS is 8.29802741698e-07; the whole-night EMG RMS is 5.50541279185e-06; the event/whole-night RMS ratio is 0.150725. The REM-like threshold is event RMS <= 0.8 * whole-night RMS. The correct final REM-like conclusion is: REM-like. The correct response must state that the event-window EMG RMS is significantly lower than the whole-night EMG RMS and must judge the window as REM or REM-like. Mark the response as correct only if it uses the RMS comparison consistently and gives the correct REM-like versus non-REM-like conclusion. If the response gives the opposite conclusion, ignores the RMS comparison, claims REM-like for a true Non-REM segment, or contradicts these rules, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case23/case23_01.json b/sleep_assessment/cases/case23/case23_01.json new file mode 100644 index 0000000000000000000000000000000000000000..85e3076ce83f1217994510dc6ba59a57d9dd7228 --- /dev/null +++ b/sleep_assessment/cases/case23/case23_01.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "instruction": "Please analyze the provided overnight sleep EEG recording. Select all genuine EEG channels only. Apply a 0.3-40Hz bandpass filter to the selected EEG signals. Split the available recording duration into the first half of the night and the second half of the night by the temporal midpoint. For each half, compute the delta-band relative energy percentage, defined as the 0.5-4Hz band energy divided by the total 0.3-40Hz EEG energy, multiplied by 100. Report the first-half and second-half delta relative energy percentages, state which half shows stronger delta activity, and judge whether this pattern is consistent with classic sleep architecture." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two delta relative energy percentage values and preserve the complete report text for semantic judging:\n1) first-half delta relative energy percentage\n2) second-half delta relative energy percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_delta_relative_power_percent\", \"second_half_delta_relative_power_percent\", and \"full_text\".\n5. Both delta values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n8. If one delta value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_delta_relative_power_percent\": , \"second_half_delta_relative_power_percent\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "first_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "first_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 67.33469949370581, + "tolerance": 3.366734974685291 + } + }, + { + "metric_id": "second_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "second_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 56.33204300356318, + "tolerance": 2.8166021501781593 + } + }, + { + "metric_id": "half_night_delta_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_01.edf. The ground-truth first-half delta relative energy is 67.3346994937%, and the ground-truth second-half delta relative energy is 56.3320430036%. Based on these values, delta activity is stronger in the first half of the night. Given the classic sleep architecture rule that deep sleep/slow-wave activity occurs mainly in the first half of the night, judge whether the agent's reasoning is consistent with the ground-truth delta activity trend and whether its conclusion about classic sleep architecture is correct. If the true delta activity is stronger in the second half of the night, the agent must conclude that the pattern is not consistent with classic sleep architecture. Only when true delta activity is stronger in the first half, and the agent uses that trend to conclude that the pattern is consistent with classic sleep architecture, should the final logic be considered correct. The correct final conclusion for this sample is: consistent with classic sleep architecture. Do not penalize small numeric rounding differences, but mark the response incorrect if it gives the opposite trend, ignores the trend, or treats second-half-stronger or no-first-half-predominance findings as consistent with classic architecture. Output only a valid JSON object with exactly one key named status whose value is true or false. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case23/case23_02.json b/sleep_assessment/cases/case23/case23_02.json new file mode 100644 index 0000000000000000000000000000000000000000..1b079b30a91c61ed19fdbc7df16df64288d65c68 --- /dev/null +++ b/sleep_assessment/cases/case23/case23_02.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "instruction": "Please analyze the provided overnight sleep EEG recording. Select all genuine EEG channels only. Apply a 0.3-40Hz bandpass filter to the selected EEG signals. Split the available recording duration into the first half of the night and the second half of the night by the temporal midpoint. For each half, compute the delta-band relative energy percentage, defined as the 0.5-4Hz band energy divided by the total 0.3-40Hz EEG energy, multiplied by 100. Report the first-half and second-half delta relative energy percentages, state which half shows stronger delta activity, and judge whether this pattern is consistent with classic sleep architecture." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two delta relative energy percentage values and preserve the complete report text for semantic judging:\n1) first-half delta relative energy percentage\n2) second-half delta relative energy percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_delta_relative_power_percent\", \"second_half_delta_relative_power_percent\", and \"full_text\".\n5. Both delta values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n8. If one delta value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_delta_relative_power_percent\": , \"second_half_delta_relative_power_percent\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "first_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "first_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 59.74017061524712, + "tolerance": 2.987008530762356 + } + }, + { + "metric_id": "second_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "second_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 61.552457872627485, + "tolerance": 3.0776228936313745 + } + }, + { + "metric_id": "half_night_delta_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_02.edf. The ground-truth first-half delta relative energy is 59.7401706152%, and the ground-truth second-half delta relative energy is 61.5524578726%. Based on these values, delta activity is stronger in the second half of the night. Given the classic sleep architecture rule that deep sleep/slow-wave activity occurs mainly in the first half of the night, judge whether the agent's reasoning is consistent with the ground-truth delta activity trend and whether its conclusion about classic sleep architecture is correct. If the true delta activity is stronger in the second half of the night, the agent must conclude that the pattern is not consistent with classic sleep architecture. Only when true delta activity is stronger in the first half, and the agent uses that trend to conclude that the pattern is consistent with classic sleep architecture, should the final logic be considered correct. The correct final conclusion for this sample is: not consistent with classic sleep architecture. Do not penalize small numeric rounding differences, but mark the response incorrect if it gives the opposite trend, ignores the trend, or treats second-half-stronger or no-first-half-predominance findings as consistent with classic architecture. Output only a valid JSON object with exactly one key named status whose value is true or false. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case23/case23_03.json b/sleep_assessment/cases/case23/case23_03.json new file mode 100644 index 0000000000000000000000000000000000000000..da5eaeaf66c23fa2b99641f8d0942c90b0d6eb9c --- /dev/null +++ b/sleep_assessment/cases/case23/case23_03.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "instruction": "Please analyze the provided overnight sleep EEG recording. Select all genuine EEG channels only. Apply a 0.3-40Hz bandpass filter to the selected EEG signals. Split the available recording duration into the first half of the night and the second half of the night by the temporal midpoint. For each half, compute the delta-band relative energy percentage, defined as the 0.5-4Hz band energy divided by the total 0.3-40Hz EEG energy, multiplied by 100. Report the first-half and second-half delta relative energy percentages, state which half shows stronger delta activity, and judge whether this pattern is consistent with classic sleep architecture." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two delta relative energy percentage values and preserve the complete report text for semantic judging:\n1) first-half delta relative energy percentage\n2) second-half delta relative energy percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_delta_relative_power_percent\", \"second_half_delta_relative_power_percent\", and \"full_text\".\n5. Both delta values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n8. If one delta value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_delta_relative_power_percent\": , \"second_half_delta_relative_power_percent\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "first_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "first_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 64.41381533056132, + "tolerance": 3.220690766528066 + } + }, + { + "metric_id": "second_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "second_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 54.92228225706158, + "tolerance": 2.7461141128530793 + } + }, + { + "metric_id": "half_night_delta_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_03.edf. The ground-truth first-half delta relative energy is 64.4138153306%, and the ground-truth second-half delta relative energy is 54.9222822571%. Based on these values, delta activity is stronger in the first half of the night. Given the classic sleep architecture rule that deep sleep/slow-wave activity occurs mainly in the first half of the night, judge whether the agent's reasoning is consistent with the ground-truth delta activity trend and whether its conclusion about classic sleep architecture is correct. If the true delta activity is stronger in the second half of the night, the agent must conclude that the pattern is not consistent with classic sleep architecture. Only when true delta activity is stronger in the first half, and the agent uses that trend to conclude that the pattern is consistent with classic sleep architecture, should the final logic be considered correct. The correct final conclusion for this sample is: consistent with classic sleep architecture. Do not penalize small numeric rounding differences, but mark the response incorrect if it gives the opposite trend, ignores the trend, or treats second-half-stronger or no-first-half-predominance findings as consistent with classic architecture. Output only a valid JSON object with exactly one key named status whose value is true or false. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case23/case23_04.json b/sleep_assessment/cases/case23/case23_04.json new file mode 100644 index 0000000000000000000000000000000000000000..3811e8824b3b500ad03ee62eb373372a2d01ad60 --- /dev/null +++ b/sleep_assessment/cases/case23/case23_04.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "ISRUC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_04.edf", + "instruction": "Please analyze the provided overnight sleep EEG recording. Select all genuine EEG channels only. Apply a 0.3-40Hz bandpass filter to the selected EEG signals. Split the available recording duration into the first half of the night and the second half of the night by the temporal midpoint. For each half, compute the delta-band relative energy percentage, defined as the 0.5-4Hz band energy divided by the total 0.3-40Hz EEG energy, multiplied by 100. Report the first-half and second-half delta relative energy percentages, state which half shows stronger delta activity, and judge whether this pattern is consistent with classic sleep architecture." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two delta relative energy percentage values and preserve the complete report text for semantic judging:\n1) first-half delta relative energy percentage\n2) second-half delta relative energy percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_delta_relative_power_percent\", \"second_half_delta_relative_power_percent\", and \"full_text\".\n5. Both delta values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n8. If one delta value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_delta_relative_power_percent\": , \"second_half_delta_relative_power_percent\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "first_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "first_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 73.64779024300691, + "tolerance": 3.682389512150346 + } + }, + { + "metric_id": "second_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "second_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 64.5161251475125, + "tolerance": 3.225806257375625 + } + }, + { + "metric_id": "half_night_delta_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_04.edf. The ground-truth first-half delta relative energy is 73.647790243%, and the ground-truth second-half delta relative energy is 64.5161251475%. Based on these values, delta activity is stronger in the first half of the night. Given the classic sleep architecture rule that deep sleep/slow-wave activity occurs mainly in the first half of the night, judge whether the agent's reasoning is consistent with the ground-truth delta activity trend and whether its conclusion about classic sleep architecture is correct. If the true delta activity is stronger in the second half of the night, the agent must conclude that the pattern is not consistent with classic sleep architecture. Only when true delta activity is stronger in the first half, and the agent uses that trend to conclude that the pattern is consistent with classic sleep architecture, should the final logic be considered correct. The correct final conclusion for this sample is: consistent with classic sleep architecture. Do not penalize small numeric rounding differences, but mark the response incorrect if it gives the opposite trend, ignores the trend, or treats second-half-stronger or no-first-half-predominance findings as consistent with classic architecture. Output only a valid JSON object with exactly one key named status whose value is true or false. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case23/case23_05.json b/sleep_assessment/cases/case23/case23_05.json new file mode 100644 index 0000000000000000000000000000000000000000..f8a131ee59cdf2e99758ec25f39834326a921504 --- /dev/null +++ b/sleep_assessment/cases/case23/case23_05.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "instruction": "Please analyze the provided overnight sleep EEG recording. Select all genuine EEG channels only. Apply a 0.3-40Hz bandpass filter to the selected EEG signals. Split the available recording duration into the first half of the night and the second half of the night by the temporal midpoint. For each half, compute the delta-band relative energy percentage, defined as the 0.5-4Hz band energy divided by the total 0.3-40Hz EEG energy, multiplied by 100. Report the first-half and second-half delta relative energy percentages, state which half shows stronger delta activity, and judge whether this pattern is consistent with classic sleep architecture." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two delta relative energy percentage values and preserve the complete report text for semantic judging:\n1) first-half delta relative energy percentage\n2) second-half delta relative energy percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_delta_relative_power_percent\", \"second_half_delta_relative_power_percent\", and \"full_text\".\n5. Both delta values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n8. If one delta value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_delta_relative_power_percent\": , \"second_half_delta_relative_power_percent\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "first_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "first_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 60.99836719455589, + "tolerance": 3.0499183597277946 + } + }, + { + "metric_id": "second_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "second_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 58.41735443656893, + "tolerance": 2.9208677218284467 + } + }, + { + "metric_id": "half_night_delta_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_05.edf. The ground-truth first-half delta relative energy is 60.9983671946%, and the ground-truth second-half delta relative energy is 58.4173544366%. Based on these values, delta activity is stronger in the first half of the night. Given the classic sleep architecture rule that deep sleep/slow-wave activity occurs mainly in the first half of the night, judge whether the agent's reasoning is consistent with the ground-truth delta activity trend and whether its conclusion about classic sleep architecture is correct. If the true delta activity is stronger in the second half of the night, the agent must conclude that the pattern is not consistent with classic sleep architecture. Only when true delta activity is stronger in the first half, and the agent uses that trend to conclude that the pattern is consistent with classic sleep architecture, should the final logic be considered correct. The correct final conclusion for this sample is: consistent with classic sleep architecture. Do not penalize small numeric rounding differences, but mark the response incorrect if it gives the opposite trend, ignores the trend, or treats second-half-stronger or no-first-half-predominance findings as consistent with classic architecture. Output only a valid JSON object with exactly one key named status whose value is true or false. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case23/case23_06.json b/sleep_assessment/cases/case23/case23_06.json new file mode 100644 index 0000000000000000000000000000000000000000..d1c54249bee5ef320616a579f2a58c7dae52a650 --- /dev/null +++ b/sleep_assessment/cases/case23/case23_06.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "HMC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_01.edf", + "instruction": "Please analyze the provided overnight sleep EEG recording. Select all genuine EEG channels only. Apply a 0.3-40Hz bandpass filter to the selected EEG signals. Split the available recording duration into the first half of the night and the second half of the night by the temporal midpoint. For each half, compute the delta-band relative energy percentage, defined as the 0.5-4Hz band energy divided by the total 0.3-40Hz EEG energy, multiplied by 100. Report the first-half and second-half delta relative energy percentages, state which half shows stronger delta activity, and judge whether this pattern is consistent with classic sleep architecture." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two delta relative energy percentage values and preserve the complete report text for semantic judging:\n1) first-half delta relative energy percentage\n2) second-half delta relative energy percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_delta_relative_power_percent\", \"second_half_delta_relative_power_percent\", and \"full_text\".\n5. Both delta values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n8. If one delta value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_delta_relative_power_percent\": , \"second_half_delta_relative_power_percent\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "first_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "first_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 53.2454651557235, + "tolerance": 2.6622732577861754 + } + }, + { + "metric_id": "second_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "second_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 53.89565679480508, + "tolerance": 2.6947828397402542 + } + }, + { + "metric_id": "half_night_delta_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is HMC_01.edf. The ground-truth first-half delta relative energy is 53.2454651557%, and the ground-truth second-half delta relative energy is 53.8956567948%. Based on these values, delta activity is stronger in the second half of the night. Given the classic sleep architecture rule that deep sleep/slow-wave activity occurs mainly in the first half of the night, judge whether the agent's reasoning is consistent with the ground-truth delta activity trend and whether its conclusion about classic sleep architecture is correct. If the true delta activity is stronger in the second half of the night, the agent must conclude that the pattern is not consistent with classic sleep architecture. Only when true delta activity is stronger in the first half, and the agent uses that trend to conclude that the pattern is consistent with classic sleep architecture, should the final logic be considered correct. The correct final conclusion for this sample is: not consistent with classic sleep architecture. Do not penalize small numeric rounding differences, but mark the response incorrect if it gives the opposite trend, ignores the trend, or treats second-half-stronger or no-first-half-predominance findings as consistent with classic architecture. Output only a valid JSON object with exactly one key named status whose value is true or false. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case23/case23_07.json b/sleep_assessment/cases/case23/case23_07.json new file mode 100644 index 0000000000000000000000000000000000000000..43b3441ca374ebfa139738ba341c2910f7ce4f57 --- /dev/null +++ b/sleep_assessment/cases/case23/case23_07.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "HMC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_02.edf", + "instruction": "Please analyze the provided overnight sleep EEG recording. Select all genuine EEG channels only. Apply a 0.3-40Hz bandpass filter to the selected EEG signals. Split the available recording duration into the first half of the night and the second half of the night by the temporal midpoint. For each half, compute the delta-band relative energy percentage, defined as the 0.5-4Hz band energy divided by the total 0.3-40Hz EEG energy, multiplied by 100. Report the first-half and second-half delta relative energy percentages, state which half shows stronger delta activity, and judge whether this pattern is consistent with classic sleep architecture." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two delta relative energy percentage values and preserve the complete report text for semantic judging:\n1) first-half delta relative energy percentage\n2) second-half delta relative energy percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_delta_relative_power_percent\", \"second_half_delta_relative_power_percent\", and \"full_text\".\n5. Both delta values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n8. If one delta value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_delta_relative_power_percent\": , \"second_half_delta_relative_power_percent\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "first_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "first_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 68.2079546881756, + "tolerance": 3.4103977344087806 + } + }, + { + "metric_id": "second_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "second_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 64.71481590110388, + "tolerance": 3.235740795055194 + } + }, + { + "metric_id": "half_night_delta_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is HMC_02.edf. The ground-truth first-half delta relative energy is 68.2079546882%, and the ground-truth second-half delta relative energy is 64.7148159011%. Based on these values, delta activity is stronger in the first half of the night. Given the classic sleep architecture rule that deep sleep/slow-wave activity occurs mainly in the first half of the night, judge whether the agent's reasoning is consistent with the ground-truth delta activity trend and whether its conclusion about classic sleep architecture is correct. If the true delta activity is stronger in the second half of the night, the agent must conclude that the pattern is not consistent with classic sleep architecture. Only when true delta activity is stronger in the first half, and the agent uses that trend to conclude that the pattern is consistent with classic sleep architecture, should the final logic be considered correct. The correct final conclusion for this sample is: consistent with classic sleep architecture. Do not penalize small numeric rounding differences, but mark the response incorrect if it gives the opposite trend, ignores the trend, or treats second-half-stronger or no-first-half-predominance findings as consistent with classic architecture. Output only a valid JSON object with exactly one key named status whose value is true or false. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case23/case23_08.json b/sleep_assessment/cases/case23/case23_08.json new file mode 100644 index 0000000000000000000000000000000000000000..d7260248ed64ee5911ee640a945695b0085f3517 --- /dev/null +++ b/sleep_assessment/cases/case23/case23_08.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "HMC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_03.edf", + "instruction": "Please analyze the provided overnight sleep EEG recording. Select all genuine EEG channels only. Apply a 0.3-40Hz bandpass filter to the selected EEG signals. Split the available recording duration into the first half of the night and the second half of the night by the temporal midpoint. For each half, compute the delta-band relative energy percentage, defined as the 0.5-4Hz band energy divided by the total 0.3-40Hz EEG energy, multiplied by 100. Report the first-half and second-half delta relative energy percentages, state which half shows stronger delta activity, and judge whether this pattern is consistent with classic sleep architecture." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two delta relative energy percentage values and preserve the complete report text for semantic judging:\n1) first-half delta relative energy percentage\n2) second-half delta relative energy percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_delta_relative_power_percent\", \"second_half_delta_relative_power_percent\", and \"full_text\".\n5. Both delta values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n8. If one delta value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_delta_relative_power_percent\": , \"second_half_delta_relative_power_percent\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "first_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "first_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 63.84305095250431, + "tolerance": 3.192152547625216 + } + }, + { + "metric_id": "second_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "second_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 59.32618439168844, + "tolerance": 2.9663092195844225 + } + }, + { + "metric_id": "half_night_delta_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is HMC_03.edf. The ground-truth first-half delta relative energy is 63.8430509525%, and the ground-truth second-half delta relative energy is 59.3261843917%. Based on these values, delta activity is stronger in the first half of the night. Given the classic sleep architecture rule that deep sleep/slow-wave activity occurs mainly in the first half of the night, judge whether the agent's reasoning is consistent with the ground-truth delta activity trend and whether its conclusion about classic sleep architecture is correct. If the true delta activity is stronger in the second half of the night, the agent must conclude that the pattern is not consistent with classic sleep architecture. Only when true delta activity is stronger in the first half, and the agent uses that trend to conclude that the pattern is consistent with classic sleep architecture, should the final logic be considered correct. The correct final conclusion for this sample is: consistent with classic sleep architecture. Do not penalize small numeric rounding differences, but mark the response incorrect if it gives the opposite trend, ignores the trend, or treats second-half-stronger or no-first-half-predominance findings as consistent with classic architecture. Output only a valid JSON object with exactly one key named status whose value is true or false. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case23/case23_09.json b/sleep_assessment/cases/case23/case23_09.json new file mode 100644 index 0000000000000000000000000000000000000000..a53dc933e042c332d9d5fc57711557b764131ff3 --- /dev/null +++ b/sleep_assessment/cases/case23/case23_09.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "HMC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_04.edf", + "instruction": "Please analyze the provided overnight sleep EEG recording. Select all genuine EEG channels only. Apply a 0.3-40Hz bandpass filter to the selected EEG signals. Split the available recording duration into the first half of the night and the second half of the night by the temporal midpoint. For each half, compute the delta-band relative energy percentage, defined as the 0.5-4Hz band energy divided by the total 0.3-40Hz EEG energy, multiplied by 100. Report the first-half and second-half delta relative energy percentages, state which half shows stronger delta activity, and judge whether this pattern is consistent with classic sleep architecture." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two delta relative energy percentage values and preserve the complete report text for semantic judging:\n1) first-half delta relative energy percentage\n2) second-half delta relative energy percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_delta_relative_power_percent\", \"second_half_delta_relative_power_percent\", and \"full_text\".\n5. Both delta values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n8. If one delta value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_delta_relative_power_percent\": , \"second_half_delta_relative_power_percent\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "first_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "first_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 59.29554008793102, + "tolerance": 2.9647770043965513 + } + }, + { + "metric_id": "second_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "second_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 52.63817079488876, + "tolerance": 2.631908539744438 + } + }, + { + "metric_id": "half_night_delta_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is HMC_04.edf. The ground-truth first-half delta relative energy is 59.2955400879%, and the ground-truth second-half delta relative energy is 52.6381707949%. Based on these values, delta activity is stronger in the first half of the night. Given the classic sleep architecture rule that deep sleep/slow-wave activity occurs mainly in the first half of the night, judge whether the agent's reasoning is consistent with the ground-truth delta activity trend and whether its conclusion about classic sleep architecture is correct. If the true delta activity is stronger in the second half of the night, the agent must conclude that the pattern is not consistent with classic sleep architecture. Only when true delta activity is stronger in the first half, and the agent uses that trend to conclude that the pattern is consistent with classic sleep architecture, should the final logic be considered correct. The correct final conclusion for this sample is: consistent with classic sleep architecture. Do not penalize small numeric rounding differences, but mark the response incorrect if it gives the opposite trend, ignores the trend, or treats second-half-stronger or no-first-half-predominance findings as consistent with classic architecture. Output only a valid JSON object with exactly one key named status whose value is true or false. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case23/case23_10.json b/sleep_assessment/cases/case23/case23_10.json new file mode 100644 index 0000000000000000000000000000000000000000..8b765da5bf4acf78ab3c5c5cc97968d28e898e91 --- /dev/null +++ b/sleep_assessment/cases/case23/case23_10.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "HMC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_05.edf", + "instruction": "Please analyze the provided overnight sleep EEG recording. Select all genuine EEG channels only. Apply a 0.3-40Hz bandpass filter to the selected EEG signals. Split the available recording duration into the first half of the night and the second half of the night by the temporal midpoint. For each half, compute the delta-band relative energy percentage, defined as the 0.5-4Hz band energy divided by the total 0.3-40Hz EEG energy, multiplied by 100. Report the first-half and second-half delta relative energy percentages, state which half shows stronger delta activity, and judge whether this pattern is consistent with classic sleep architecture." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two delta relative energy percentage values and preserve the complete report text for semantic judging:\n1) first-half delta relative energy percentage\n2) second-half delta relative energy percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_delta_relative_power_percent\", \"second_half_delta_relative_power_percent\", and \"full_text\".\n5. Both delta values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n8. If one delta value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_delta_relative_power_percent\": , \"second_half_delta_relative_power_percent\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "first_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "first_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 62.38383614739112, + "tolerance": 3.1191918073695564 + } + }, + { + "metric_id": "second_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "second_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 63.20857081319707, + "tolerance": 3.160428540659854 + } + }, + { + "metric_id": "half_night_delta_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is HMC_05.edf. The ground-truth first-half delta relative energy is 62.3838361474%, and the ground-truth second-half delta relative energy is 63.2085708132%. Based on these values, delta activity is stronger in the second half of the night. Given the classic sleep architecture rule that deep sleep/slow-wave activity occurs mainly in the first half of the night, judge whether the agent's reasoning is consistent with the ground-truth delta activity trend and whether its conclusion about classic sleep architecture is correct. If the true delta activity is stronger in the second half of the night, the agent must conclude that the pattern is not consistent with classic sleep architecture. Only when true delta activity is stronger in the first half, and the agent uses that trend to conclude that the pattern is consistent with classic sleep architecture, should the final logic be considered correct. The correct final conclusion for this sample is: not consistent with classic sleep architecture. Do not penalize small numeric rounding differences, but mark the response incorrect if it gives the opposite trend, ignores the trend, or treats second-half-stronger or no-first-half-predominance findings as consistent with classic architecture. Output only a valid JSON object with exactly one key named status whose value is true or false. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case23/case23_11.json b/sleep_assessment/cases/case23/case23_11.json new file mode 100644 index 0000000000000000000000000000000000000000..1c03f5a334e898c9ca79b226538c6ec054df283e --- /dev/null +++ b/sleep_assessment/cases/case23/case23_11.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "instruction": "Please analyze the provided overnight sleep EEG recording. Select all genuine EEG channels only. Apply a 0.3-40Hz bandpass filter to the selected EEG signals. Split the available recording duration into the first half of the night and the second half of the night by the temporal midpoint. For each half, compute the delta-band relative energy percentage, defined as the 0.5-4Hz band energy divided by the total 0.3-40Hz EEG energy, multiplied by 100. Report the first-half and second-half delta relative energy percentages, state which half shows stronger delta activity, and judge whether this pattern is consistent with classic sleep architecture." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two delta relative energy percentage values and preserve the complete report text for semantic judging:\n1) first-half delta relative energy percentage\n2) second-half delta relative energy percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_delta_relative_power_percent\", \"second_half_delta_relative_power_percent\", and \"full_text\".\n5. Both delta values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n8. If one delta value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_delta_relative_power_percent\": , \"second_half_delta_relative_power_percent\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "first_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "first_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 54.94309662111741, + "tolerance": 2.747154831055871 + } + }, + { + "metric_id": "second_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "second_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 31.940083384351215, + "tolerance": 1.597004169217561 + } + }, + { + "metric_id": "half_night_delta_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_01.edf. The ground-truth first-half delta relative energy is 54.9430966211%, and the ground-truth second-half delta relative energy is 31.9400833844%. Based on these values, delta activity is stronger in the first half of the night. Given the classic sleep architecture rule that deep sleep/slow-wave activity occurs mainly in the first half of the night, judge whether the agent's reasoning is consistent with the ground-truth delta activity trend and whether its conclusion about classic sleep architecture is correct. If the true delta activity is stronger in the second half of the night, the agent must conclude that the pattern is not consistent with classic sleep architecture. Only when true delta activity is stronger in the first half, and the agent uses that trend to conclude that the pattern is consistent with classic sleep architecture, should the final logic be considered correct. The correct final conclusion for this sample is: consistent with classic sleep architecture. Do not penalize small numeric rounding differences, but mark the response incorrect if it gives the opposite trend, ignores the trend, or treats second-half-stronger or no-first-half-predominance findings as consistent with classic architecture. Output only a valid JSON object with exactly one key named status whose value is true or false. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case23/case23_12.json b/sleep_assessment/cases/case23/case23_12.json new file mode 100644 index 0000000000000000000000000000000000000000..d9c413b53256cf1c0cfaae21ec5f646ac74e216a --- /dev/null +++ b/sleep_assessment/cases/case23/case23_12.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "instruction": "Please analyze the provided overnight sleep EEG recording. Select all genuine EEG channels only. Apply a 0.3-40Hz bandpass filter to the selected EEG signals. Split the available recording duration into the first half of the night and the second half of the night by the temporal midpoint. For each half, compute the delta-band relative energy percentage, defined as the 0.5-4Hz band energy divided by the total 0.3-40Hz EEG energy, multiplied by 100. Report the first-half and second-half delta relative energy percentages, state which half shows stronger delta activity, and judge whether this pattern is consistent with classic sleep architecture." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two delta relative energy percentage values and preserve the complete report text for semantic judging:\n1) first-half delta relative energy percentage\n2) second-half delta relative energy percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_delta_relative_power_percent\", \"second_half_delta_relative_power_percent\", and \"full_text\".\n5. Both delta values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n8. If one delta value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_delta_relative_power_percent\": , \"second_half_delta_relative_power_percent\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "first_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "first_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 36.49717974900702, + "tolerance": 1.824858987450351 + } + }, + { + "metric_id": "second_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "second_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 6.116601437124146, + "tolerance": 0.3058300718562073 + } + }, + { + "metric_id": "half_night_delta_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_02.edf. The ground-truth first-half delta relative energy is 36.497179749%, and the ground-truth second-half delta relative energy is 6.11660143712%. Based on these values, delta activity is stronger in the first half of the night. Given the classic sleep architecture rule that deep sleep/slow-wave activity occurs mainly in the first half of the night, judge whether the agent's reasoning is consistent with the ground-truth delta activity trend and whether its conclusion about classic sleep architecture is correct. If the true delta activity is stronger in the second half of the night, the agent must conclude that the pattern is not consistent with classic sleep architecture. Only when true delta activity is stronger in the first half, and the agent uses that trend to conclude that the pattern is consistent with classic sleep architecture, should the final logic be considered correct. The correct final conclusion for this sample is: consistent with classic sleep architecture. Do not penalize small numeric rounding differences, but mark the response incorrect if it gives the opposite trend, ignores the trend, or treats second-half-stronger or no-first-half-predominance findings as consistent with classic architecture. Output only a valid JSON object with exactly one key named status whose value is true or false. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case23/case23_13.json b/sleep_assessment/cases/case23/case23_13.json new file mode 100644 index 0000000000000000000000000000000000000000..8793a5e39e5db783f2146ef0140f940b786561f5 --- /dev/null +++ b/sleep_assessment/cases/case23/case23_13.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "instruction": "Please analyze the provided overnight sleep EEG recording. Select all genuine EEG channels only. Apply a 0.3-40Hz bandpass filter to the selected EEG signals. Split the available recording duration into the first half of the night and the second half of the night by the temporal midpoint. For each half, compute the delta-band relative energy percentage, defined as the 0.5-4Hz band energy divided by the total 0.3-40Hz EEG energy, multiplied by 100. Report the first-half and second-half delta relative energy percentages, state which half shows stronger delta activity, and judge whether this pattern is consistent with classic sleep architecture." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two delta relative energy percentage values and preserve the complete report text for semantic judging:\n1) first-half delta relative energy percentage\n2) second-half delta relative energy percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_delta_relative_power_percent\", \"second_half_delta_relative_power_percent\", and \"full_text\".\n5. Both delta values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n8. If one delta value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_delta_relative_power_percent\": , \"second_half_delta_relative_power_percent\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "first_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "first_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 54.80242716070924, + "tolerance": 2.7401213580354624 + } + }, + { + "metric_id": "second_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "second_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 62.87760275000422, + "tolerance": 3.1438801375002114 + } + }, + { + "metric_id": "half_night_delta_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_03.edf. The ground-truth first-half delta relative energy is 54.8024271607%, and the ground-truth second-half delta relative energy is 62.87760275%. Based on these values, delta activity is stronger in the second half of the night. Given the classic sleep architecture rule that deep sleep/slow-wave activity occurs mainly in the first half of the night, judge whether the agent's reasoning is consistent with the ground-truth delta activity trend and whether its conclusion about classic sleep architecture is correct. If the true delta activity is stronger in the second half of the night, the agent must conclude that the pattern is not consistent with classic sleep architecture. Only when true delta activity is stronger in the first half, and the agent uses that trend to conclude that the pattern is consistent with classic sleep architecture, should the final logic be considered correct. The correct final conclusion for this sample is: not consistent with classic sleep architecture. Do not penalize small numeric rounding differences, but mark the response incorrect if it gives the opposite trend, ignores the trend, or treats second-half-stronger or no-first-half-predominance findings as consistent with classic architecture. Output only a valid JSON object with exactly one key named status whose value is true or false. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case23/case23_14.json b/sleep_assessment/cases/case23/case23_14.json new file mode 100644 index 0000000000000000000000000000000000000000..feb624774fcfaef2775a77393c2a9aebc8c74d85 --- /dev/null +++ b/sleep_assessment/cases/case23/case23_14.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "instruction": "Please analyze the provided overnight sleep EEG recording. Select all genuine EEG channels only. Apply a 0.3-40Hz bandpass filter to the selected EEG signals. Split the available recording duration into the first half of the night and the second half of the night by the temporal midpoint. For each half, compute the delta-band relative energy percentage, defined as the 0.5-4Hz band energy divided by the total 0.3-40Hz EEG energy, multiplied by 100. Report the first-half and second-half delta relative energy percentages, state which half shows stronger delta activity, and judge whether this pattern is consistent with classic sleep architecture." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two delta relative energy percentage values and preserve the complete report text for semantic judging:\n1) first-half delta relative energy percentage\n2) second-half delta relative energy percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_delta_relative_power_percent\", \"second_half_delta_relative_power_percent\", and \"full_text\".\n5. Both delta values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n8. If one delta value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_delta_relative_power_percent\": , \"second_half_delta_relative_power_percent\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "first_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "first_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 74.60040758551183, + "tolerance": 3.730020379275592 + } + }, + { + "metric_id": "second_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "second_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 59.16313938394403, + "tolerance": 2.9581569691972014 + } + }, + { + "metric_id": "half_night_delta_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_04.edf. The ground-truth first-half delta relative energy is 74.6004075855%, and the ground-truth second-half delta relative energy is 59.1631393839%. Based on these values, delta activity is stronger in the first half of the night. Given the classic sleep architecture rule that deep sleep/slow-wave activity occurs mainly in the first half of the night, judge whether the agent's reasoning is consistent with the ground-truth delta activity trend and whether its conclusion about classic sleep architecture is correct. If the true delta activity is stronger in the second half of the night, the agent must conclude that the pattern is not consistent with classic sleep architecture. Only when true delta activity is stronger in the first half, and the agent uses that trend to conclude that the pattern is consistent with classic sleep architecture, should the final logic be considered correct. The correct final conclusion for this sample is: consistent with classic sleep architecture. Do not penalize small numeric rounding differences, but mark the response incorrect if it gives the opposite trend, ignores the trend, or treats second-half-stronger or no-first-half-predominance findings as consistent with classic architecture. Output only a valid JSON object with exactly one key named status whose value is true or false. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case23/case23_15.json b/sleep_assessment/cases/case23/case23_15.json new file mode 100644 index 0000000000000000000000000000000000000000..8d7f7e6735fdf2ba43b683bfb0a22425a2822ad2 --- /dev/null +++ b/sleep_assessment/cases/case23/case23_15.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "instruction": "Please analyze the provided overnight sleep EEG recording. Select all genuine EEG channels only. Apply a 0.3-40Hz bandpass filter to the selected EEG signals. Split the available recording duration into the first half of the night and the second half of the night by the temporal midpoint. For each half, compute the delta-band relative energy percentage, defined as the 0.5-4Hz band energy divided by the total 0.3-40Hz EEG energy, multiplied by 100. Report the first-half and second-half delta relative energy percentages, state which half shows stronger delta activity, and judge whether this pattern is consistent with classic sleep architecture." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two delta relative energy percentage values and preserve the complete report text for semantic judging:\n1) first-half delta relative energy percentage\n2) second-half delta relative energy percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_delta_relative_power_percent\", \"second_half_delta_relative_power_percent\", and \"full_text\".\n5. Both delta values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n8. If one delta value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_delta_relative_power_percent\": , \"second_half_delta_relative_power_percent\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "first_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "first_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 58.7792740401091, + "tolerance": 2.9389637020054553 + } + }, + { + "metric_id": "second_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "second_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 51.92668533423487, + "tolerance": 2.5963342667117435 + } + }, + { + "metric_id": "half_night_delta_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_05.edf. The ground-truth first-half delta relative energy is 58.7792740401%, and the ground-truth second-half delta relative energy is 51.9266853342%. Based on these values, delta activity is stronger in the first half of the night. Given the classic sleep architecture rule that deep sleep/slow-wave activity occurs mainly in the first half of the night, judge whether the agent's reasoning is consistent with the ground-truth delta activity trend and whether its conclusion about classic sleep architecture is correct. If the true delta activity is stronger in the second half of the night, the agent must conclude that the pattern is not consistent with classic sleep architecture. Only when true delta activity is stronger in the first half, and the agent uses that trend to conclude that the pattern is consistent with classic sleep architecture, should the final logic be considered correct. The correct final conclusion for this sample is: consistent with classic sleep architecture. Do not penalize small numeric rounding differences, but mark the response incorrect if it gives the opposite trend, ignores the trend, or treats second-half-stronger or no-first-half-predominance findings as consistent with classic architecture. Output only a valid JSON object with exactly one key named status whose value is true or false. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case23/case23_16.json b/sleep_assessment/cases/case23/case23_16.json new file mode 100644 index 0000000000000000000000000000000000000000..001fe2bf15efb96494293ab22ca09f6b152e56c7 --- /dev/null +++ b/sleep_assessment/cases/case23/case23_16.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "MASSSS3_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_01.edf", + "instruction": "Please analyze the provided overnight sleep EEG recording. Select all genuine EEG channels only. Apply a 0.3-40Hz bandpass filter to the selected EEG signals. Split the available recording duration into the first half of the night and the second half of the night by the temporal midpoint. For each half, compute the delta-band relative energy percentage, defined as the 0.5-4Hz band energy divided by the total 0.3-40Hz EEG energy, multiplied by 100. Report the first-half and second-half delta relative energy percentages, state which half shows stronger delta activity, and judge whether this pattern is consistent with classic sleep architecture." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two delta relative energy percentage values and preserve the complete report text for semantic judging:\n1) first-half delta relative energy percentage\n2) second-half delta relative energy percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_delta_relative_power_percent\", \"second_half_delta_relative_power_percent\", and \"full_text\".\n5. Both delta values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n8. If one delta value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_delta_relative_power_percent\": , \"second_half_delta_relative_power_percent\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "first_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "first_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 76.73606841909654, + "tolerance": 3.836803420954827 + } + }, + { + "metric_id": "second_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "second_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 67.25899972754054, + "tolerance": 3.362949986377027 + } + }, + { + "metric_id": "half_night_delta_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is MASSSS3_01.edf. The ground-truth first-half delta relative energy is 76.7360684191%, and the ground-truth second-half delta relative energy is 67.2589997275%. Based on these values, delta activity is stronger in the first half of the night. Given the classic sleep architecture rule that deep sleep/slow-wave activity occurs mainly in the first half of the night, judge whether the agent's reasoning is consistent with the ground-truth delta activity trend and whether its conclusion about classic sleep architecture is correct. If the true delta activity is stronger in the second half of the night, the agent must conclude that the pattern is not consistent with classic sleep architecture. Only when true delta activity is stronger in the first half, and the agent uses that trend to conclude that the pattern is consistent with classic sleep architecture, should the final logic be considered correct. The correct final conclusion for this sample is: consistent with classic sleep architecture. Do not penalize small numeric rounding differences, but mark the response incorrect if it gives the opposite trend, ignores the trend, or treats second-half-stronger or no-first-half-predominance findings as consistent with classic architecture. Output only a valid JSON object with exactly one key named status whose value is true or false. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case23/case23_17.json b/sleep_assessment/cases/case23/case23_17.json new file mode 100644 index 0000000000000000000000000000000000000000..7c79dcee26e2b43afba24a50988be0f4a3f06a31 --- /dev/null +++ b/sleep_assessment/cases/case23/case23_17.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "MASSSS3_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_02.edf", + "instruction": "Please analyze the provided overnight sleep EEG recording. Select all genuine EEG channels only. Apply a 0.3-40Hz bandpass filter to the selected EEG signals. Split the available recording duration into the first half of the night and the second half of the night by the temporal midpoint. For each half, compute the delta-band relative energy percentage, defined as the 0.5-4Hz band energy divided by the total 0.3-40Hz EEG energy, multiplied by 100. Report the first-half and second-half delta relative energy percentages, state which half shows stronger delta activity, and judge whether this pattern is consistent with classic sleep architecture." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two delta relative energy percentage values and preserve the complete report text for semantic judging:\n1) first-half delta relative energy percentage\n2) second-half delta relative energy percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_delta_relative_power_percent\", \"second_half_delta_relative_power_percent\", and \"full_text\".\n5. Both delta values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n8. If one delta value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_delta_relative_power_percent\": , \"second_half_delta_relative_power_percent\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "first_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "first_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 72.53159928810908, + "tolerance": 3.626579964405454 + } + }, + { + "metric_id": "second_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "second_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 63.274320009356465, + "tolerance": 3.1637160004678235 + } + }, + { + "metric_id": "half_night_delta_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is MASSSS3_02.edf. The ground-truth first-half delta relative energy is 72.5315992881%, and the ground-truth second-half delta relative energy is 63.2743200094%. Based on these values, delta activity is stronger in the first half of the night. Given the classic sleep architecture rule that deep sleep/slow-wave activity occurs mainly in the first half of the night, judge whether the agent's reasoning is consistent with the ground-truth delta activity trend and whether its conclusion about classic sleep architecture is correct. If the true delta activity is stronger in the second half of the night, the agent must conclude that the pattern is not consistent with classic sleep architecture. Only when true delta activity is stronger in the first half, and the agent uses that trend to conclude that the pattern is consistent with classic sleep architecture, should the final logic be considered correct. The correct final conclusion for this sample is: consistent with classic sleep architecture. Do not penalize small numeric rounding differences, but mark the response incorrect if it gives the opposite trend, ignores the trend, or treats second-half-stronger or no-first-half-predominance findings as consistent with classic architecture. Output only a valid JSON object with exactly one key named status whose value is true or false. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case23/case23_18.json b/sleep_assessment/cases/case23/case23_18.json new file mode 100644 index 0000000000000000000000000000000000000000..dd423f792e06d2fd8153f33d1c0691d4c8dad007 --- /dev/null +++ b/sleep_assessment/cases/case23/case23_18.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "MASSSS3_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_03.edf", + "instruction": "Please analyze the provided overnight sleep EEG recording. Select all genuine EEG channels only. Apply a 0.3-40Hz bandpass filter to the selected EEG signals. Split the available recording duration into the first half of the night and the second half of the night by the temporal midpoint. For each half, compute the delta-band relative energy percentage, defined as the 0.5-4Hz band energy divided by the total 0.3-40Hz EEG energy, multiplied by 100. Report the first-half and second-half delta relative energy percentages, state which half shows stronger delta activity, and judge whether this pattern is consistent with classic sleep architecture." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two delta relative energy percentage values and preserve the complete report text for semantic judging:\n1) first-half delta relative energy percentage\n2) second-half delta relative energy percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_delta_relative_power_percent\", \"second_half_delta_relative_power_percent\", and \"full_text\".\n5. Both delta values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n8. If one delta value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_delta_relative_power_percent\": , \"second_half_delta_relative_power_percent\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "first_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "first_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 71.80063165302475, + "tolerance": 3.5900315826512377 + } + }, + { + "metric_id": "second_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "second_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 69.30178185761056, + "tolerance": 3.4650890928805285 + } + }, + { + "metric_id": "half_night_delta_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is MASSSS3_03.edf. The ground-truth first-half delta relative energy is 71.800631653%, and the ground-truth second-half delta relative energy is 69.3017818576%. Based on these values, delta activity is stronger in the first half of the night. Given the classic sleep architecture rule that deep sleep/slow-wave activity occurs mainly in the first half of the night, judge whether the agent's reasoning is consistent with the ground-truth delta activity trend and whether its conclusion about classic sleep architecture is correct. If the true delta activity is stronger in the second half of the night, the agent must conclude that the pattern is not consistent with classic sleep architecture. Only when true delta activity is stronger in the first half, and the agent uses that trend to conclude that the pattern is consistent with classic sleep architecture, should the final logic be considered correct. The correct final conclusion for this sample is: consistent with classic sleep architecture. Do not penalize small numeric rounding differences, but mark the response incorrect if it gives the opposite trend, ignores the trend, or treats second-half-stronger or no-first-half-predominance findings as consistent with classic architecture. Output only a valid JSON object with exactly one key named status whose value is true or false. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case23/case23_19.json b/sleep_assessment/cases/case23/case23_19.json new file mode 100644 index 0000000000000000000000000000000000000000..d2bbe429a1f1989b99d72f6640601caeaf61aea9 --- /dev/null +++ b/sleep_assessment/cases/case23/case23_19.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "MASSSS3_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_04.edf", + "instruction": "Please analyze the provided overnight sleep EEG recording. Select all genuine EEG channels only. Apply a 0.3-40Hz bandpass filter to the selected EEG signals. Split the available recording duration into the first half of the night and the second half of the night by the temporal midpoint. For each half, compute the delta-band relative energy percentage, defined as the 0.5-4Hz band energy divided by the total 0.3-40Hz EEG energy, multiplied by 100. Report the first-half and second-half delta relative energy percentages, state which half shows stronger delta activity, and judge whether this pattern is consistent with classic sleep architecture." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two delta relative energy percentage values and preserve the complete report text for semantic judging:\n1) first-half delta relative energy percentage\n2) second-half delta relative energy percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_delta_relative_power_percent\", \"second_half_delta_relative_power_percent\", and \"full_text\".\n5. Both delta values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n8. If one delta value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_delta_relative_power_percent\": , \"second_half_delta_relative_power_percent\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "first_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "first_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 75.72108377007636, + "tolerance": 3.7860541885038184 + } + }, + { + "metric_id": "second_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "second_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 62.98382025329574, + "tolerance": 3.149191012664787 + } + }, + { + "metric_id": "half_night_delta_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is MASSSS3_04.edf. The ground-truth first-half delta relative energy is 75.7210837701%, and the ground-truth second-half delta relative energy is 62.9838202533%. Based on these values, delta activity is stronger in the first half of the night. Given the classic sleep architecture rule that deep sleep/slow-wave activity occurs mainly in the first half of the night, judge whether the agent's reasoning is consistent with the ground-truth delta activity trend and whether its conclusion about classic sleep architecture is correct. If the true delta activity is stronger in the second half of the night, the agent must conclude that the pattern is not consistent with classic sleep architecture. Only when true delta activity is stronger in the first half, and the agent uses that trend to conclude that the pattern is consistent with classic sleep architecture, should the final logic be considered correct. The correct final conclusion for this sample is: consistent with classic sleep architecture. Do not penalize small numeric rounding differences, but mark the response incorrect if it gives the opposite trend, ignores the trend, or treats second-half-stronger or no-first-half-predominance findings as consistent with classic architecture. Output only a valid JSON object with exactly one key named status whose value is true or false. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case23/case23_20.json b/sleep_assessment/cases/case23/case23_20.json new file mode 100644 index 0000000000000000000000000000000000000000..f9f881562dd4985fea77fa11110d934924de9765 --- /dev/null +++ b/sleep_assessment/cases/case23/case23_20.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "MASSSS3_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_05.edf", + "instruction": "Please analyze the provided overnight sleep EEG recording. Select all genuine EEG channels only. Apply a 0.3-40Hz bandpass filter to the selected EEG signals. Split the available recording duration into the first half of the night and the second half of the night by the temporal midpoint. For each half, compute the delta-band relative energy percentage, defined as the 0.5-4Hz band energy divided by the total 0.3-40Hz EEG energy, multiplied by 100. Report the first-half and second-half delta relative energy percentages, state which half shows stronger delta activity, and judge whether this pattern is consistent with classic sleep architecture." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two delta relative energy percentage values and preserve the complete report text for semantic judging:\n1) first-half delta relative energy percentage\n2) second-half delta relative energy percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_delta_relative_power_percent\", \"second_half_delta_relative_power_percent\", and \"full_text\".\n5. Both delta values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n8. If one delta value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_delta_relative_power_percent\": , \"second_half_delta_relative_power_percent\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "first_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "first_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 81.66016329954789, + "tolerance": 4.083008164977395 + } + }, + { + "metric_id": "second_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "second_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 65.26062456325221, + "tolerance": 3.2630312281626104 + } + }, + { + "metric_id": "half_night_delta_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is MASSSS3_05.edf. The ground-truth first-half delta relative energy is 81.6601632995%, and the ground-truth second-half delta relative energy is 65.2606245633%. Based on these values, delta activity is stronger in the first half of the night. Given the classic sleep architecture rule that deep sleep/slow-wave activity occurs mainly in the first half of the night, judge whether the agent's reasoning is consistent with the ground-truth delta activity trend and whether its conclusion about classic sleep architecture is correct. If the true delta activity is stronger in the second half of the night, the agent must conclude that the pattern is not consistent with classic sleep architecture. Only when true delta activity is stronger in the first half, and the agent uses that trend to conclude that the pattern is consistent with classic sleep architecture, should the final logic be considered correct. The correct final conclusion for this sample is: consistent with classic sleep architecture. Do not penalize small numeric rounding differences, but mark the response incorrect if it gives the opposite trend, ignores the trend, or treats second-half-stronger or no-first-half-predominance findings as consistent with classic architecture. Output only a valid JSON object with exactly one key named status whose value is true or false. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case23/case23_21.json b/sleep_assessment/cases/case23/case23_21.json new file mode 100644 index 0000000000000000000000000000000000000000..f203657178cfaa4eb03e0438b7b78c05169ce8c5 --- /dev/null +++ b/sleep_assessment/cases/case23/case23_21.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "instruction": "Please analyze the provided overnight sleep EEG recording. Select all genuine EEG channels only. Apply a 0.3-40Hz bandpass filter to the selected EEG signals. Split the available recording duration into the first half of the night and the second half of the night by the temporal midpoint. For each half, compute the delta-band relative energy percentage, defined as the 0.5-4Hz band energy divided by the total 0.3-40Hz EEG energy, multiplied by 100. Report the first-half and second-half delta relative energy percentages, state which half shows stronger delta activity, and judge whether this pattern is consistent with classic sleep architecture." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two delta relative energy percentage values and preserve the complete report text for semantic judging:\n1) first-half delta relative energy percentage\n2) second-half delta relative energy percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_delta_relative_power_percent\", \"second_half_delta_relative_power_percent\", and \"full_text\".\n5. Both delta values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n8. If one delta value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_delta_relative_power_percent\": , \"second_half_delta_relative_power_percent\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "first_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "first_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 66.11556455168528, + "tolerance": 3.3057782275842644 + } + }, + { + "metric_id": "second_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "second_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 61.548092145181144, + "tolerance": 3.0774046072590573 + } + }, + { + "metric_id": "half_night_delta_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is Physionet2018_01.edf. The ground-truth first-half delta relative energy is 66.1155645517%, and the ground-truth second-half delta relative energy is 61.5480921452%. Based on these values, delta activity is stronger in the first half of the night. Given the classic sleep architecture rule that deep sleep/slow-wave activity occurs mainly in the first half of the night, judge whether the agent's reasoning is consistent with the ground-truth delta activity trend and whether its conclusion about classic sleep architecture is correct. If the true delta activity is stronger in the second half of the night, the agent must conclude that the pattern is not consistent with classic sleep architecture. Only when true delta activity is stronger in the first half, and the agent uses that trend to conclude that the pattern is consistent with classic sleep architecture, should the final logic be considered correct. The correct final conclusion for this sample is: consistent with classic sleep architecture. Do not penalize small numeric rounding differences, but mark the response incorrect if it gives the opposite trend, ignores the trend, or treats second-half-stronger or no-first-half-predominance findings as consistent with classic architecture. Output only a valid JSON object with exactly one key named status whose value is true or false. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case23/case23_22.json b/sleep_assessment/cases/case23/case23_22.json new file mode 100644 index 0000000000000000000000000000000000000000..54621e67c3a7e8ce6d40a5b2b92273597a6ee628 --- /dev/null +++ b/sleep_assessment/cases/case23/case23_22.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "instruction": "Please analyze the provided overnight sleep EEG recording. Select all genuine EEG channels only. Apply a 0.3-40Hz bandpass filter to the selected EEG signals. Split the available recording duration into the first half of the night and the second half of the night by the temporal midpoint. For each half, compute the delta-band relative energy percentage, defined as the 0.5-4Hz band energy divided by the total 0.3-40Hz EEG energy, multiplied by 100. Report the first-half and second-half delta relative energy percentages, state which half shows stronger delta activity, and judge whether this pattern is consistent with classic sleep architecture." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two delta relative energy percentage values and preserve the complete report text for semantic judging:\n1) first-half delta relative energy percentage\n2) second-half delta relative energy percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_delta_relative_power_percent\", \"second_half_delta_relative_power_percent\", and \"full_text\".\n5. Both delta values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n8. If one delta value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_delta_relative_power_percent\": , \"second_half_delta_relative_power_percent\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "first_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "first_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 51.363191485886006, + "tolerance": 2.5681595742943006 + } + }, + { + "metric_id": "second_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "second_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 68.09016875833315, + "tolerance": 3.4045084379166575 + } + }, + { + "metric_id": "half_night_delta_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is Physionet2018_02.edf. The ground-truth first-half delta relative energy is 51.3631914859%, and the ground-truth second-half delta relative energy is 68.0901687583%. Based on these values, delta activity is stronger in the second half of the night. Given the classic sleep architecture rule that deep sleep/slow-wave activity occurs mainly in the first half of the night, judge whether the agent's reasoning is consistent with the ground-truth delta activity trend and whether its conclusion about classic sleep architecture is correct. If the true delta activity is stronger in the second half of the night, the agent must conclude that the pattern is not consistent with classic sleep architecture. Only when true delta activity is stronger in the first half, and the agent uses that trend to conclude that the pattern is consistent with classic sleep architecture, should the final logic be considered correct. The correct final conclusion for this sample is: not consistent with classic sleep architecture. Do not penalize small numeric rounding differences, but mark the response incorrect if it gives the opposite trend, ignores the trend, or treats second-half-stronger or no-first-half-predominance findings as consistent with classic architecture. Output only a valid JSON object with exactly one key named status whose value is true or false. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case23/case23_23.json b/sleep_assessment/cases/case23/case23_23.json new file mode 100644 index 0000000000000000000000000000000000000000..bf13899660e7b1eb67dd4d1b31b60bd3e14b5229 --- /dev/null +++ b/sleep_assessment/cases/case23/case23_23.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "Physionet2018_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_03.edf", + "instruction": "Please analyze the provided overnight sleep EEG recording. Select all genuine EEG channels only. Apply a 0.3-40Hz bandpass filter to the selected EEG signals. Split the available recording duration into the first half of the night and the second half of the night by the temporal midpoint. For each half, compute the delta-band relative energy percentage, defined as the 0.5-4Hz band energy divided by the total 0.3-40Hz EEG energy, multiplied by 100. Report the first-half and second-half delta relative energy percentages, state which half shows stronger delta activity, and judge whether this pattern is consistent with classic sleep architecture." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two delta relative energy percentage values and preserve the complete report text for semantic judging:\n1) first-half delta relative energy percentage\n2) second-half delta relative energy percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_delta_relative_power_percent\", \"second_half_delta_relative_power_percent\", and \"full_text\".\n5. Both delta values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n8. If one delta value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_delta_relative_power_percent\": , \"second_half_delta_relative_power_percent\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "first_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "first_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 58.808126294987204, + "tolerance": 2.94040631474936 + } + }, + { + "metric_id": "second_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "second_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 55.85273881278936, + "tolerance": 2.7926369406394684 + } + }, + { + "metric_id": "half_night_delta_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is Physionet2018_03.edf. The ground-truth first-half delta relative energy is 58.808126295%, and the ground-truth second-half delta relative energy is 55.8527388128%. Based on these values, delta activity is stronger in the first half of the night. Given the classic sleep architecture rule that deep sleep/slow-wave activity occurs mainly in the first half of the night, judge whether the agent's reasoning is consistent with the ground-truth delta activity trend and whether its conclusion about classic sleep architecture is correct. If the true delta activity is stronger in the second half of the night, the agent must conclude that the pattern is not consistent with classic sleep architecture. Only when true delta activity is stronger in the first half, and the agent uses that trend to conclude that the pattern is consistent with classic sleep architecture, should the final logic be considered correct. The correct final conclusion for this sample is: consistent with classic sleep architecture. Do not penalize small numeric rounding differences, but mark the response incorrect if it gives the opposite trend, ignores the trend, or treats second-half-stronger or no-first-half-predominance findings as consistent with classic architecture. Output only a valid JSON object with exactly one key named status whose value is true or false. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case23/case23_24.json b/sleep_assessment/cases/case23/case23_24.json new file mode 100644 index 0000000000000000000000000000000000000000..1693264bb075b7c327479d691ce169390a297637 --- /dev/null +++ b/sleep_assessment/cases/case23/case23_24.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "instruction": "Please analyze the provided overnight sleep EEG recording. Select all genuine EEG channels only. Apply a 0.3-40Hz bandpass filter to the selected EEG signals. Split the available recording duration into the first half of the night and the second half of the night by the temporal midpoint. For each half, compute the delta-band relative energy percentage, defined as the 0.5-4Hz band energy divided by the total 0.3-40Hz EEG energy, multiplied by 100. Report the first-half and second-half delta relative energy percentages, state which half shows stronger delta activity, and judge whether this pattern is consistent with classic sleep architecture." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two delta relative energy percentage values and preserve the complete report text for semantic judging:\n1) first-half delta relative energy percentage\n2) second-half delta relative energy percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_delta_relative_power_percent\", \"second_half_delta_relative_power_percent\", and \"full_text\".\n5. Both delta values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n8. If one delta value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_delta_relative_power_percent\": , \"second_half_delta_relative_power_percent\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "first_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "first_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 71.07078653093416, + "tolerance": 3.553539326546708 + } + }, + { + "metric_id": "second_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "second_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 56.952817277383325, + "tolerance": 2.8476408638691666 + } + }, + { + "metric_id": "half_night_delta_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is Physionet2018_04.edf. The ground-truth first-half delta relative energy is 71.0707865309%, and the ground-truth second-half delta relative energy is 56.9528172774%. Based on these values, delta activity is stronger in the first half of the night. Given the classic sleep architecture rule that deep sleep/slow-wave activity occurs mainly in the first half of the night, judge whether the agent's reasoning is consistent with the ground-truth delta activity trend and whether its conclusion about classic sleep architecture is correct. If the true delta activity is stronger in the second half of the night, the agent must conclude that the pattern is not consistent with classic sleep architecture. Only when true delta activity is stronger in the first half, and the agent uses that trend to conclude that the pattern is consistent with classic sleep architecture, should the final logic be considered correct. The correct final conclusion for this sample is: consistent with classic sleep architecture. Do not penalize small numeric rounding differences, but mark the response incorrect if it gives the opposite trend, ignores the trend, or treats second-half-stronger or no-first-half-predominance findings as consistent with classic architecture. Output only a valid JSON object with exactly one key named status whose value is true or false. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case23/case23_25.json b/sleep_assessment/cases/case23/case23_25.json new file mode 100644 index 0000000000000000000000000000000000000000..f09aac83d12dc48b26110249a6a901a94e35f424 --- /dev/null +++ b/sleep_assessment/cases/case23/case23_25.json @@ -0,0 +1,46 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "instruction": "Please analyze the provided overnight sleep EEG recording. Select all genuine EEG channels only. Apply a 0.3-40Hz bandpass filter to the selected EEG signals. Split the available recording duration into the first half of the night and the second half of the night by the temporal midpoint. For each half, compute the delta-band relative energy percentage, defined as the 0.5-4Hz band energy divided by the total 0.3-40Hz EEG energy, multiplied by 100. Report the first-half and second-half delta relative energy percentages, state which half shows stronger delta activity, and judge whether this pattern is consistent with classic sleep architecture." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EEG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two delta relative energy percentage values and preserve the complete report text for semantic judging:\n1) first-half delta relative energy percentage\n2) second-half delta relative energy percentage\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"first_half_delta_relative_power_percent\", \"second_half_delta_relative_power_percent\", and \"full_text\".\n5. Both delta values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n8. If one delta value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"first_half_delta_relative_power_percent\": , \"second_half_delta_relative_power_percent\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "first_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "first_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 63.27427463871565, + "tolerance": 3.1637137319357826 + } + }, + { + "metric_id": "second_half_delta_relative_power_accuracy", + "type": "numeric_check", + "target_key": "second_half_delta_relative_power_percent", + "weight": 30, + "params": { + "gt_value": 61.09546931626567, + "tolerance": 3.054773465813284 + } + }, + { + "metric_id": "half_night_delta_pattern_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is Physionet2018_05.edf. The ground-truth first-half delta relative energy is 63.2742746387%, and the ground-truth second-half delta relative energy is 61.0954693163%. Based on these values, delta activity is stronger in the first half of the night. Given the classic sleep architecture rule that deep sleep/slow-wave activity occurs mainly in the first half of the night, judge whether the agent's reasoning is consistent with the ground-truth delta activity trend and whether its conclusion about classic sleep architecture is correct. If the true delta activity is stronger in the second half of the night, the agent must conclude that the pattern is not consistent with classic sleep architecture. Only when true delta activity is stronger in the first half, and the agent uses that trend to conclude that the pattern is consistent with classic sleep architecture, should the final logic be considered correct. The correct final conclusion for this sample is: consistent with classic sleep architecture. Do not penalize small numeric rounding differences, but mark the response incorrect if it gives the opposite trend, ignores the trend, or treats second-half-stronger or no-first-half-predominance findings as consistent with classic architecture. Output only a valid JSON object with exactly one key named status whose value is true or false. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case24/case24_01.json b/sleep_assessment/cases/case24/case24_01.json new file mode 100644 index 0000000000000000000000000000000000000000..4522142a1ef7ac508b068e58d33227ba2c214f19 --- /dev/null +++ b/sleep_assessment/cases/case24/case24_01.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "instruction": "Please use the provided overnight sleep file to perform automatic sleep staging for the entire recording. Based on the resulting sleep-stage sequence, report three final percentage values: sleep efficiency (SE), the percentage of total sleep time spent in N3 deep sleep, and the percentage of total sleep time spent in REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final percentage values from the agent's report:\n1) sleep efficiency (SE)\n2) the percentage of total sleep time spent in N3 deep sleep\n3) the percentage of total sleep time spent in REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"sleep_efficiency_percent\", \"n3_sleep_ratio_percent\", and \"rem_sleep_ratio_percent\".\n5. All three values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"sleep_efficiency_percent\": , \"n3_sleep_ratio_percent\": , \"rem_sleep_ratio_percent\": }", + "metrics": [ + { + "metric_id": "sleep_efficiency_accuracy", + "type": "numeric_check", + "target_key": "sleep_efficiency_percent", + "weight": 40, + "params": { + "gt_value": 70.0, + "tolerance": 8.0 + } + }, + { + "metric_id": "n3_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "n3_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 37.5, + "tolerance": 5.0 + } + }, + { + "metric_id": "rem_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "rem_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 19.155844155844157, + "tolerance": 5.0 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case24/case24_02.json b/sleep_assessment/cases/case24/case24_02.json new file mode 100644 index 0000000000000000000000000000000000000000..384520e84b51e1b19bd67261572c365f64259357 --- /dev/null +++ b/sleep_assessment/cases/case24/case24_02.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "instruction": "Please use the provided overnight sleep file to perform automatic sleep staging for the entire recording. Based on the resulting sleep-stage sequence, report three final percentage values: sleep efficiency (SE), the percentage of total sleep time spent in N3 deep sleep, and the percentage of total sleep time spent in REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final percentage values from the agent's report:\n1) sleep efficiency (SE)\n2) the percentage of total sleep time spent in N3 deep sleep\n3) the percentage of total sleep time spent in REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"sleep_efficiency_percent\", \"n3_sleep_ratio_percent\", and \"rem_sleep_ratio_percent\".\n5. All three values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"sleep_efficiency_percent\": , \"n3_sleep_ratio_percent\": , \"rem_sleep_ratio_percent\": }", + "metrics": [ + { + "metric_id": "sleep_efficiency_accuracy", + "type": "numeric_check", + "target_key": "sleep_efficiency_percent", + "weight": 40, + "params": { + "gt_value": 74.5850622406639, + "tolerance": 8.0 + } + }, + { + "metric_id": "n3_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "n3_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 21.835883171070932, + "tolerance": 5.0 + } + }, + { + "metric_id": "rem_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "rem_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 14.186369958275382, + "tolerance": 5.0 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case24/case24_03.json b/sleep_assessment/cases/case24/case24_03.json new file mode 100644 index 0000000000000000000000000000000000000000..40714a62bf54a4e604953bd66651bb6cd7b95130 --- /dev/null +++ b/sleep_assessment/cases/case24/case24_03.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "instruction": "Please use the provided overnight sleep file to perform automatic sleep staging for the entire recording. Based on the resulting sleep-stage sequence, report three final percentage values: sleep efficiency (SE), the percentage of total sleep time spent in N3 deep sleep, and the percentage of total sleep time spent in REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final percentage values from the agent's report:\n1) sleep efficiency (SE)\n2) the percentage of total sleep time spent in N3 deep sleep\n3) the percentage of total sleep time spent in REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"sleep_efficiency_percent\", \"n3_sleep_ratio_percent\", and \"rem_sleep_ratio_percent\".\n5. All three values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"sleep_efficiency_percent\": , \"n3_sleep_ratio_percent\": , \"rem_sleep_ratio_percent\": }", + "metrics": [ + { + "metric_id": "sleep_efficiency_accuracy", + "type": "numeric_check", + "target_key": "sleep_efficiency_percent", + "weight": 40, + "params": { + "gt_value": 86.00212089077412, + "tolerance": 5.0 + } + }, + { + "metric_id": "n3_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "n3_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 21.331689272503084, + "tolerance": 5.0 + } + }, + { + "metric_id": "rem_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "rem_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 27.990135635018497, + "tolerance": 6.0 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case24/case24_04.json b/sleep_assessment/cases/case24/case24_04.json new file mode 100644 index 0000000000000000000000000000000000000000..0d1a2220e4e239434ac2c0857989f914ea3204c0 --- /dev/null +++ b/sleep_assessment/cases/case24/case24_04.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "ISRUC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_04.edf", + "instruction": "Please use the provided overnight sleep file to perform automatic sleep staging for the entire recording. Based on the resulting sleep-stage sequence, report three final percentage values: sleep efficiency (SE), the percentage of total sleep time spent in N3 deep sleep, and the percentage of total sleep time spent in REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final percentage values from the agent's report:\n1) sleep efficiency (SE)\n2) the percentage of total sleep time spent in N3 deep sleep\n3) the percentage of total sleep time spent in REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"sleep_efficiency_percent\", \"n3_sleep_ratio_percent\", and \"rem_sleep_ratio_percent\".\n5. All three values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"sleep_efficiency_percent\": , \"n3_sleep_ratio_percent\": , \"rem_sleep_ratio_percent\": }", + "metrics": [ + { + "metric_id": "sleep_efficiency_accuracy", + "type": "numeric_check", + "target_key": "sleep_efficiency_percent", + "weight": 40, + "params": { + "gt_value": 97.09241952232607, + "tolerance": 5.0 + } + }, + { + "metric_id": "n3_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "n3_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 22.887700534759357, + "tolerance": 5.0 + } + }, + { + "metric_id": "rem_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "rem_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 24.59893048128342, + "tolerance": 5.0 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case24/case24_05.json b/sleep_assessment/cases/case24/case24_05.json new file mode 100644 index 0000000000000000000000000000000000000000..183e2965a55c01cbf11b679bf3301f2af9192e63 --- /dev/null +++ b/sleep_assessment/cases/case24/case24_05.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "instruction": "Please use the provided overnight sleep file to perform automatic sleep staging for the entire recording. Based on the resulting sleep-stage sequence, report three final percentage values: sleep efficiency (SE), the percentage of total sleep time spent in N3 deep sleep, and the percentage of total sleep time spent in REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final percentage values from the agent's report:\n1) sleep efficiency (SE)\n2) the percentage of total sleep time spent in N3 deep sleep\n3) the percentage of total sleep time spent in REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"sleep_efficiency_percent\", \"n3_sleep_ratio_percent\", and \"rem_sleep_ratio_percent\".\n5. All three values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"sleep_efficiency_percent\": , \"n3_sleep_ratio_percent\": , \"rem_sleep_ratio_percent\": }", + "metrics": [ + { + "metric_id": "sleep_efficiency_accuracy", + "type": "numeric_check", + "target_key": "sleep_efficiency_percent", + "weight": 40, + "params": { + "gt_value": 66.17142857142858, + "tolerance": 8.0 + } + }, + { + "metric_id": "n3_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "n3_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 28.324697754749568, + "tolerance": 5.0 + } + }, + { + "metric_id": "rem_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "rem_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 7.253886010362694, + "tolerance": 4.0 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case24/case24_06.json b/sleep_assessment/cases/case24/case24_06.json new file mode 100644 index 0000000000000000000000000000000000000000..8c5a9c382b9679114c9f8edfca8781c30449cb06 --- /dev/null +++ b/sleep_assessment/cases/case24/case24_06.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "HMC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_01.edf", + "instruction": "Please use the provided overnight sleep file to perform automatic sleep staging for the entire recording. Based on the resulting sleep-stage sequence, report three final percentage values: sleep efficiency (SE), the percentage of total sleep time spent in N3 deep sleep, and the percentage of total sleep time spent in REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final percentage values from the agent's report:\n1) sleep efficiency (SE)\n2) the percentage of total sleep time spent in N3 deep sleep\n3) the percentage of total sleep time spent in REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"sleep_efficiency_percent\", \"n3_sleep_ratio_percent\", and \"rem_sleep_ratio_percent\".\n5. All three values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"sleep_efficiency_percent\": , \"n3_sleep_ratio_percent\": , \"rem_sleep_ratio_percent\": }", + "metrics": [ + { + "metric_id": "sleep_efficiency_accuracy", + "type": "numeric_check", + "target_key": "sleep_efficiency_percent", + "weight": 40, + "params": { + "gt_value": 82.31850117096019, + "tolerance": 5.0 + } + }, + { + "metric_id": "n3_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "n3_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 3.271692745376956, + "tolerance": 4.0 + } + }, + { + "metric_id": "rem_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "rem_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 20.056899004267425, + "tolerance": 5.0 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case24/case24_07.json b/sleep_assessment/cases/case24/case24_07.json new file mode 100644 index 0000000000000000000000000000000000000000..25e9ebe043e3a5aa53c0f95847dd1a1f09fcc3b1 --- /dev/null +++ b/sleep_assessment/cases/case24/case24_07.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "HMC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_02.edf", + "instruction": "Please use the provided overnight sleep file to perform automatic sleep staging for the entire recording. Based on the resulting sleep-stage sequence, report three final percentage values: sleep efficiency (SE), the percentage of total sleep time spent in N3 deep sleep, and the percentage of total sleep time spent in REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final percentage values from the agent's report:\n1) sleep efficiency (SE)\n2) the percentage of total sleep time spent in N3 deep sleep\n3) the percentage of total sleep time spent in REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"sleep_efficiency_percent\", \"n3_sleep_ratio_percent\", and \"rem_sleep_ratio_percent\".\n5. All three values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"sleep_efficiency_percent\": , \"n3_sleep_ratio_percent\": , \"rem_sleep_ratio_percent\": }", + "metrics": [ + { + "metric_id": "sleep_efficiency_accuracy", + "type": "numeric_check", + "target_key": "sleep_efficiency_percent", + "weight": 40, + "params": { + "gt_value": 90.42056074766356, + "tolerance": 5.0 + } + }, + { + "metric_id": "n3_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "n3_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 28.940568475452196, + "tolerance": 5.0 + } + }, + { + "metric_id": "rem_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "rem_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 22.48062015503876, + "tolerance": 5.0 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case24/case24_08.json b/sleep_assessment/cases/case24/case24_08.json new file mode 100644 index 0000000000000000000000000000000000000000..a6e541aa6025debe3a7940b21ba12275dd59e3d5 --- /dev/null +++ b/sleep_assessment/cases/case24/case24_08.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "HMC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_03.edf", + "instruction": "Please use the provided overnight sleep file to perform automatic sleep staging for the entire recording. Based on the resulting sleep-stage sequence, report three final percentage values: sleep efficiency (SE), the percentage of total sleep time spent in N3 deep sleep, and the percentage of total sleep time spent in REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final percentage values from the agent's report:\n1) sleep efficiency (SE)\n2) the percentage of total sleep time spent in N3 deep sleep\n3) the percentage of total sleep time spent in REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"sleep_efficiency_percent\", \"n3_sleep_ratio_percent\", and \"rem_sleep_ratio_percent\".\n5. All three values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"sleep_efficiency_percent\": , \"n3_sleep_ratio_percent\": , \"rem_sleep_ratio_percent\": }", + "metrics": [ + { + "metric_id": "sleep_efficiency_accuracy", + "type": "numeric_check", + "target_key": "sleep_efficiency_percent", + "weight": 40, + "params": { + "gt_value": 57.75681341719078, + "tolerance": 8.0 + } + }, + { + "metric_id": "n3_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "n3_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 27.22323049001815, + "tolerance": 5.0 + } + }, + { + "metric_id": "rem_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "rem_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 27.586206896551722, + "tolerance": 6.0 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case24/case24_09.json b/sleep_assessment/cases/case24/case24_09.json new file mode 100644 index 0000000000000000000000000000000000000000..b21c0a77f189bfdea2dfb57f3cbe12b81e130331 --- /dev/null +++ b/sleep_assessment/cases/case24/case24_09.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "HMC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_04.edf", + "instruction": "Please use the provided overnight sleep file to perform automatic sleep staging for the entire recording. Based on the resulting sleep-stage sequence, report three final percentage values: sleep efficiency (SE), the percentage of total sleep time spent in N3 deep sleep, and the percentage of total sleep time spent in REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final percentage values from the agent's report:\n1) sleep efficiency (SE)\n2) the percentage of total sleep time spent in N3 deep sleep\n3) the percentage of total sleep time spent in REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"sleep_efficiency_percent\", \"n3_sleep_ratio_percent\", and \"rem_sleep_ratio_percent\".\n5. All three values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"sleep_efficiency_percent\": , \"n3_sleep_ratio_percent\": , \"rem_sleep_ratio_percent\": }", + "metrics": [ + { + "metric_id": "sleep_efficiency_accuracy", + "type": "numeric_check", + "target_key": "sleep_efficiency_percent", + "weight": 40, + "params": { + "gt_value": 92.1259842519685, + "tolerance": 5.0 + } + }, + { + "metric_id": "n3_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "n3_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 6.837606837606837, + "tolerance": 4.0 + } + }, + { + "metric_id": "rem_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "rem_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 26.38888888888889, + "tolerance": 6.0 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case24/case24_10.json b/sleep_assessment/cases/case24/case24_10.json new file mode 100644 index 0000000000000000000000000000000000000000..da4c2ef86831b18d6acdb51406992618199531c4 --- /dev/null +++ b/sleep_assessment/cases/case24/case24_10.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "HMC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_05.edf", + "instruction": "Please use the provided overnight sleep file to perform automatic sleep staging for the entire recording. Based on the resulting sleep-stage sequence, report three final percentage values: sleep efficiency (SE), the percentage of total sleep time spent in N3 deep sleep, and the percentage of total sleep time spent in REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final percentage values from the agent's report:\n1) sleep efficiency (SE)\n2) the percentage of total sleep time spent in N3 deep sleep\n3) the percentage of total sleep time spent in REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"sleep_efficiency_percent\", \"n3_sleep_ratio_percent\", and \"rem_sleep_ratio_percent\".\n5. All three values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"sleep_efficiency_percent\": , \"n3_sleep_ratio_percent\": , \"rem_sleep_ratio_percent\": }", + "metrics": [ + { + "metric_id": "sleep_efficiency_accuracy", + "type": "numeric_check", + "target_key": "sleep_efficiency_percent", + "weight": 40, + "params": { + "gt_value": 79.04066736183525, + "tolerance": 5.0 + } + }, + { + "metric_id": "n3_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "n3_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 19.65699208443272, + "tolerance": 5.0 + } + }, + { + "metric_id": "rem_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "rem_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 16.75461741424802, + "tolerance": 5.0 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case24/case24_11.json b/sleep_assessment/cases/case24/case24_11.json new file mode 100644 index 0000000000000000000000000000000000000000..8bbc818610a0f2fd46437b3eb7819f343a99e66a --- /dev/null +++ b/sleep_assessment/cases/case24/case24_11.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "instruction": "Please use the provided overnight sleep file to perform automatic sleep staging for the entire recording. Based on the resulting sleep-stage sequence, report three final percentage values: sleep efficiency (SE), the percentage of total sleep time spent in N3 deep sleep, and the percentage of total sleep time spent in REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final percentage values from the agent's report:\n1) sleep efficiency (SE)\n2) the percentage of total sleep time spent in N3 deep sleep\n3) the percentage of total sleep time spent in REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"sleep_efficiency_percent\", \"n3_sleep_ratio_percent\", and \"rem_sleep_ratio_percent\".\n5. All three values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"sleep_efficiency_percent\": , \"n3_sleep_ratio_percent\": , \"rem_sleep_ratio_percent\": }", + "metrics": [ + { + "metric_id": "sleep_efficiency_accuracy", + "type": "numeric_check", + "target_key": "sleep_efficiency_percent", + "weight": 40, + "params": { + "gt_value": 69.28044280442805, + "tolerance": 8.0 + } + }, + { + "metric_id": "n3_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "n3_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 19.30758988015979, + "tolerance": 5.0 + } + }, + { + "metric_id": "rem_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "rem_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 13.581890812250332, + "tolerance": 5.0 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case24/case24_12.json b/sleep_assessment/cases/case24/case24_12.json new file mode 100644 index 0000000000000000000000000000000000000000..7be9b85cecdce207f91e3de03a7d13de9fe8c304 --- /dev/null +++ b/sleep_assessment/cases/case24/case24_12.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "instruction": "Please use the provided overnight sleep file to perform automatic sleep staging for the entire recording. Based on the resulting sleep-stage sequence, report three final percentage values: sleep efficiency (SE), the percentage of total sleep time spent in N3 deep sleep, and the percentage of total sleep time spent in REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final percentage values from the agent's report:\n1) sleep efficiency (SE)\n2) the percentage of total sleep time spent in N3 deep sleep\n3) the percentage of total sleep time spent in REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"sleep_efficiency_percent\", \"n3_sleep_ratio_percent\", and \"rem_sleep_ratio_percent\".\n5. All three values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"sleep_efficiency_percent\": , \"n3_sleep_ratio_percent\": , \"rem_sleep_ratio_percent\": }", + "metrics": [ + { + "metric_id": "sleep_efficiency_accuracy", + "type": "numeric_check", + "target_key": "sleep_efficiency_percent", + "weight": 40, + "params": { + "gt_value": 33.734939759036145, + "tolerance": 10.0 + } + }, + { + "metric_id": "n3_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "n3_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 16.75824175824176, + "tolerance": 5.0 + } + }, + { + "metric_id": "rem_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "rem_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 16.75824175824176, + "tolerance": 5.0 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case24/case24_13.json b/sleep_assessment/cases/case24/case24_13.json new file mode 100644 index 0000000000000000000000000000000000000000..90814a0b6599c8ece09072c538196ee455c38a04 --- /dev/null +++ b/sleep_assessment/cases/case24/case24_13.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "instruction": "Please use the provided overnight sleep file to perform automatic sleep staging for the entire recording. Based on the resulting sleep-stage sequence, report three final percentage values: sleep efficiency (SE), the percentage of total sleep time spent in N3 deep sleep, and the percentage of total sleep time spent in REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final percentage values from the agent's report:\n1) sleep efficiency (SE)\n2) the percentage of total sleep time spent in N3 deep sleep\n3) the percentage of total sleep time spent in REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"sleep_efficiency_percent\", \"n3_sleep_ratio_percent\", and \"rem_sleep_ratio_percent\".\n5. All three values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"sleep_efficiency_percent\": , \"n3_sleep_ratio_percent\": , \"rem_sleep_ratio_percent\": }", + "metrics": [ + { + "metric_id": "sleep_efficiency_accuracy", + "type": "numeric_check", + "target_key": "sleep_efficiency_percent", + "weight": 40, + "params": { + "gt_value": 68.35081029551954, + "tolerance": 8.0 + } + }, + { + "metric_id": "n3_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "n3_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 42.81729428172943, + "tolerance": 5.0 + } + }, + { + "metric_id": "rem_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "rem_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 11.99442119944212, + "tolerance": 5.0 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case24/case24_14.json b/sleep_assessment/cases/case24/case24_14.json new file mode 100644 index 0000000000000000000000000000000000000000..117c0085e5b79ae73c6bebf46332b7b3e5928f33 --- /dev/null +++ b/sleep_assessment/cases/case24/case24_14.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "instruction": "Please use the provided overnight sleep file to perform automatic sleep staging for the entire recording. Based on the resulting sleep-stage sequence, report three final percentage values: sleep efficiency (SE), the percentage of total sleep time spent in N3 deep sleep, and the percentage of total sleep time spent in REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final percentage values from the agent's report:\n1) sleep efficiency (SE)\n2) the percentage of total sleep time spent in N3 deep sleep\n3) the percentage of total sleep time spent in REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"sleep_efficiency_percent\", \"n3_sleep_ratio_percent\", and \"rem_sleep_ratio_percent\".\n5. All three values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"sleep_efficiency_percent\": , \"n3_sleep_ratio_percent\": , \"rem_sleep_ratio_percent\": }", + "metrics": [ + { + "metric_id": "sleep_efficiency_accuracy", + "type": "numeric_check", + "target_key": "sleep_efficiency_percent", + "weight": 40, + "params": { + "gt_value": 68.8, + "tolerance": 8.0 + } + }, + { + "metric_id": "n3_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "n3_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 52.325581395348834, + "tolerance": 5.0 + } + }, + { + "metric_id": "rem_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "rem_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 15.282392026578073, + "tolerance": 5.0 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case24/case24_15.json b/sleep_assessment/cases/case24/case24_15.json new file mode 100644 index 0000000000000000000000000000000000000000..1159f13faa986ecc2deba7d370d586018803ef70 --- /dev/null +++ b/sleep_assessment/cases/case24/case24_15.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "instruction": "Please use the provided overnight sleep file to perform automatic sleep staging for the entire recording. Based on the resulting sleep-stage sequence, report three final percentage values: sleep efficiency (SE), the percentage of total sleep time spent in N3 deep sleep, and the percentage of total sleep time spent in REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final percentage values from the agent's report:\n1) sleep efficiency (SE)\n2) the percentage of total sleep time spent in N3 deep sleep\n3) the percentage of total sleep time spent in REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"sleep_efficiency_percent\", \"n3_sleep_ratio_percent\", and \"rem_sleep_ratio_percent\".\n5. All three values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"sleep_efficiency_percent\": , \"n3_sleep_ratio_percent\": , \"rem_sleep_ratio_percent\": }", + "metrics": [ + { + "metric_id": "sleep_efficiency_accuracy", + "type": "numeric_check", + "target_key": "sleep_efficiency_percent", + "weight": 40, + "params": { + "gt_value": 68.26568265682657, + "tolerance": 8.0 + } + }, + { + "metric_id": "n3_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "n3_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 13.378378378378379, + "tolerance": 4.0 + } + }, + { + "metric_id": "rem_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "rem_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 12.297297297297296, + "tolerance": 5.0 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case24/case24_16.json b/sleep_assessment/cases/case24/case24_16.json new file mode 100644 index 0000000000000000000000000000000000000000..ca0ef066da4f63298486a6884c790ee251896c61 --- /dev/null +++ b/sleep_assessment/cases/case24/case24_16.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "MASSSS3_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_01.edf", + "instruction": "Please use the provided overnight sleep file to perform automatic sleep staging for the entire recording. Based on the resulting sleep-stage sequence, report three final percentage values: sleep efficiency (SE), the percentage of total sleep time spent in N3 deep sleep, and the percentage of total sleep time spent in REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final percentage values from the agent's report:\n1) sleep efficiency (SE)\n2) the percentage of total sleep time spent in N3 deep sleep\n3) the percentage of total sleep time spent in REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"sleep_efficiency_percent\", \"n3_sleep_ratio_percent\", and \"rem_sleep_ratio_percent\".\n5. All three values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"sleep_efficiency_percent\": , \"n3_sleep_ratio_percent\": , \"rem_sleep_ratio_percent\": }", + "metrics": [ + { + "metric_id": "sleep_efficiency_accuracy", + "type": "numeric_check", + "target_key": "sleep_efficiency_percent", + "weight": 40, + "params": { + "gt_value": 93.1547619047619, + "tolerance": 5.0 + } + }, + { + "metric_id": "n3_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "n3_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 15.548455804046858, + "tolerance": 5.0 + } + }, + { + "metric_id": "rem_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "rem_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 19.275825346112885, + "tolerance": 5.0 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case24/case24_17.json b/sleep_assessment/cases/case24/case24_17.json new file mode 100644 index 0000000000000000000000000000000000000000..b228f3aa1b153c1a066550075cea103f2bab3c2f --- /dev/null +++ b/sleep_assessment/cases/case24/case24_17.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "MASSSS3_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_02.edf", + "instruction": "Please use the provided overnight sleep file to perform automatic sleep staging for the entire recording. Based on the resulting sleep-stage sequence, report three final percentage values: sleep efficiency (SE), the percentage of total sleep time spent in N3 deep sleep, and the percentage of total sleep time spent in REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final percentage values from the agent's report:\n1) sleep efficiency (SE)\n2) the percentage of total sleep time spent in N3 deep sleep\n3) the percentage of total sleep time spent in REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"sleep_efficiency_percent\", \"n3_sleep_ratio_percent\", and \"rem_sleep_ratio_percent\".\n5. All three values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"sleep_efficiency_percent\": , \"n3_sleep_ratio_percent\": , \"rem_sleep_ratio_percent\": }", + "metrics": [ + { + "metric_id": "sleep_efficiency_accuracy", + "type": "numeric_check", + "target_key": "sleep_efficiency_percent", + "weight": 40, + "params": { + "gt_value": 89.08582089552239, + "tolerance": 5.0 + } + }, + { + "metric_id": "n3_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "n3_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 18.219895287958114, + "tolerance": 5.0 + } + }, + { + "metric_id": "rem_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "rem_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 25.340314136125656, + "tolerance": 6.0 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case24/case24_18.json b/sleep_assessment/cases/case24/case24_18.json new file mode 100644 index 0000000000000000000000000000000000000000..15dad6ee729f7b57474eb04cccbec010f4722603 --- /dev/null +++ b/sleep_assessment/cases/case24/case24_18.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "MASSSS3_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_03.edf", + "instruction": "Please use the provided overnight sleep file to perform automatic sleep staging for the entire recording. Based on the resulting sleep-stage sequence, report three final percentage values: sleep efficiency (SE), the percentage of total sleep time spent in N3 deep sleep, and the percentage of total sleep time spent in REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final percentage values from the agent's report:\n1) sleep efficiency (SE)\n2) the percentage of total sleep time spent in N3 deep sleep\n3) the percentage of total sleep time spent in REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"sleep_efficiency_percent\", \"n3_sleep_ratio_percent\", and \"rem_sleep_ratio_percent\".\n5. All three values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"sleep_efficiency_percent\": , \"n3_sleep_ratio_percent\": , \"rem_sleep_ratio_percent\": }", + "metrics": [ + { + "metric_id": "sleep_efficiency_accuracy", + "type": "numeric_check", + "target_key": "sleep_efficiency_percent", + "weight": 40, + "params": { + "gt_value": 94.55497382198953, + "tolerance": 5.0 + } + }, + { + "metric_id": "n3_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "n3_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 13.732004429678849, + "tolerance": 4.0 + } + }, + { + "metric_id": "rem_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "rem_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 29.678848283499445, + "tolerance": 6.0 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case24/case24_19.json b/sleep_assessment/cases/case24/case24_19.json new file mode 100644 index 0000000000000000000000000000000000000000..b00bc66ecda782c2b63ded49c48c6ef34640614a --- /dev/null +++ b/sleep_assessment/cases/case24/case24_19.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "MASSSS3_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_04.edf", + "instruction": "Please use the provided overnight sleep file to perform automatic sleep staging for the entire recording. Based on the resulting sleep-stage sequence, report three final percentage values: sleep efficiency (SE), the percentage of total sleep time spent in N3 deep sleep, and the percentage of total sleep time spent in REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final percentage values from the agent's report:\n1) sleep efficiency (SE)\n2) the percentage of total sleep time spent in N3 deep sleep\n3) the percentage of total sleep time spent in REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"sleep_efficiency_percent\", \"n3_sleep_ratio_percent\", and \"rem_sleep_ratio_percent\".\n5. All three values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"sleep_efficiency_percent\": , \"n3_sleep_ratio_percent\": , \"rem_sleep_ratio_percent\": }", + "metrics": [ + { + "metric_id": "sleep_efficiency_accuracy", + "type": "numeric_check", + "target_key": "sleep_efficiency_percent", + "weight": 40, + "params": { + "gt_value": 98.79385964912281, + "tolerance": 5.0 + } + }, + { + "metric_id": "n3_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "n3_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 7.103218645948946, + "tolerance": 4.0 + } + }, + { + "metric_id": "rem_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "rem_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 27.19200887902331, + "tolerance": 6.0 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case24/case24_20.json b/sleep_assessment/cases/case24/case24_20.json new file mode 100644 index 0000000000000000000000000000000000000000..00cb79be3ba90e596944b24bb8069bfdbd00e069 --- /dev/null +++ b/sleep_assessment/cases/case24/case24_20.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "MASSSS3_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_05.edf", + "instruction": "Please use the provided overnight sleep file to perform automatic sleep staging for the entire recording. Based on the resulting sleep-stage sequence, report three final percentage values: sleep efficiency (SE), the percentage of total sleep time spent in N3 deep sleep, and the percentage of total sleep time spent in REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final percentage values from the agent's report:\n1) sleep efficiency (SE)\n2) the percentage of total sleep time spent in N3 deep sleep\n3) the percentage of total sleep time spent in REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"sleep_efficiency_percent\", \"n3_sleep_ratio_percent\", and \"rem_sleep_ratio_percent\".\n5. All three values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"sleep_efficiency_percent\": , \"n3_sleep_ratio_percent\": , \"rem_sleep_ratio_percent\": }", + "metrics": [ + { + "metric_id": "sleep_efficiency_accuracy", + "type": "numeric_check", + "target_key": "sleep_efficiency_percent", + "weight": 40, + "params": { + "gt_value": 95.0420954162769, + "tolerance": 5.0 + } + }, + { + "metric_id": "n3_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "n3_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 19.19291338582677, + "tolerance": 5.0 + } + }, + { + "metric_id": "rem_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "rem_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 27.263779527559056, + "tolerance": 6.0 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case24/case24_21.json b/sleep_assessment/cases/case24/case24_21.json new file mode 100644 index 0000000000000000000000000000000000000000..cb889e9bba4c7471cf02cdf4b97e2728fd28e39d --- /dev/null +++ b/sleep_assessment/cases/case24/case24_21.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "instruction": "Please use the provided overnight sleep file to perform automatic sleep staging for the entire recording. Based on the resulting sleep-stage sequence, report three final percentage values: sleep efficiency (SE), the percentage of total sleep time spent in N3 deep sleep, and the percentage of total sleep time spent in REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final percentage values from the agent's report:\n1) sleep efficiency (SE)\n2) the percentage of total sleep time spent in N3 deep sleep\n3) the percentage of total sleep time spent in REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"sleep_efficiency_percent\", \"n3_sleep_ratio_percent\", and \"rem_sleep_ratio_percent\".\n5. All three values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"sleep_efficiency_percent\": , \"n3_sleep_ratio_percent\": , \"rem_sleep_ratio_percent\": }", + "metrics": [ + { + "metric_id": "sleep_efficiency_accuracy", + "type": "numeric_check", + "target_key": "sleep_efficiency_percent", + "weight": 40, + "params": { + "gt_value": 56.56804733727811, + "tolerance": 8.0 + } + }, + { + "metric_id": "n3_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "n3_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 15.06276150627615, + "tolerance": 5.0 + } + }, + { + "metric_id": "rem_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "rem_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 16.94560669456067, + "tolerance": 5.0 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case24/case24_22.json b/sleep_assessment/cases/case24/case24_22.json new file mode 100644 index 0000000000000000000000000000000000000000..30f1af3359ff5034c19cf14ad1091b25c51e6ca2 --- /dev/null +++ b/sleep_assessment/cases/case24/case24_22.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "instruction": "Please use the provided overnight sleep file to perform automatic sleep staging for the entire recording. Based on the resulting sleep-stage sequence, report three final percentage values: sleep efficiency (SE), the percentage of total sleep time spent in N3 deep sleep, and the percentage of total sleep time spent in REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final percentage values from the agent's report:\n1) sleep efficiency (SE)\n2) the percentage of total sleep time spent in N3 deep sleep\n3) the percentage of total sleep time spent in REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"sleep_efficiency_percent\", \"n3_sleep_ratio_percent\", and \"rem_sleep_ratio_percent\".\n5. All three values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"sleep_efficiency_percent\": , \"n3_sleep_ratio_percent\": , \"rem_sleep_ratio_percent\": }", + "metrics": [ + { + "metric_id": "sleep_efficiency_accuracy", + "type": "numeric_check", + "target_key": "sleep_efficiency_percent", + "weight": 40, + "params": { + "gt_value": 59.2020592020592, + "tolerance": 8.0 + } + }, + { + "metric_id": "n3_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "n3_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 4.565217391304348, + "tolerance": 4.0 + } + }, + { + "metric_id": "rem_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "rem_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 0.21739130434782608, + "tolerance": 4.0 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case24/case24_23.json b/sleep_assessment/cases/case24/case24_23.json new file mode 100644 index 0000000000000000000000000000000000000000..8828fb52d954049fc2f4f401f0a0f763bcf281f1 --- /dev/null +++ b/sleep_assessment/cases/case24/case24_23.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "Physionet2018_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_03.edf", + "instruction": "Please use the provided overnight sleep file to perform automatic sleep staging for the entire recording. Based on the resulting sleep-stage sequence, report three final percentage values: sleep efficiency (SE), the percentage of total sleep time spent in N3 deep sleep, and the percentage of total sleep time spent in REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final percentage values from the agent's report:\n1) sleep efficiency (SE)\n2) the percentage of total sleep time spent in N3 deep sleep\n3) the percentage of total sleep time spent in REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"sleep_efficiency_percent\", \"n3_sleep_ratio_percent\", and \"rem_sleep_ratio_percent\".\n5. All three values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"sleep_efficiency_percent\": , \"n3_sleep_ratio_percent\": , \"rem_sleep_ratio_percent\": }", + "metrics": [ + { + "metric_id": "sleep_efficiency_accuracy", + "type": "numeric_check", + "target_key": "sleep_efficiency_percent", + "weight": 40, + "params": { + "gt_value": 46.21676891615542, + "tolerance": 10.0 + } + }, + { + "metric_id": "n3_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "n3_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 13.716814159292035, + "tolerance": 4.0 + } + }, + { + "metric_id": "rem_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "rem_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 0.22123893805309736, + "tolerance": 4.0 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case24/case24_24.json b/sleep_assessment/cases/case24/case24_24.json new file mode 100644 index 0000000000000000000000000000000000000000..03198a7dd7dd12bb3a693524be42fe3303ee4f4b --- /dev/null +++ b/sleep_assessment/cases/case24/case24_24.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "instruction": "Please use the provided overnight sleep file to perform automatic sleep staging for the entire recording. Based on the resulting sleep-stage sequence, report three final percentage values: sleep efficiency (SE), the percentage of total sleep time spent in N3 deep sleep, and the percentage of total sleep time spent in REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final percentage values from the agent's report:\n1) sleep efficiency (SE)\n2) the percentage of total sleep time spent in N3 deep sleep\n3) the percentage of total sleep time spent in REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"sleep_efficiency_percent\", \"n3_sleep_ratio_percent\", and \"rem_sleep_ratio_percent\".\n5. All three values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"sleep_efficiency_percent\": , \"n3_sleep_ratio_percent\": , \"rem_sleep_ratio_percent\": }", + "metrics": [ + { + "metric_id": "sleep_efficiency_accuracy", + "type": "numeric_check", + "target_key": "sleep_efficiency_percent", + "weight": 40, + "params": { + "gt_value": 90.40284360189574, + "tolerance": 5.0 + } + }, + { + "metric_id": "n3_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "n3_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 22.542595019659238, + "tolerance": 5.0 + } + }, + { + "metric_id": "rem_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "rem_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 11.795543905635649, + "tolerance": 5.0 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case24/case24_25.json b/sleep_assessment/cases/case24/case24_25.json new file mode 100644 index 0000000000000000000000000000000000000000..439377677242d20383d5c3f4f3b6c535c3c0d959 --- /dev/null +++ b/sleep_assessment/cases/case24/case24_25.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "instruction": "Please use the provided overnight sleep file to perform automatic sleep staging for the entire recording. Based on the resulting sleep-stage sequence, report three final percentage values: sleep efficiency (SE), the percentage of total sleep time spent in N3 deep sleep, and the percentage of total sleep time spent in REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final percentage values from the agent's report:\n1) sleep efficiency (SE)\n2) the percentage of total sleep time spent in N3 deep sleep\n3) the percentage of total sleep time spent in REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The keys must be EXACTLY \"sleep_efficiency_percent\", \"n3_sleep_ratio_percent\", and \"rem_sleep_ratio_percent\".\n5. All three values must be FLOAT percentages or null. Do not include percent signs.\n6. If the report gives a fraction or proportion between 0 and 1, convert it to a percentage by multiplying by 100.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"sleep_efficiency_percent\": , \"n3_sleep_ratio_percent\": , \"rem_sleep_ratio_percent\": }", + "metrics": [ + { + "metric_id": "sleep_efficiency_accuracy", + "type": "numeric_check", + "target_key": "sleep_efficiency_percent", + "weight": 40, + "params": { + "gt_value": 93.84775808133472, + "tolerance": 5.0 + } + }, + { + "metric_id": "n3_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "n3_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 11.0, + "tolerance": 4.0 + } + }, + { + "metric_id": "rem_sleep_ratio_accuracy", + "type": "numeric_check", + "target_key": "rem_sleep_ratio_percent", + "weight": 30, + "params": { + "gt_value": 11.333333333333334, + "tolerance": 5.0 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case25/case25_01.json b/sleep_assessment/cases/case25/case25_01.json new file mode 100644 index 0000000000000000000000000000000000000000..b0a47a0b858a42971312388de138097dcc5cc767 --- /dev/null +++ b/sleep_assessment/cases/case25/case25_01.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "instruction": "Please analyze the provided sleep file and examine the EEG segment from minute 181.0 to minute 181.5. Evaluate the claim that the segment's dominant EEG frequency band is alpha and that its sleep stage is R. Report whether the claim is supported, and provide the dominant frequency band and sleep stage you determine for this segment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to preserve the agent's answer for semantic judging.\n\n### TASK\nExtract the full substantive text of the agent's report, including its validity judgement, identified errors, and corrected answer if present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The key must be EXACTLY \"full_text\".\n5. If the report is empty or no substantive answer is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "false_conclusion_verification_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_01.edf. The agent was asked to verify this conclusion: in the EEG segment from minute 181.0 to minute 181.5, the dominant EEG frequency band is alpha, and the sleep stage is R. Ground truth: the true dominant EEG frequency band is theta, and the true sleep stage is R. The provided conclusion is invalid. The incorrect component(s) are: dominant_band. Score the response in two parts. Give 0.5 points if the response clearly recognizes that the provided conclusion is false or partially false and identifies which component is wrong. Give 0.5 points if the response provides the corrected answer, meaning the true dominant band and true sleep stage, or correctly fixes each wrong component while preserving any component that was already correct. If the response agrees with the provided false conclusion or says the conclusion is valid without finding any error, the score must be 0.0. The total score must therefore be exactly one of 0.0, 0.5, or 1.0. Output only one valid JSON object: {\"score\": 0.0}, {\"score\": 0.5}, or {\"score\": 1.0}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case25/case25_02.json b/sleep_assessment/cases/case25/case25_02.json new file mode 100644 index 0000000000000000000000000000000000000000..be2591c4e578122bd301e431a965c0ccd5ae4f5e --- /dev/null +++ b/sleep_assessment/cases/case25/case25_02.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "instruction": "Please analyze the provided sleep file and examine the EEG segment from minute 65.0 to minute 65.5. Evaluate the claim that the segment's dominant EEG frequency band is delta and that its sleep stage is W. Report whether the claim is supported, and provide the dominant frequency band and sleep stage you determine for this segment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to preserve the agent's answer for semantic judging.\n\n### TASK\nExtract the full substantive text of the agent's report, including its validity judgement, identified errors, and corrected answer if present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The key must be EXACTLY \"full_text\".\n5. If the report is empty or no substantive answer is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "false_conclusion_verification_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_02.edf. The agent was asked to verify this conclusion: in the EEG segment from minute 65.0 to minute 65.5, the dominant EEG frequency band is delta, and the sleep stage is W. Ground truth: the true dominant EEG frequency band is alpha, and the true sleep stage is W. The provided conclusion is invalid. The incorrect component(s) are: dominant_band. Score the response in two parts. Give 0.5 points if the response clearly recognizes that the provided conclusion is false or partially false and identifies which component is wrong. Give 0.5 points if the response provides the corrected answer, meaning the true dominant band and true sleep stage, or correctly fixes each wrong component while preserving any component that was already correct. If the response agrees with the provided false conclusion or says the conclusion is valid without finding any error, the score must be 0.0. The total score must therefore be exactly one of 0.0, 0.5, or 1.0. Output only one valid JSON object: {\"score\": 0.0}, {\"score\": 0.5}, or {\"score\": 1.0}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case25/case25_03.json b/sleep_assessment/cases/case25/case25_03.json new file mode 100644 index 0000000000000000000000000000000000000000..7289f2e8c60cec009b971ca07131349cea95f060 --- /dev/null +++ b/sleep_assessment/cases/case25/case25_03.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "instruction": "Please analyze the provided sleep file and examine the EEG segment from minute 251.0 to minute 251.5. Evaluate the claim that the segment's dominant EEG frequency band is alpha and that its sleep stage is N3. Report whether the claim is supported, and provide the dominant frequency band and sleep stage you determine for this segment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to preserve the agent's answer for semantic judging.\n\n### TASK\nExtract the full substantive text of the agent's report, including its validity judgement, identified errors, and corrected answer if present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The key must be EXACTLY \"full_text\".\n5. If the report is empty or no substantive answer is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "false_conclusion_verification_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_03.edf. The agent was asked to verify this conclusion: in the EEG segment from minute 251.0 to minute 251.5, the dominant EEG frequency band is alpha, and the sleep stage is N3. Ground truth: the true dominant EEG frequency band is alpha, and the true sleep stage is W. The provided conclusion is invalid. The incorrect component(s) are: sleep_stage. Score the response in two parts. Give 0.5 points if the response clearly recognizes that the provided conclusion is false or partially false and identifies which component is wrong. Give 0.5 points if the response provides the corrected answer, meaning the true dominant band and true sleep stage, or correctly fixes each wrong component while preserving any component that was already correct. If the response agrees with the provided false conclusion or says the conclusion is valid without finding any error, the score must be 0.0. The total score must therefore be exactly one of 0.0, 0.5, or 1.0. Output only one valid JSON object: {\"score\": 0.0}, {\"score\": 0.5}, or {\"score\": 1.0}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case25/case25_04.json b/sleep_assessment/cases/case25/case25_04.json new file mode 100644 index 0000000000000000000000000000000000000000..16b2fb09e0b26491e753104c681796828dbf03ed --- /dev/null +++ b/sleep_assessment/cases/case25/case25_04.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "ISRUC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_04.edf", + "instruction": "Please analyze the provided sleep file and examine the EEG segment from minute 147.5 to minute 148.5. Evaluate the claim that the segment's dominant EEG frequency band is theta and that its sleep stage is N3. Report whether the claim is supported, and provide the dominant frequency band and sleep stage you determine for this segment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to preserve the agent's answer for semantic judging.\n\n### TASK\nExtract the full substantive text of the agent's report, including its validity judgement, identified errors, and corrected answer if present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The key must be EXACTLY \"full_text\".\n5. If the report is empty or no substantive answer is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "false_conclusion_verification_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_04.edf. The agent was asked to verify this conclusion: in the EEG segment from minute 147.5 to minute 148.5, the dominant EEG frequency band is theta, and the sleep stage is N3. Ground truth: the true dominant EEG frequency band is delta, and the true sleep stage is N3. The provided conclusion is invalid. The incorrect component(s) are: dominant_band. Score the response in two parts. Give 0.5 points if the response clearly recognizes that the provided conclusion is false or partially false and identifies which component is wrong. Give 0.5 points if the response provides the corrected answer, meaning the true dominant band and true sleep stage, or correctly fixes each wrong component while preserving any component that was already correct. If the response agrees with the provided false conclusion or says the conclusion is valid without finding any error, the score must be 0.0. The total score must therefore be exactly one of 0.0, 0.5, or 1.0. Output only one valid JSON object: {\"score\": 0.0}, {\"score\": 0.5}, or {\"score\": 1.0}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case25/case25_05.json b/sleep_assessment/cases/case25/case25_05.json new file mode 100644 index 0000000000000000000000000000000000000000..5fff28064442024a4bb252e452d973f3ecddf584 --- /dev/null +++ b/sleep_assessment/cases/case25/case25_05.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "instruction": "Please analyze the provided sleep file and examine the EEG segment from minute 415.5 to minute 416.0. Evaluate the claim that the segment's dominant EEG frequency band is beta and that its sleep stage is N3. Report whether the claim is supported, and provide the dominant frequency band and sleep stage you determine for this segment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to preserve the agent's answer for semantic judging.\n\n### TASK\nExtract the full substantive text of the agent's report, including its validity judgement, identified errors, and corrected answer if present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The key must be EXACTLY \"full_text\".\n5. If the report is empty or no substantive answer is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "false_conclusion_verification_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_05.edf. The agent was asked to verify this conclusion: in the EEG segment from minute 415.5 to minute 416.0, the dominant EEG frequency band is beta, and the sleep stage is N3. Ground truth: the true dominant EEG frequency band is beta, and the true sleep stage is W. The provided conclusion is invalid. The incorrect component(s) are: sleep_stage. Score the response in two parts. Give 0.5 points if the response clearly recognizes that the provided conclusion is false or partially false and identifies which component is wrong. Give 0.5 points if the response provides the corrected answer, meaning the true dominant band and true sleep stage, or correctly fixes each wrong component while preserving any component that was already correct. If the response agrees with the provided false conclusion or says the conclusion is valid without finding any error, the score must be 0.0. The total score must therefore be exactly one of 0.0, 0.5, or 1.0. Output only one valid JSON object: {\"score\": 0.0}, {\"score\": 0.5}, or {\"score\": 1.0}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case25/case25_06.json b/sleep_assessment/cases/case25/case25_06.json new file mode 100644 index 0000000000000000000000000000000000000000..daeaba2c994c32457275c05314bf52c5fd8f4af9 --- /dev/null +++ b/sleep_assessment/cases/case25/case25_06.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "HMC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_01.edf", + "instruction": "Please analyze the provided sleep file and examine the EEG segment from minute 233.5 to minute 234.5. Evaluate the claim that the segment's dominant EEG frequency band is delta and that its sleep stage is W. Report whether the claim is supported, and provide the dominant frequency band and sleep stage you determine for this segment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to preserve the agent's answer for semantic judging.\n\n### TASK\nExtract the full substantive text of the agent's report, including its validity judgement, identified errors, and corrected answer if present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The key must be EXACTLY \"full_text\".\n5. If the report is empty or no substantive answer is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "false_conclusion_verification_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is HMC_01.edf. The agent was asked to verify this conclusion: in the EEG segment from minute 233.5 to minute 234.5, the dominant EEG frequency band is delta, and the sleep stage is W. Ground truth: the true dominant EEG frequency band is delta, and the true sleep stage is N3. The provided conclusion is invalid. The incorrect component(s) are: sleep_stage. Score the response in two parts. Give 0.5 points if the response clearly recognizes that the provided conclusion is false or partially false and identifies which component is wrong. Give 0.5 points if the response provides the corrected answer, meaning the true dominant band and true sleep stage, or correctly fixes each wrong component while preserving any component that was already correct. If the response agrees with the provided false conclusion or says the conclusion is valid without finding any error, the score must be 0.0. The total score must therefore be exactly one of 0.0, 0.5, or 1.0. Output only one valid JSON object: {\"score\": 0.0}, {\"score\": 0.5}, or {\"score\": 1.0}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case25/case25_07.json b/sleep_assessment/cases/case25/case25_07.json new file mode 100644 index 0000000000000000000000000000000000000000..106d03e97994b17261182e57258b848a09516c72 --- /dev/null +++ b/sleep_assessment/cases/case25/case25_07.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "HMC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_02.edf", + "instruction": "Please analyze the provided sleep file and examine the EEG segment from minute 161.0 to minute 162.0. Evaluate the claim that the segment's dominant EEG frequency band is alpha and that its sleep stage is N3. Report whether the claim is supported, and provide the dominant frequency band and sleep stage you determine for this segment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to preserve the agent's answer for semantic judging.\n\n### TASK\nExtract the full substantive text of the agent's report, including its validity judgement, identified errors, and corrected answer if present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The key must be EXACTLY \"full_text\".\n5. If the report is empty or no substantive answer is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "false_conclusion_verification_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is HMC_02.edf. The agent was asked to verify this conclusion: in the EEG segment from minute 161.0 to minute 162.0, the dominant EEG frequency band is alpha, and the sleep stage is N3. Ground truth: the true dominant EEG frequency band is delta, and the true sleep stage is N3. The provided conclusion is invalid. The incorrect component(s) are: dominant_band. Score the response in two parts. Give 0.5 points if the response clearly recognizes that the provided conclusion is false or partially false and identifies which component is wrong. Give 0.5 points if the response provides the corrected answer, meaning the true dominant band and true sleep stage, or correctly fixes each wrong component while preserving any component that was already correct. If the response agrees with the provided false conclusion or says the conclusion is valid without finding any error, the score must be 0.0. The total score must therefore be exactly one of 0.0, 0.5, or 1.0. Output only one valid JSON object: {\"score\": 0.0}, {\"score\": 0.5}, or {\"score\": 1.0}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case25/case25_08.json b/sleep_assessment/cases/case25/case25_08.json new file mode 100644 index 0000000000000000000000000000000000000000..19bee5bc134a7d5f391fe1a999bd095160c887e1 --- /dev/null +++ b/sleep_assessment/cases/case25/case25_08.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "HMC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_03.edf", + "instruction": "Please analyze the provided sleep file and examine the EEG segment from minute 235.5 to minute 237.0. Evaluate the claim that the segment's dominant EEG frequency band is theta and that its sleep stage is R. Report whether the claim is supported, and provide the dominant frequency band and sleep stage you determine for this segment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to preserve the agent's answer for semantic judging.\n\n### TASK\nExtract the full substantive text of the agent's report, including its validity judgement, identified errors, and corrected answer if present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The key must be EXACTLY \"full_text\".\n5. If the report is empty or no substantive answer is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "false_conclusion_verification_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is HMC_03.edf. The agent was asked to verify this conclusion: in the EEG segment from minute 235.5 to minute 237.0, the dominant EEG frequency band is theta, and the sleep stage is R. Ground truth: the true dominant EEG frequency band is delta, and the true sleep stage is N3. The provided conclusion is invalid. The incorrect component(s) are: dominant_band, sleep_stage. Score the response in two parts. Give 0.5 points if the response clearly recognizes that the provided conclusion is false or partially false and identifies which component is wrong. Give 0.5 points if the response provides the corrected answer, meaning the true dominant band and true sleep stage, or correctly fixes each wrong component while preserving any component that was already correct. If the response agrees with the provided false conclusion or says the conclusion is valid without finding any error, the score must be 0.0. The total score must therefore be exactly one of 0.0, 0.5, or 1.0. Output only one valid JSON object: {\"score\": 0.0}, {\"score\": 0.5}, or {\"score\": 1.0}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case25/case25_09.json b/sleep_assessment/cases/case25/case25_09.json new file mode 100644 index 0000000000000000000000000000000000000000..b9d955e530d1c0aeddd4b050545cee40c2c8c5e8 --- /dev/null +++ b/sleep_assessment/cases/case25/case25_09.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "HMC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_04.edf", + "instruction": "Please analyze the provided sleep file and examine the EEG segment from minute 504.5 to minute 505.0. Evaluate the claim that the segment's dominant EEG frequency band is delta and that its sleep stage is W. Report whether the claim is supported, and provide the dominant frequency band and sleep stage you determine for this segment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to preserve the agent's answer for semantic judging.\n\n### TASK\nExtract the full substantive text of the agent's report, including its validity judgement, identified errors, and corrected answer if present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The key must be EXACTLY \"full_text\".\n5. If the report is empty or no substantive answer is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "false_conclusion_verification_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is HMC_04.edf. The agent was asked to verify this conclusion: in the EEG segment from minute 504.5 to minute 505.0, the dominant EEG frequency band is delta, and the sleep stage is W. Ground truth: the true dominant EEG frequency band is beta, and the true sleep stage is W. The provided conclusion is invalid. The incorrect component(s) are: dominant_band. Score the response in two parts. Give 0.5 points if the response clearly recognizes that the provided conclusion is false or partially false and identifies which component is wrong. Give 0.5 points if the response provides the corrected answer, meaning the true dominant band and true sleep stage, or correctly fixes each wrong component while preserving any component that was already correct. If the response agrees with the provided false conclusion or says the conclusion is valid without finding any error, the score must be 0.0. The total score must therefore be exactly one of 0.0, 0.5, or 1.0. Output only one valid JSON object: {\"score\": 0.0}, {\"score\": 0.5}, or {\"score\": 1.0}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case25/case25_10.json b/sleep_assessment/cases/case25/case25_10.json new file mode 100644 index 0000000000000000000000000000000000000000..4411b2b2b8640f6242f5ca1efe8a7d1dcdcd2c0a --- /dev/null +++ b/sleep_assessment/cases/case25/case25_10.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "HMC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_05.edf", + "instruction": "Please analyze the provided sleep file and examine the EEG segment from minute 144.0 to minute 144.5. Evaluate the claim that the segment's dominant EEG frequency band is theta and that its sleep stage is W. Report whether the claim is supported, and provide the dominant frequency band and sleep stage you determine for this segment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to preserve the agent's answer for semantic judging.\n\n### TASK\nExtract the full substantive text of the agent's report, including its validity judgement, identified errors, and corrected answer if present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The key must be EXACTLY \"full_text\".\n5. If the report is empty or no substantive answer is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "false_conclusion_verification_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is HMC_05.edf. The agent was asked to verify this conclusion: in the EEG segment from minute 144.0 to minute 144.5, the dominant EEG frequency band is theta, and the sleep stage is W. Ground truth: the true dominant EEG frequency band is theta, and the true sleep stage is N1. The provided conclusion is invalid. The incorrect component(s) are: sleep_stage. Score the response in two parts. Give 0.5 points if the response clearly recognizes that the provided conclusion is false or partially false and identifies which component is wrong. Give 0.5 points if the response provides the corrected answer, meaning the true dominant band and true sleep stage, or correctly fixes each wrong component while preserving any component that was already correct. If the response agrees with the provided false conclusion or says the conclusion is valid without finding any error, the score must be 0.0. The total score must therefore be exactly one of 0.0, 0.5, or 1.0. Output only one valid JSON object: {\"score\": 0.0}, {\"score\": 0.5}, or {\"score\": 1.0}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case25/case25_11.json b/sleep_assessment/cases/case25/case25_11.json new file mode 100644 index 0000000000000000000000000000000000000000..1397433003fd8f77b2b40dfcd30e775410580278 --- /dev/null +++ b/sleep_assessment/cases/case25/case25_11.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "instruction": "Please analyze the provided sleep file and examine the EEG segment from minute 6.0 to minute 6.5. Evaluate the claim that the segment's dominant EEG frequency band is alpha and that its sleep stage is W. Report whether the claim is supported, and provide the dominant frequency band and sleep stage you determine for this segment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to preserve the agent's answer for semantic judging.\n\n### TASK\nExtract the full substantive text of the agent's report, including its validity judgement, identified errors, and corrected answer if present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The key must be EXACTLY \"full_text\".\n5. If the report is empty or no substantive answer is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "false_conclusion_verification_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_01.edf. The agent was asked to verify this conclusion: in the EEG segment from minute 6.0 to minute 6.5, the dominant EEG frequency band is alpha, and the sleep stage is W. Ground truth: the true dominant EEG frequency band is beta, and the true sleep stage is W. The provided conclusion is invalid. The incorrect component(s) are: dominant_band. Score the response in two parts. Give 0.5 points if the response clearly recognizes that the provided conclusion is false or partially false and identifies which component is wrong. Give 0.5 points if the response provides the corrected answer, meaning the true dominant band and true sleep stage, or correctly fixes each wrong component while preserving any component that was already correct. If the response agrees with the provided false conclusion or says the conclusion is valid without finding any error, the score must be 0.0. The total score must therefore be exactly one of 0.0, 0.5, or 1.0. Output only one valid JSON object: {\"score\": 0.0}, {\"score\": 0.5}, or {\"score\": 1.0}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case25/case25_12.json b/sleep_assessment/cases/case25/case25_12.json new file mode 100644 index 0000000000000000000000000000000000000000..1eea095d7a9d55faa4e804efb802cb17641579e6 --- /dev/null +++ b/sleep_assessment/cases/case25/case25_12.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "instruction": "Please analyze the provided sleep file and examine the EEG segment from minute 87.5 to minute 88.0. Evaluate the claim that the segment's dominant EEG frequency band is alpha and that its sleep stage is R. Report whether the claim is supported, and provide the dominant frequency band and sleep stage you determine for this segment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to preserve the agent's answer for semantic judging.\n\n### TASK\nExtract the full substantive text of the agent's report, including its validity judgement, identified errors, and corrected answer if present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The key must be EXACTLY \"full_text\".\n5. If the report is empty or no substantive answer is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "false_conclusion_verification_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_02.edf. The agent was asked to verify this conclusion: in the EEG segment from minute 87.5 to minute 88.0, the dominant EEG frequency band is alpha, and the sleep stage is R. Ground truth: the true dominant EEG frequency band is alpha, and the true sleep stage is W. The provided conclusion is invalid. The incorrect component(s) are: sleep_stage. Score the response in two parts. Give 0.5 points if the response clearly recognizes that the provided conclusion is false or partially false and identifies which component is wrong. Give 0.5 points if the response provides the corrected answer, meaning the true dominant band and true sleep stage, or correctly fixes each wrong component while preserving any component that was already correct. If the response agrees with the provided false conclusion or says the conclusion is valid without finding any error, the score must be 0.0. The total score must therefore be exactly one of 0.0, 0.5, or 1.0. Output only one valid JSON object: {\"score\": 0.0}, {\"score\": 0.5}, or {\"score\": 1.0}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case25/case25_13.json b/sleep_assessment/cases/case25/case25_13.json new file mode 100644 index 0000000000000000000000000000000000000000..4269a0b63b4c2c3cc85b75ea29c4fc6fde49ee8f --- /dev/null +++ b/sleep_assessment/cases/case25/case25_13.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "instruction": "Please analyze the provided sleep file and examine the EEG segment from minute 439.0 to minute 439.5. Evaluate the claim that the segment's dominant EEG frequency band is theta and that its sleep stage is N1. Report whether the claim is supported, and provide the dominant frequency band and sleep stage you determine for this segment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to preserve the agent's answer for semantic judging.\n\n### TASK\nExtract the full substantive text of the agent's report, including its validity judgement, identified errors, and corrected answer if present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The key must be EXACTLY \"full_text\".\n5. If the report is empty or no substantive answer is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "false_conclusion_verification_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_03.edf. The agent was asked to verify this conclusion: in the EEG segment from minute 439.0 to minute 439.5, the dominant EEG frequency band is theta, and the sleep stage is N1. Ground truth: the true dominant EEG frequency band is alpha, and the true sleep stage is N1. The provided conclusion is invalid. The incorrect component(s) are: dominant_band. Score the response in two parts. Give 0.5 points if the response clearly recognizes that the provided conclusion is false or partially false and identifies which component is wrong. Give 0.5 points if the response provides the corrected answer, meaning the true dominant band and true sleep stage, or correctly fixes each wrong component while preserving any component that was already correct. If the response agrees with the provided false conclusion or says the conclusion is valid without finding any error, the score must be 0.0. The total score must therefore be exactly one of 0.0, 0.5, or 1.0. Output only one valid JSON object: {\"score\": 0.0}, {\"score\": 0.5}, or {\"score\": 1.0}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case25/case25_14.json b/sleep_assessment/cases/case25/case25_14.json new file mode 100644 index 0000000000000000000000000000000000000000..729ffa4161532b402716513927ad4e5d7bf574f6 --- /dev/null +++ b/sleep_assessment/cases/case25/case25_14.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "instruction": "Please analyze the provided sleep file and examine the EEG segment from minute 150.5 to minute 151.0. Evaluate the claim that the segment's dominant EEG frequency band is theta and that its sleep stage is W. Report whether the claim is supported, and provide the dominant frequency band and sleep stage you determine for this segment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to preserve the agent's answer for semantic judging.\n\n### TASK\nExtract the full substantive text of the agent's report, including its validity judgement, identified errors, and corrected answer if present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The key must be EXACTLY \"full_text\".\n5. If the report is empty or no substantive answer is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "false_conclusion_verification_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_04.edf. The agent was asked to verify this conclusion: in the EEG segment from minute 150.5 to minute 151.0, the dominant EEG frequency band is theta, and the sleep stage is W. Ground truth: the true dominant EEG frequency band is theta, and the true sleep stage is R. The provided conclusion is invalid. The incorrect component(s) are: sleep_stage. Score the response in two parts. Give 0.5 points if the response clearly recognizes that the provided conclusion is false or partially false and identifies which component is wrong. Give 0.5 points if the response provides the corrected answer, meaning the true dominant band and true sleep stage, or correctly fixes each wrong component while preserving any component that was already correct. If the response agrees with the provided false conclusion or says the conclusion is valid without finding any error, the score must be 0.0. The total score must therefore be exactly one of 0.0, 0.5, or 1.0. Output only one valid JSON object: {\"score\": 0.0}, {\"score\": 0.5}, or {\"score\": 1.0}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case25/case25_15.json b/sleep_assessment/cases/case25/case25_15.json new file mode 100644 index 0000000000000000000000000000000000000000..35a7a9fe635ede7753a3b1dc014100170c0b0e22 --- /dev/null +++ b/sleep_assessment/cases/case25/case25_15.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "instruction": "Please analyze the provided sleep file and examine the EEG segment from minute 54.0 to minute 55.0. Evaluate the claim that the segment's dominant EEG frequency band is beta and that its sleep stage is N2. Report whether the claim is supported, and provide the dominant frequency band and sleep stage you determine for this segment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to preserve the agent's answer for semantic judging.\n\n### TASK\nExtract the full substantive text of the agent's report, including its validity judgement, identified errors, and corrected answer if present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The key must be EXACTLY \"full_text\".\n5. If the report is empty or no substantive answer is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "false_conclusion_verification_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_05.edf. The agent was asked to verify this conclusion: in the EEG segment from minute 54.0 to minute 55.0, the dominant EEG frequency band is beta, and the sleep stage is N2. Ground truth: the true dominant EEG frequency band is beta, and the true sleep stage is W. The provided conclusion is invalid. The incorrect component(s) are: sleep_stage. Score the response in two parts. Give 0.5 points if the response clearly recognizes that the provided conclusion is false or partially false and identifies which component is wrong. Give 0.5 points if the response provides the corrected answer, meaning the true dominant band and true sleep stage, or correctly fixes each wrong component while preserving any component that was already correct. If the response agrees with the provided false conclusion or says the conclusion is valid without finding any error, the score must be 0.0. The total score must therefore be exactly one of 0.0, 0.5, or 1.0. Output only one valid JSON object: {\"score\": 0.0}, {\"score\": 0.5}, or {\"score\": 1.0}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case25/case25_16.json b/sleep_assessment/cases/case25/case25_16.json new file mode 100644 index 0000000000000000000000000000000000000000..2a9b7e764e9cce1f948713e38e1010f873073a1b --- /dev/null +++ b/sleep_assessment/cases/case25/case25_16.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "MASSSS3_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_01.edf", + "instruction": "Please analyze the provided sleep file and examine the EEG segment from minute 337.5 to minute 338.5. Evaluate the claim that the segment's dominant EEG frequency band is delta and that its sleep stage is W. Report whether the claim is supported, and provide the dominant frequency band and sleep stage you determine for this segment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to preserve the agent's answer for semantic judging.\n\n### TASK\nExtract the full substantive text of the agent's report, including its validity judgement, identified errors, and corrected answer if present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The key must be EXACTLY \"full_text\".\n5. If the report is empty or no substantive answer is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "false_conclusion_verification_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is MASSSS3_01.edf. The agent was asked to verify this conclusion: in the EEG segment from minute 337.5 to minute 338.5, the dominant EEG frequency band is delta, and the sleep stage is W. Ground truth: the true dominant EEG frequency band is delta, and the true sleep stage is N3. The provided conclusion is invalid. The incorrect component(s) are: sleep_stage. Score the response in two parts. Give 0.5 points if the response clearly recognizes that the provided conclusion is false or partially false and identifies which component is wrong. Give 0.5 points if the response provides the corrected answer, meaning the true dominant band and true sleep stage, or correctly fixes each wrong component while preserving any component that was already correct. If the response agrees with the provided false conclusion or says the conclusion is valid without finding any error, the score must be 0.0. The total score must therefore be exactly one of 0.0, 0.5, or 1.0. Output only one valid JSON object: {\"score\": 0.0}, {\"score\": 0.5}, or {\"score\": 1.0}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case25/case25_17.json b/sleep_assessment/cases/case25/case25_17.json new file mode 100644 index 0000000000000000000000000000000000000000..f17d5ed24e08dac37251c1e75f020ef975e5e488 --- /dev/null +++ b/sleep_assessment/cases/case25/case25_17.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "MASSSS3_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_02.edf", + "instruction": "Please analyze the provided sleep file and examine the EEG segment from minute 517.5 to minute 518.0. Evaluate the claim that the segment's dominant EEG frequency band is beta and that its sleep stage is N1. Report whether the claim is supported, and provide the dominant frequency band and sleep stage you determine for this segment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to preserve the agent's answer for semantic judging.\n\n### TASK\nExtract the full substantive text of the agent's report, including its validity judgement, identified errors, and corrected answer if present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The key must be EXACTLY \"full_text\".\n5. If the report is empty or no substantive answer is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "false_conclusion_verification_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is MASSSS3_02.edf. The agent was asked to verify this conclusion: in the EEG segment from minute 517.5 to minute 518.0, the dominant EEG frequency band is beta, and the sleep stage is N1. Ground truth: the true dominant EEG frequency band is theta, and the true sleep stage is N1. The provided conclusion is invalid. The incorrect component(s) are: dominant_band. Score the response in two parts. Give 0.5 points if the response clearly recognizes that the provided conclusion is false or partially false and identifies which component is wrong. Give 0.5 points if the response provides the corrected answer, meaning the true dominant band and true sleep stage, or correctly fixes each wrong component while preserving any component that was already correct. If the response agrees with the provided false conclusion or says the conclusion is valid without finding any error, the score must be 0.0. The total score must therefore be exactly one of 0.0, 0.5, or 1.0. Output only one valid JSON object: {\"score\": 0.0}, {\"score\": 0.5}, or {\"score\": 1.0}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case25/case25_18.json b/sleep_assessment/cases/case25/case25_18.json new file mode 100644 index 0000000000000000000000000000000000000000..b8d0ad03ace23f316c2dfefb64e374201213a395 --- /dev/null +++ b/sleep_assessment/cases/case25/case25_18.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "MASSSS3_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_03.edf", + "instruction": "Please analyze the provided sleep file and examine the EEG segment from minute 78.0 to minute 79.0. Evaluate the claim that the segment's dominant EEG frequency band is theta and that its sleep stage is N3. Report whether the claim is supported, and provide the dominant frequency band and sleep stage you determine for this segment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to preserve the agent's answer for semantic judging.\n\n### TASK\nExtract the full substantive text of the agent's report, including its validity judgement, identified errors, and corrected answer if present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The key must be EXACTLY \"full_text\".\n5. If the report is empty or no substantive answer is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "false_conclusion_verification_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is MASSSS3_03.edf. The agent was asked to verify this conclusion: in the EEG segment from minute 78.0 to minute 79.0, the dominant EEG frequency band is theta, and the sleep stage is N3. Ground truth: the true dominant EEG frequency band is delta, and the true sleep stage is N3. The provided conclusion is invalid. The incorrect component(s) are: dominant_band. Score the response in two parts. Give 0.5 points if the response clearly recognizes that the provided conclusion is false or partially false and identifies which component is wrong. Give 0.5 points if the response provides the corrected answer, meaning the true dominant band and true sleep stage, or correctly fixes each wrong component while preserving any component that was already correct. If the response agrees with the provided false conclusion or says the conclusion is valid without finding any error, the score must be 0.0. The total score must therefore be exactly one of 0.0, 0.5, or 1.0. Output only one valid JSON object: {\"score\": 0.0}, {\"score\": 0.5}, or {\"score\": 1.0}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case25/case25_19.json b/sleep_assessment/cases/case25/case25_19.json new file mode 100644 index 0000000000000000000000000000000000000000..95a1aec0454f451ca319628f282a9af0a49f9cca --- /dev/null +++ b/sleep_assessment/cases/case25/case25_19.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "MASSSS3_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_04.edf", + "instruction": "Please analyze the provided sleep file and examine the EEG segment from minute 57.5 to minute 58.5. Evaluate the claim that the segment's dominant EEG frequency band is delta and that its sleep stage is R. Report whether the claim is supported, and provide the dominant frequency band and sleep stage you determine for this segment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to preserve the agent's answer for semantic judging.\n\n### TASK\nExtract the full substantive text of the agent's report, including its validity judgement, identified errors, and corrected answer if present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The key must be EXACTLY \"full_text\".\n5. If the report is empty or no substantive answer is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "false_conclusion_verification_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is MASSSS3_04.edf. The agent was asked to verify this conclusion: in the EEG segment from minute 57.5 to minute 58.5, the dominant EEG frequency band is delta, and the sleep stage is R. Ground truth: the true dominant EEG frequency band is delta, and the true sleep stage is N3. The provided conclusion is invalid. The incorrect component(s) are: sleep_stage. Score the response in two parts. Give 0.5 points if the response clearly recognizes that the provided conclusion is false or partially false and identifies which component is wrong. Give 0.5 points if the response provides the corrected answer, meaning the true dominant band and true sleep stage, or correctly fixes each wrong component while preserving any component that was already correct. If the response agrees with the provided false conclusion or says the conclusion is valid without finding any error, the score must be 0.0. The total score must therefore be exactly one of 0.0, 0.5, or 1.0. Output only one valid JSON object: {\"score\": 0.0}, {\"score\": 0.5}, or {\"score\": 1.0}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case25/case25_20.json b/sleep_assessment/cases/case25/case25_20.json new file mode 100644 index 0000000000000000000000000000000000000000..a29f8bfaf0e37e804976c308f00ba3e8aaf6182e --- /dev/null +++ b/sleep_assessment/cases/case25/case25_20.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "MASSSS3_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_05.edf", + "instruction": "Please analyze the provided sleep file and examine the EEG segment from minute 80.0 to minute 81.0. Evaluate the claim that the segment's dominant EEG frequency band is alpha and that its sleep stage is W. Report whether the claim is supported, and provide the dominant frequency band and sleep stage you determine for this segment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to preserve the agent's answer for semantic judging.\n\n### TASK\nExtract the full substantive text of the agent's report, including its validity judgement, identified errors, and corrected answer if present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The key must be EXACTLY \"full_text\".\n5. If the report is empty or no substantive answer is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "false_conclusion_verification_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is MASSSS3_05.edf. The agent was asked to verify this conclusion: in the EEG segment from minute 80.0 to minute 81.0, the dominant EEG frequency band is alpha, and the sleep stage is W. Ground truth: the true dominant EEG frequency band is delta, and the true sleep stage is N3. The provided conclusion is invalid. The incorrect component(s) are: dominant_band, sleep_stage. Score the response in two parts. Give 0.5 points if the response clearly recognizes that the provided conclusion is false or partially false and identifies which component is wrong. Give 0.5 points if the response provides the corrected answer, meaning the true dominant band and true sleep stage, or correctly fixes each wrong component while preserving any component that was already correct. If the response agrees with the provided false conclusion or says the conclusion is valid without finding any error, the score must be 0.0. The total score must therefore be exactly one of 0.0, 0.5, or 1.0. Output only one valid JSON object: {\"score\": 0.0}, {\"score\": 0.5}, or {\"score\": 1.0}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case25/case25_21.json b/sleep_assessment/cases/case25/case25_21.json new file mode 100644 index 0000000000000000000000000000000000000000..4a960f710f7a634745e6667c789821e66f697293 --- /dev/null +++ b/sleep_assessment/cases/case25/case25_21.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "instruction": "Please analyze the provided sleep file and examine the EEG segment from minute 8.5 to minute 9.0. Evaluate the claim that the segment's dominant EEG frequency band is alpha and that its sleep stage is W. Report whether the claim is supported, and provide the dominant frequency band and sleep stage you determine for this segment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to preserve the agent's answer for semantic judging.\n\n### TASK\nExtract the full substantive text of the agent's report, including its validity judgement, identified errors, and corrected answer if present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The key must be EXACTLY \"full_text\".\n5. If the report is empty or no substantive answer is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "false_conclusion_verification_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is Physionet2018_01.edf. The agent was asked to verify this conclusion: in the EEG segment from minute 8.5 to minute 9.0, the dominant EEG frequency band is alpha, and the sleep stage is W. Ground truth: the true dominant EEG frequency band is delta, and the true sleep stage is N1. The provided conclusion is invalid. The incorrect component(s) are: dominant_band, sleep_stage. Score the response in two parts. Give 0.5 points if the response clearly recognizes that the provided conclusion is false or partially false and identifies which component is wrong. Give 0.5 points if the response provides the corrected answer, meaning the true dominant band and true sleep stage, or correctly fixes each wrong component while preserving any component that was already correct. If the response agrees with the provided false conclusion or says the conclusion is valid without finding any error, the score must be 0.0. The total score must therefore be exactly one of 0.0, 0.5, or 1.0. Output only one valid JSON object: {\"score\": 0.0}, {\"score\": 0.5}, or {\"score\": 1.0}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case25/case25_22.json b/sleep_assessment/cases/case25/case25_22.json new file mode 100644 index 0000000000000000000000000000000000000000..bc2b51987cee8947e2172f40eaa615fb489fd60b --- /dev/null +++ b/sleep_assessment/cases/case25/case25_22.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "instruction": "Please analyze the provided sleep file and examine the EEG segment from minute 131.5 to minute 132.0. Evaluate the claim that the segment's dominant EEG frequency band is theta and that its sleep stage is W. Report whether the claim is supported, and provide the dominant frequency band and sleep stage you determine for this segment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to preserve the agent's answer for semantic judging.\n\n### TASK\nExtract the full substantive text of the agent's report, including its validity judgement, identified errors, and corrected answer if present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The key must be EXACTLY \"full_text\".\n5. If the report is empty or no substantive answer is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "false_conclusion_verification_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is Physionet2018_02.edf. The agent was asked to verify this conclusion: in the EEG segment from minute 131.5 to minute 132.0, the dominant EEG frequency band is theta, and the sleep stage is W. Ground truth: the true dominant EEG frequency band is delta, and the true sleep stage is W. The provided conclusion is invalid. The incorrect component(s) are: dominant_band. Score the response in two parts. Give 0.5 points if the response clearly recognizes that the provided conclusion is false or partially false and identifies which component is wrong. Give 0.5 points if the response provides the corrected answer, meaning the true dominant band and true sleep stage, or correctly fixes each wrong component while preserving any component that was already correct. If the response agrees with the provided false conclusion or says the conclusion is valid without finding any error, the score must be 0.0. The total score must therefore be exactly one of 0.0, 0.5, or 1.0. Output only one valid JSON object: {\"score\": 0.0}, {\"score\": 0.5}, or {\"score\": 1.0}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case25/case25_23.json b/sleep_assessment/cases/case25/case25_23.json new file mode 100644 index 0000000000000000000000000000000000000000..b576ad7b93a0965c6961231b37d3dd0604d85437 --- /dev/null +++ b/sleep_assessment/cases/case25/case25_23.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Physionet2018_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_03.edf", + "instruction": "Please analyze the provided sleep file and examine the EEG segment from minute 452.0 to minute 452.5. Evaluate the claim that the segment's dominant EEG frequency band is theta and that its sleep stage is W. Report whether the claim is supported, and provide the dominant frequency band and sleep stage you determine for this segment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to preserve the agent's answer for semantic judging.\n\n### TASK\nExtract the full substantive text of the agent's report, including its validity judgement, identified errors, and corrected answer if present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The key must be EXACTLY \"full_text\".\n5. If the report is empty or no substantive answer is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "false_conclusion_verification_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is Physionet2018_03.edf. The agent was asked to verify this conclusion: in the EEG segment from minute 452.0 to minute 452.5, the dominant EEG frequency band is theta, and the sleep stage is W. Ground truth: the true dominant EEG frequency band is alpha, and the true sleep stage is W. The provided conclusion is invalid. The incorrect component(s) are: dominant_band. Score the response in two parts. Give 0.5 points if the response clearly recognizes that the provided conclusion is false or partially false and identifies which component is wrong. Give 0.5 points if the response provides the corrected answer, meaning the true dominant band and true sleep stage, or correctly fixes each wrong component while preserving any component that was already correct. If the response agrees with the provided false conclusion or says the conclusion is valid without finding any error, the score must be 0.0. The total score must therefore be exactly one of 0.0, 0.5, or 1.0. Output only one valid JSON object: {\"score\": 0.0}, {\"score\": 0.5}, or {\"score\": 1.0}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case25/case25_24.json b/sleep_assessment/cases/case25/case25_24.json new file mode 100644 index 0000000000000000000000000000000000000000..87fad54b624b59715d314ec9ddc5acd5c9fe9e92 --- /dev/null +++ b/sleep_assessment/cases/case25/case25_24.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "instruction": "Please analyze the provided sleep file and examine the EEG segment from minute 121.5 to minute 124.5. Evaluate the claim that the segment's dominant EEG frequency band is theta and that its sleep stage is N1. Report whether the claim is supported, and provide the dominant frequency band and sleep stage you determine for this segment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to preserve the agent's answer for semantic judging.\n\n### TASK\nExtract the full substantive text of the agent's report, including its validity judgement, identified errors, and corrected answer if present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The key must be EXACTLY \"full_text\".\n5. If the report is empty or no substantive answer is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "false_conclusion_verification_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is Physionet2018_04.edf. The agent was asked to verify this conclusion: in the EEG segment from minute 121.5 to minute 124.5, the dominant EEG frequency band is theta, and the sleep stage is N1. Ground truth: the true dominant EEG frequency band is delta, and the true sleep stage is N3. The provided conclusion is invalid. The incorrect component(s) are: dominant_band, sleep_stage. Score the response in two parts. Give 0.5 points if the response clearly recognizes that the provided conclusion is false or partially false and identifies which component is wrong. Give 0.5 points if the response provides the corrected answer, meaning the true dominant band and true sleep stage, or correctly fixes each wrong component while preserving any component that was already correct. If the response agrees with the provided false conclusion or says the conclusion is valid without finding any error, the score must be 0.0. The total score must therefore be exactly one of 0.0, 0.5, or 1.0. Output only one valid JSON object: {\"score\": 0.0}, {\"score\": 0.5}, or {\"score\": 1.0}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case25/case25_25.json b/sleep_assessment/cases/case25/case25_25.json new file mode 100644 index 0000000000000000000000000000000000000000..5d5056c1e1b42ac0361b33ea417c30876b0d7ec1 --- /dev/null +++ b/sleep_assessment/cases/case25/case25_25.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "instruction": "Please analyze the provided sleep file and examine the EEG segment from minute 364.5 to minute 367.5. Evaluate the claim that the segment's dominant EEG frequency band is theta and that its sleep stage is N3. Report whether the claim is supported, and provide the dominant frequency band and sleep stage you determine for this segment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to preserve the agent's answer for semantic judging.\n\n### TASK\nExtract the full substantive text of the agent's report, including its validity judgement, identified errors, and corrected answer if present.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The key must be EXACTLY \"full_text\".\n5. If the report is empty or no substantive answer is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "false_conclusion_verification_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is Physionet2018_05.edf. The agent was asked to verify this conclusion: in the EEG segment from minute 364.5 to minute 367.5, the dominant EEG frequency band is theta, and the sleep stage is N3. Ground truth: the true dominant EEG frequency band is delta, and the true sleep stage is N3. The provided conclusion is invalid. The incorrect component(s) are: dominant_band. Score the response in two parts. Give 0.5 points if the response clearly recognizes that the provided conclusion is false or partially false and identifies which component is wrong. Give 0.5 points if the response provides the corrected answer, meaning the true dominant band and true sleep stage, or correctly fixes each wrong component while preserving any component that was already correct. If the response agrees with the provided false conclusion or says the conclusion is valid without finding any error, the score must be 0.0. The total score must therefore be exactly one of 0.0, 0.5, or 1.0. Output only one valid JSON object: {\"score\": 0.0}, {\"score\": 0.5}, or {\"score\": 1.0}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case26/case26_01.json b/sleep_assessment/cases/case26/case26_01.json new file mode 100644 index 0000000000000000000000000000000000000000..a67b56f0228556607b0343cbb4262afb834f1129 --- /dev/null +++ b/sleep_assessment/cases/case26/case26_01.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 274.5 to 275.5, B = minute 63.0 to 64.0, C = minute 437.0 to 438.0. Compare the EOG/eye-movement activity strength across A, B, and C. Report the segment labels in descending order of EOG activity strength, and state whether the strongest segment is REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the descending order of EOG activity strength across segments A, B, and C\n2) whether the strongest segment is REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"eog_activity_order_desc\" and \"strongest_segment_is_rem\".\n5. \"eog_activity_order_desc\" must be a JSON array with EXACTLY 3 elements.\n6. The array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. \"strongest_segment_is_rem\" must be exactly one of: \"yes\", \"no\", or null.\n8. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n9. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"eog_activity_order_desc\": [\"B\", \"A\", \"C\"], \"strongest_segment_is_rem\": \"yes\"}", + "metrics": [ + { + "metric_id": "eog_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "eog_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "A", + "B", + "C" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "strongest_segment_is_rem_accuracy", + "type": "categorical_check", + "target_key": "strongest_segment_is_rem", + "weight": 40, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case26/case26_02.json b/sleep_assessment/cases/case26/case26_02.json new file mode 100644 index 0000000000000000000000000000000000000000..84b302f5d87920aa259eafb50eb41baa153bc76f --- /dev/null +++ b/sleep_assessment/cases/case26/case26_02.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 473.5 to 474.5, B = minute 359.5 to 360.5, C = minute 135.5 to 136.5. Compare the EOG/eye-movement activity strength across A, B, and C. Report the segment labels in descending order of EOG activity strength, and state whether the strongest segment is REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the descending order of EOG activity strength across segments A, B, and C\n2) whether the strongest segment is REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"eog_activity_order_desc\" and \"strongest_segment_is_rem\".\n5. \"eog_activity_order_desc\" must be a JSON array with EXACTLY 3 elements.\n6. The array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. \"strongest_segment_is_rem\" must be exactly one of: \"yes\", \"no\", or null.\n8. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n9. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"eog_activity_order_desc\": [\"B\", \"A\", \"C\"], \"strongest_segment_is_rem\": \"yes\"}", + "metrics": [ + { + "metric_id": "eog_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "eog_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "B", + "C", + "A" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "strongest_segment_is_rem_accuracy", + "type": "categorical_check", + "target_key": "strongest_segment_is_rem", + "weight": 40, + "params": { + "gt_value": "yes" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case26/case26_03.json b/sleep_assessment/cases/case26/case26_03.json new file mode 100644 index 0000000000000000000000000000000000000000..a6349a2e307a7f5bcebbca0f5fe78a0e735154b0 --- /dev/null +++ b/sleep_assessment/cases/case26/case26_03.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 210.5 to 211.5, B = minute 334.0 to 335.0, C = minute 322.0 to 323.0. Compare the EOG/eye-movement activity strength across A, B, and C. Report the segment labels in descending order of EOG activity strength, and state whether the strongest segment is REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the descending order of EOG activity strength across segments A, B, and C\n2) whether the strongest segment is REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"eog_activity_order_desc\" and \"strongest_segment_is_rem\".\n5. \"eog_activity_order_desc\" must be a JSON array with EXACTLY 3 elements.\n6. The array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. \"strongest_segment_is_rem\" must be exactly one of: \"yes\", \"no\", or null.\n8. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n9. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"eog_activity_order_desc\": [\"B\", \"A\", \"C\"], \"strongest_segment_is_rem\": \"yes\"}", + "metrics": [ + { + "metric_id": "eog_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "eog_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "C", + "A", + "B" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "strongest_segment_is_rem_accuracy", + "type": "categorical_check", + "target_key": "strongest_segment_is_rem", + "weight": 40, + "params": { + "gt_value": "yes" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case26/case26_04.json b/sleep_assessment/cases/case26/case26_04.json new file mode 100644 index 0000000000000000000000000000000000000000..accb13895fb27c4f4e1cb6cb75e14c079c63b1fc --- /dev/null +++ b/sleep_assessment/cases/case26/case26_04.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "ISRUC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_04.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 349.0 to 350.0, B = minute 375.0 to 376.0, C = minute 97.0 to 98.0. Compare the EOG/eye-movement activity strength across A, B, and C. Report the segment labels in descending order of EOG activity strength, and state whether the strongest segment is REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the descending order of EOG activity strength across segments A, B, and C\n2) whether the strongest segment is REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"eog_activity_order_desc\" and \"strongest_segment_is_rem\".\n5. \"eog_activity_order_desc\" must be a JSON array with EXACTLY 3 elements.\n6. The array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. \"strongest_segment_is_rem\" must be exactly one of: \"yes\", \"no\", or null.\n8. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n9. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"eog_activity_order_desc\": [\"B\", \"A\", \"C\"], \"strongest_segment_is_rem\": \"yes\"}", + "metrics": [ + { + "metric_id": "eog_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "eog_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "A", + "C", + "B" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "strongest_segment_is_rem_accuracy", + "type": "categorical_check", + "target_key": "strongest_segment_is_rem", + "weight": 40, + "params": { + "gt_value": "yes" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case26/case26_05.json b/sleep_assessment/cases/case26/case26_05.json new file mode 100644 index 0000000000000000000000000000000000000000..155aeee9fb7fdb5e067995ecaa18a1edd8cc359a --- /dev/null +++ b/sleep_assessment/cases/case26/case26_05.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 64.0 to 65.0, B = minute 423.0 to 424.0, C = minute 366.5 to 367.5. Compare the EOG/eye-movement activity strength across A, B, and C. Report the segment labels in descending order of EOG activity strength, and state whether the strongest segment is REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the descending order of EOG activity strength across segments A, B, and C\n2) whether the strongest segment is REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"eog_activity_order_desc\" and \"strongest_segment_is_rem\".\n5. \"eog_activity_order_desc\" must be a JSON array with EXACTLY 3 elements.\n6. The array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. \"strongest_segment_is_rem\" must be exactly one of: \"yes\", \"no\", or null.\n8. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n9. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"eog_activity_order_desc\": [\"B\", \"A\", \"C\"], \"strongest_segment_is_rem\": \"yes\"}", + "metrics": [ + { + "metric_id": "eog_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "eog_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "B", + "A", + "C" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "strongest_segment_is_rem_accuracy", + "type": "categorical_check", + "target_key": "strongest_segment_is_rem", + "weight": 40, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case26/case26_06.json b/sleep_assessment/cases/case26/case26_06.json new file mode 100644 index 0000000000000000000000000000000000000000..0bb0228b82c024f6ecb6736418279b7d1c687d2d --- /dev/null +++ b/sleep_assessment/cases/case26/case26_06.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "HMC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_01.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 278.5 to 279.5, B = minute 24.5 to 25.5, C = minute 398.5 to 399.5. Compare the EOG/eye-movement activity strength across A, B, and C. Report the segment labels in descending order of EOG activity strength, and state whether the strongest segment is REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the descending order of EOG activity strength across segments A, B, and C\n2) whether the strongest segment is REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"eog_activity_order_desc\" and \"strongest_segment_is_rem\".\n5. \"eog_activity_order_desc\" must be a JSON array with EXACTLY 3 elements.\n6. The array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. \"strongest_segment_is_rem\" must be exactly one of: \"yes\", \"no\", or null.\n8. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n9. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"eog_activity_order_desc\": [\"B\", \"A\", \"C\"], \"strongest_segment_is_rem\": \"yes\"}", + "metrics": [ + { + "metric_id": "eog_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "eog_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "A", + "B", + "C" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "strongest_segment_is_rem_accuracy", + "type": "categorical_check", + "target_key": "strongest_segment_is_rem", + "weight": 40, + "params": { + "gt_value": "yes" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case26/case26_07.json b/sleep_assessment/cases/case26/case26_07.json new file mode 100644 index 0000000000000000000000000000000000000000..274b21fb17676bd0904993eba807fb14a5f8344e --- /dev/null +++ b/sleep_assessment/cases/case26/case26_07.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "HMC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_02.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 409.0 to 410.0, B = minute 217.5 to 218.5, C = minute 60.5 to 61.5. Compare the EOG/eye-movement activity strength across A, B, and C. Report the segment labels in descending order of EOG activity strength, and state whether the strongest segment is REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the descending order of EOG activity strength across segments A, B, and C\n2) whether the strongest segment is REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"eog_activity_order_desc\" and \"strongest_segment_is_rem\".\n5. \"eog_activity_order_desc\" must be a JSON array with EXACTLY 3 elements.\n6. The array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. \"strongest_segment_is_rem\" must be exactly one of: \"yes\", \"no\", or null.\n8. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n9. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"eog_activity_order_desc\": [\"B\", \"A\", \"C\"], \"strongest_segment_is_rem\": \"yes\"}", + "metrics": [ + { + "metric_id": "eog_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "eog_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "B", + "C", + "A" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "strongest_segment_is_rem_accuracy", + "type": "categorical_check", + "target_key": "strongest_segment_is_rem", + "weight": 40, + "params": { + "gt_value": "yes" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case26/case26_08.json b/sleep_assessment/cases/case26/case26_08.json new file mode 100644 index 0000000000000000000000000000000000000000..7e6579364e87104aea515dd50233042b52016582 --- /dev/null +++ b/sleep_assessment/cases/case26/case26_08.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "HMC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_03.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 110.0 to 111.0, B = minute 461.5 to 462.5, C = minute 65.5 to 66.5. Compare the EOG/eye-movement activity strength across A, B, and C. Report the segment labels in descending order of EOG activity strength, and state whether the strongest segment is REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the descending order of EOG activity strength across segments A, B, and C\n2) whether the strongest segment is REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"eog_activity_order_desc\" and \"strongest_segment_is_rem\".\n5. \"eog_activity_order_desc\" must be a JSON array with EXACTLY 3 elements.\n6. The array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. \"strongest_segment_is_rem\" must be exactly one of: \"yes\", \"no\", or null.\n8. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n9. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"eog_activity_order_desc\": [\"B\", \"A\", \"C\"], \"strongest_segment_is_rem\": \"yes\"}", + "metrics": [ + { + "metric_id": "eog_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "eog_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "C", + "A", + "B" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "strongest_segment_is_rem_accuracy", + "type": "categorical_check", + "target_key": "strongest_segment_is_rem", + "weight": 40, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case26/case26_09.json b/sleep_assessment/cases/case26/case26_09.json new file mode 100644 index 0000000000000000000000000000000000000000..204edc575f3da3de65d3eec05f2eb773bce6e52d --- /dev/null +++ b/sleep_assessment/cases/case26/case26_09.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "HMC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_04.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 353.5 to 354.5, B = minute 387.5 to 388.5, C = minute 5.0 to 6.0. Compare the EOG/eye-movement activity strength across A, B, and C. Report the segment labels in descending order of EOG activity strength, and state whether the strongest segment is REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the descending order of EOG activity strength across segments A, B, and C\n2) whether the strongest segment is REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"eog_activity_order_desc\" and \"strongest_segment_is_rem\".\n5. \"eog_activity_order_desc\" must be a JSON array with EXACTLY 3 elements.\n6. The array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. \"strongest_segment_is_rem\" must be exactly one of: \"yes\", \"no\", or null.\n8. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n9. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"eog_activity_order_desc\": [\"B\", \"A\", \"C\"], \"strongest_segment_is_rem\": \"yes\"}", + "metrics": [ + { + "metric_id": "eog_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "eog_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "A", + "C", + "B" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "strongest_segment_is_rem_accuracy", + "type": "categorical_check", + "target_key": "strongest_segment_is_rem", + "weight": 40, + "params": { + "gt_value": "yes" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case26/case26_10.json b/sleep_assessment/cases/case26/case26_10.json new file mode 100644 index 0000000000000000000000000000000000000000..3e707a5dfe4b343eea4ad6023bd00b13968332c3 --- /dev/null +++ b/sleep_assessment/cases/case26/case26_10.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "HMC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_05.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 427.0 to 428.0, B = minute 19.5 to 20.5, C = minute 244.5 to 245.5. Compare the EOG/eye-movement activity strength across A, B, and C. Report the segment labels in descending order of EOG activity strength, and state whether the strongest segment is REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the descending order of EOG activity strength across segments A, B, and C\n2) whether the strongest segment is REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"eog_activity_order_desc\" and \"strongest_segment_is_rem\".\n5. \"eog_activity_order_desc\" must be a JSON array with EXACTLY 3 elements.\n6. The array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. \"strongest_segment_is_rem\" must be exactly one of: \"yes\", \"no\", or null.\n8. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n9. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"eog_activity_order_desc\": [\"B\", \"A\", \"C\"], \"strongest_segment_is_rem\": \"yes\"}", + "metrics": [ + { + "metric_id": "eog_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "eog_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "B", + "A", + "C" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "strongest_segment_is_rem_accuracy", + "type": "categorical_check", + "target_key": "strongest_segment_is_rem", + "weight": 40, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case26/case26_11.json b/sleep_assessment/cases/case26/case26_11.json new file mode 100644 index 0000000000000000000000000000000000000000..1414372578f27ca86f4320986e9d1f50091df3c2 --- /dev/null +++ b/sleep_assessment/cases/case26/case26_11.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 432.5 to 433.5, B = minute 96.5 to 97.5, C = minute 168.5 to 169.5. Compare the EOG/eye-movement activity strength across A, B, and C. Report the segment labels in descending order of EOG activity strength, and state whether the strongest segment is REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the descending order of EOG activity strength across segments A, B, and C\n2) whether the strongest segment is REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"eog_activity_order_desc\" and \"strongest_segment_is_rem\".\n5. \"eog_activity_order_desc\" must be a JSON array with EXACTLY 3 elements.\n6. The array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. \"strongest_segment_is_rem\" must be exactly one of: \"yes\", \"no\", or null.\n8. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n9. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"eog_activity_order_desc\": [\"B\", \"A\", \"C\"], \"strongest_segment_is_rem\": \"yes\"}", + "metrics": [ + { + "metric_id": "eog_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "eog_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "A", + "B", + "C" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "strongest_segment_is_rem_accuracy", + "type": "categorical_check", + "target_key": "strongest_segment_is_rem", + "weight": 40, + "params": { + "gt_value": "yes" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case26/case26_12.json b/sleep_assessment/cases/case26/case26_12.json new file mode 100644 index 0000000000000000000000000000000000000000..192d26bafae1de1e9613a72ec2c31b0f51f31ec8 --- /dev/null +++ b/sleep_assessment/cases/case26/case26_12.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 223.0 to 224.0, B = minute 423.5 to 424.5, C = minute 230.0 to 231.0. Compare the EOG/eye-movement activity strength across A, B, and C. Report the segment labels in descending order of EOG activity strength, and state whether the strongest segment is REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the descending order of EOG activity strength across segments A, B, and C\n2) whether the strongest segment is REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"eog_activity_order_desc\" and \"strongest_segment_is_rem\".\n5. \"eog_activity_order_desc\" must be a JSON array with EXACTLY 3 elements.\n6. The array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. \"strongest_segment_is_rem\" must be exactly one of: \"yes\", \"no\", or null.\n8. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n9. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"eog_activity_order_desc\": [\"B\", \"A\", \"C\"], \"strongest_segment_is_rem\": \"yes\"}", + "metrics": [ + { + "metric_id": "eog_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "eog_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "B", + "C", + "A" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "strongest_segment_is_rem_accuracy", + "type": "categorical_check", + "target_key": "strongest_segment_is_rem", + "weight": 40, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case26/case26_13.json b/sleep_assessment/cases/case26/case26_13.json new file mode 100644 index 0000000000000000000000000000000000000000..7269b23f4b67e8456a752e54421b22d5b93446ef --- /dev/null +++ b/sleep_assessment/cases/case26/case26_13.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 144.0 to 145.0, B = minute 439.5 to 440.5, C = minute 461.0 to 462.0. Compare the EOG/eye-movement activity strength across A, B, and C. Report the segment labels in descending order of EOG activity strength, and state whether the strongest segment is REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the descending order of EOG activity strength across segments A, B, and C\n2) whether the strongest segment is REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"eog_activity_order_desc\" and \"strongest_segment_is_rem\".\n5. \"eog_activity_order_desc\" must be a JSON array with EXACTLY 3 elements.\n6. The array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. \"strongest_segment_is_rem\" must be exactly one of: \"yes\", \"no\", or null.\n8. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n9. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"eog_activity_order_desc\": [\"B\", \"A\", \"C\"], \"strongest_segment_is_rem\": \"yes\"}", + "metrics": [ + { + "metric_id": "eog_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "eog_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "C", + "A", + "B" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "strongest_segment_is_rem_accuracy", + "type": "categorical_check", + "target_key": "strongest_segment_is_rem", + "weight": 40, + "params": { + "gt_value": "yes" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case26/case26_14.json b/sleep_assessment/cases/case26/case26_14.json new file mode 100644 index 0000000000000000000000000000000000000000..f332a8c060e1f0fddf3c4152da5b281bbc65d819 --- /dev/null +++ b/sleep_assessment/cases/case26/case26_14.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 436.5 to 437.5, B = minute 305.0 to 306.0, C = minute 124.5 to 125.5. Compare the EOG/eye-movement activity strength across A, B, and C. Report the segment labels in descending order of EOG activity strength, and state whether the strongest segment is REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the descending order of EOG activity strength across segments A, B, and C\n2) whether the strongest segment is REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"eog_activity_order_desc\" and \"strongest_segment_is_rem\".\n5. \"eog_activity_order_desc\" must be a JSON array with EXACTLY 3 elements.\n6. The array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. \"strongest_segment_is_rem\" must be exactly one of: \"yes\", \"no\", or null.\n8. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n9. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"eog_activity_order_desc\": [\"B\", \"A\", \"C\"], \"strongest_segment_is_rem\": \"yes\"}", + "metrics": [ + { + "metric_id": "eog_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "eog_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "A", + "C", + "B" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "strongest_segment_is_rem_accuracy", + "type": "categorical_check", + "target_key": "strongest_segment_is_rem", + "weight": 40, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case26/case26_15.json b/sleep_assessment/cases/case26/case26_15.json new file mode 100644 index 0000000000000000000000000000000000000000..324e9dffd19913a87055faaba42152cb4532933f --- /dev/null +++ b/sleep_assessment/cases/case26/case26_15.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 472.0 to 473.0, B = minute 316.5 to 317.5, C = minute 12.0 to 13.0. Compare the EOG/eye-movement activity strength across A, B, and C. Report the segment labels in descending order of EOG activity strength, and state whether the strongest segment is REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the descending order of EOG activity strength across segments A, B, and C\n2) whether the strongest segment is REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"eog_activity_order_desc\" and \"strongest_segment_is_rem\".\n5. \"eog_activity_order_desc\" must be a JSON array with EXACTLY 3 elements.\n6. The array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. \"strongest_segment_is_rem\" must be exactly one of: \"yes\", \"no\", or null.\n8. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n9. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"eog_activity_order_desc\": [\"B\", \"A\", \"C\"], \"strongest_segment_is_rem\": \"yes\"}", + "metrics": [ + { + "metric_id": "eog_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "eog_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "B", + "A", + "C" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "strongest_segment_is_rem_accuracy", + "type": "categorical_check", + "target_key": "strongest_segment_is_rem", + "weight": 40, + "params": { + "gt_value": "yes" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case26/case26_16.json b/sleep_assessment/cases/case26/case26_16.json new file mode 100644 index 0000000000000000000000000000000000000000..0b52b2e3206e1928103806dd432b0c24b57f876a --- /dev/null +++ b/sleep_assessment/cases/case26/case26_16.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "MASSSS3_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_01.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 371.0 to 372.0, B = minute 219.0 to 220.0, C = minute 425.5 to 426.5. Compare the EOG/eye-movement activity strength across A, B, and C. Report the segment labels in descending order of EOG activity strength, and state whether the strongest segment is REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the descending order of EOG activity strength across segments A, B, and C\n2) whether the strongest segment is REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"eog_activity_order_desc\" and \"strongest_segment_is_rem\".\n5. \"eog_activity_order_desc\" must be a JSON array with EXACTLY 3 elements.\n6. The array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. \"strongest_segment_is_rem\" must be exactly one of: \"yes\", \"no\", or null.\n8. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n9. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"eog_activity_order_desc\": [\"B\", \"A\", \"C\"], \"strongest_segment_is_rem\": \"yes\"}", + "metrics": [ + { + "metric_id": "eog_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "eog_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "A", + "B", + "C" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "strongest_segment_is_rem_accuracy", + "type": "categorical_check", + "target_key": "strongest_segment_is_rem", + "weight": 40, + "params": { + "gt_value": "yes" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case26/case26_17.json b/sleep_assessment/cases/case26/case26_17.json new file mode 100644 index 0000000000000000000000000000000000000000..283c0395050a0aad873cd464907f1883960fe83f --- /dev/null +++ b/sleep_assessment/cases/case26/case26_17.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "MASSSS3_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_02.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 375.5 to 376.5, B = minute 463.5 to 464.5, C = minute 16.5 to 17.5. Compare the EOG/eye-movement activity strength across A, B, and C. Report the segment labels in descending order of EOG activity strength, and state whether the strongest segment is REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the descending order of EOG activity strength across segments A, B, and C\n2) whether the strongest segment is REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"eog_activity_order_desc\" and \"strongest_segment_is_rem\".\n5. \"eog_activity_order_desc\" must be a JSON array with EXACTLY 3 elements.\n6. The array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. \"strongest_segment_is_rem\" must be exactly one of: \"yes\", \"no\", or null.\n8. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n9. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"eog_activity_order_desc\": [\"B\", \"A\", \"C\"], \"strongest_segment_is_rem\": \"yes\"}", + "metrics": [ + { + "metric_id": "eog_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "eog_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "B", + "C", + "A" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "strongest_segment_is_rem_accuracy", + "type": "categorical_check", + "target_key": "strongest_segment_is_rem", + "weight": 40, + "params": { + "gt_value": "yes" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case26/case26_18.json b/sleep_assessment/cases/case26/case26_18.json new file mode 100644 index 0000000000000000000000000000000000000000..26070db752d9558f9504598a12c7fe55c12cbe8e --- /dev/null +++ b/sleep_assessment/cases/case26/case26_18.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "MASSSS3_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_03.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 54.5 to 55.5, B = minute 193.0 to 194.0, C = minute 410.0 to 411.0. Compare the EOG/eye-movement activity strength across A, B, and C. Report the segment labels in descending order of EOG activity strength, and state whether the strongest segment is REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the descending order of EOG activity strength across segments A, B, and C\n2) whether the strongest segment is REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"eog_activity_order_desc\" and \"strongest_segment_is_rem\".\n5. \"eog_activity_order_desc\" must be a JSON array with EXACTLY 3 elements.\n6. The array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. \"strongest_segment_is_rem\" must be exactly one of: \"yes\", \"no\", or null.\n8. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n9. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"eog_activity_order_desc\": [\"B\", \"A\", \"C\"], \"strongest_segment_is_rem\": \"yes\"}", + "metrics": [ + { + "metric_id": "eog_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "eog_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "C", + "A", + "B" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "strongest_segment_is_rem_accuracy", + "type": "categorical_check", + "target_key": "strongest_segment_is_rem", + "weight": 40, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case26/case26_19.json b/sleep_assessment/cases/case26/case26_19.json new file mode 100644 index 0000000000000000000000000000000000000000..fae1bb20523a9cba4e83529146397d9a926b28ea --- /dev/null +++ b/sleep_assessment/cases/case26/case26_19.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "MASSSS3_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_04.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 0.0 to 1.0, B = minute 176.5 to 177.5, C = minute 133.5 to 134.5. Compare the EOG/eye-movement activity strength across A, B, and C. Report the segment labels in descending order of EOG activity strength, and state whether the strongest segment is REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the descending order of EOG activity strength across segments A, B, and C\n2) whether the strongest segment is REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"eog_activity_order_desc\" and \"strongest_segment_is_rem\".\n5. \"eog_activity_order_desc\" must be a JSON array with EXACTLY 3 elements.\n6. The array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. \"strongest_segment_is_rem\" must be exactly one of: \"yes\", \"no\", or null.\n8. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n9. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"eog_activity_order_desc\": [\"B\", \"A\", \"C\"], \"strongest_segment_is_rem\": \"yes\"}", + "metrics": [ + { + "metric_id": "eog_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "eog_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "A", + "C", + "B" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "strongest_segment_is_rem_accuracy", + "type": "categorical_check", + "target_key": "strongest_segment_is_rem", + "weight": 40, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case26/case26_20.json b/sleep_assessment/cases/case26/case26_20.json new file mode 100644 index 0000000000000000000000000000000000000000..67dd49b0c7136279af41cd7bd502a5bea6c31e6e --- /dev/null +++ b/sleep_assessment/cases/case26/case26_20.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "MASSSS3_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_05.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 356.0 to 357.0, B = minute 407.0 to 408.0, C = minute 300.5 to 301.5. Compare the EOG/eye-movement activity strength across A, B, and C. Report the segment labels in descending order of EOG activity strength, and state whether the strongest segment is REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the descending order of EOG activity strength across segments A, B, and C\n2) whether the strongest segment is REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"eog_activity_order_desc\" and \"strongest_segment_is_rem\".\n5. \"eog_activity_order_desc\" must be a JSON array with EXACTLY 3 elements.\n6. The array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. \"strongest_segment_is_rem\" must be exactly one of: \"yes\", \"no\", or null.\n8. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n9. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"eog_activity_order_desc\": [\"B\", \"A\", \"C\"], \"strongest_segment_is_rem\": \"yes\"}", + "metrics": [ + { + "metric_id": "eog_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "eog_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "B", + "A", + "C" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "strongest_segment_is_rem_accuracy", + "type": "categorical_check", + "target_key": "strongest_segment_is_rem", + "weight": 40, + "params": { + "gt_value": "yes" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case26/case26_21.json b/sleep_assessment/cases/case26/case26_21.json new file mode 100644 index 0000000000000000000000000000000000000000..ba233c29aad63449fc879aeeb9d661084c625d32 --- /dev/null +++ b/sleep_assessment/cases/case26/case26_21.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 227.5 to 228.5, B = minute 196.0 to 197.0, C = minute 96.0 to 97.0. Compare the EOG/eye-movement activity strength across A, B, and C. Report the segment labels in descending order of EOG activity strength, and state whether the strongest segment is REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the descending order of EOG activity strength across segments A, B, and C\n2) whether the strongest segment is REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"eog_activity_order_desc\" and \"strongest_segment_is_rem\".\n5. \"eog_activity_order_desc\" must be a JSON array with EXACTLY 3 elements.\n6. The array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. \"strongest_segment_is_rem\" must be exactly one of: \"yes\", \"no\", or null.\n8. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n9. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"eog_activity_order_desc\": [\"B\", \"A\", \"C\"], \"strongest_segment_is_rem\": \"yes\"}", + "metrics": [ + { + "metric_id": "eog_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "eog_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "A", + "B", + "C" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "strongest_segment_is_rem_accuracy", + "type": "categorical_check", + "target_key": "strongest_segment_is_rem", + "weight": 40, + "params": { + "gt_value": "yes" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case26/case26_22.json b/sleep_assessment/cases/case26/case26_22.json new file mode 100644 index 0000000000000000000000000000000000000000..eff6b582b753f2da1c96565ee8c3db7ae2db43d8 --- /dev/null +++ b/sleep_assessment/cases/case26/case26_22.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 291.0 to 292.0, B = minute 232.5 to 233.5, C = minute 168.5 to 169.5. Compare the EOG/eye-movement activity strength across A, B, and C. Report the segment labels in descending order of EOG activity strength, and state whether the strongest segment is REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the descending order of EOG activity strength across segments A, B, and C\n2) whether the strongest segment is REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"eog_activity_order_desc\" and \"strongest_segment_is_rem\".\n5. \"eog_activity_order_desc\" must be a JSON array with EXACTLY 3 elements.\n6. The array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. \"strongest_segment_is_rem\" must be exactly one of: \"yes\", \"no\", or null.\n8. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n9. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"eog_activity_order_desc\": [\"B\", \"A\", \"C\"], \"strongest_segment_is_rem\": \"yes\"}", + "metrics": [ + { + "metric_id": "eog_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "eog_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "B", + "C", + "A" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "strongest_segment_is_rem_accuracy", + "type": "categorical_check", + "target_key": "strongest_segment_is_rem", + "weight": 40, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case26/case26_23.json b/sleep_assessment/cases/case26/case26_23.json new file mode 100644 index 0000000000000000000000000000000000000000..34c0cd06662de5a422807440bb53c4b4a43c030c --- /dev/null +++ b/sleep_assessment/cases/case26/case26_23.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "Physionet2018_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_03.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 102.0 to 103.0, B = minute 425.0 to 426.0, C = minute 475.0 to 476.0. Compare the EOG/eye-movement activity strength across A, B, and C. Report the segment labels in descending order of EOG activity strength, and state whether the strongest segment is REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the descending order of EOG activity strength across segments A, B, and C\n2) whether the strongest segment is REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"eog_activity_order_desc\" and \"strongest_segment_is_rem\".\n5. \"eog_activity_order_desc\" must be a JSON array with EXACTLY 3 elements.\n6. The array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. \"strongest_segment_is_rem\" must be exactly one of: \"yes\", \"no\", or null.\n8. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n9. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"eog_activity_order_desc\": [\"B\", \"A\", \"C\"], \"strongest_segment_is_rem\": \"yes\"}", + "metrics": [ + { + "metric_id": "eog_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "eog_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "C", + "A", + "B" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "strongest_segment_is_rem_accuracy", + "type": "categorical_check", + "target_key": "strongest_segment_is_rem", + "weight": 40, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case26/case26_24.json b/sleep_assessment/cases/case26/case26_24.json new file mode 100644 index 0000000000000000000000000000000000000000..91a0bd788369ec74e0ce20093b7d15b6c32fad94 --- /dev/null +++ b/sleep_assessment/cases/case26/case26_24.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 62.0 to 63.0, B = minute 268.0 to 269.0, C = minute 313.0 to 314.0. Compare the EOG/eye-movement activity strength across A, B, and C. Report the segment labels in descending order of EOG activity strength, and state whether the strongest segment is REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the descending order of EOG activity strength across segments A, B, and C\n2) whether the strongest segment is REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"eog_activity_order_desc\" and \"strongest_segment_is_rem\".\n5. \"eog_activity_order_desc\" must be a JSON array with EXACTLY 3 elements.\n6. The array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. \"strongest_segment_is_rem\" must be exactly one of: \"yes\", \"no\", or null.\n8. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n9. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"eog_activity_order_desc\": [\"B\", \"A\", \"C\"], \"strongest_segment_is_rem\": \"yes\"}", + "metrics": [ + { + "metric_id": "eog_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "eog_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "A", + "C", + "B" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "strongest_segment_is_rem_accuracy", + "type": "categorical_check", + "target_key": "strongest_segment_is_rem", + "weight": 40, + "params": { + "gt_value": "yes" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case26/case26_25.json b/sleep_assessment/cases/case26/case26_25.json new file mode 100644 index 0000000000000000000000000000000000000000..155525832cd604753c7da8b49ca8b60f84b831ff --- /dev/null +++ b/sleep_assessment/cases/case26/case26_25.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 389.5 to 390.5, B = minute 381.0 to 382.0, C = minute 206.5 to 207.5. Compare the EOG/eye-movement activity strength across A, B, and C. Report the segment labels in descending order of EOG activity strength, and state whether the strongest segment is REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent's report:\n1) the descending order of EOG activity strength across segments A, B, and C\n2) whether the strongest segment is REM sleep\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"eog_activity_order_desc\" and \"strongest_segment_is_rem\".\n5. \"eog_activity_order_desc\" must be a JSON array with EXACTLY 3 elements.\n6. The array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. \"strongest_segment_is_rem\" must be exactly one of: \"yes\", \"no\", or null.\n8. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n9. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"eog_activity_order_desc\": [\"B\", \"A\", \"C\"], \"strongest_segment_is_rem\": \"yes\"}", + "metrics": [ + { + "metric_id": "eog_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "eog_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "B", + "A", + "C" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "strongest_segment_is_rem_accuracy", + "type": "categorical_check", + "target_key": "strongest_segment_is_rem", + "weight": 40, + "params": { + "gt_value": "yes" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case27/case27_01.json b/sleep_assessment/cases/case27/case27_01.json new file mode 100644 index 0000000000000000000000000000000000000000..911060fa418b541ad0729d6cc125fdf3cef45b57 --- /dev/null +++ b/sleep_assessment/cases/case27/case27_01.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 274.5 to 275.5, B = minute 4.0 to 5.0, C = minute 194.5 to 195.5. For this file, X1, X3, and X4 are EMG channels. Compare the EMG activity strength across A, B, and C. Report the segment labels in descending order of EMG activity strength, and state whether any of the three segments is REM sleep. If REM sleep is present, identify which segment it is." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the descending EMG activity order across A/B/C and preserve the full report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"emg_activity_order_desc\" and \"full_text\".\n5. \"emg_activity_order_desc\" must be a JSON array with EXACTLY 3 elements or null.\n6. If present, the array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n8. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n\n### OUTPUT TEMPLATE\n{\"emg_activity_order_desc\": [\"B\", \"A\", \"C\"], \"full_text\": }", + "metrics": [ + { + "metric_id": "emg_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "emg_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "A", + "B", + "C" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "rem_presence_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. REM sleep is present among A/B/C, and the correct segment is C. The response is correct only if it states that REM sleep is present and identifies segment C as the REM segment. Mark the response incorrect if it says REM is present when it is absent, says REM is absent when it is present, identifies the wrong segment, identifies multiple segments, or gives an ambiguous conclusion about REM presence. Output only one valid JSON object: {\"status\": true} or {\"status\": false}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case27/case27_02.json b/sleep_assessment/cases/case27/case27_02.json new file mode 100644 index 0000000000000000000000000000000000000000..ef74b4b0998ece05ff27d243c950f79f5801e6d5 --- /dev/null +++ b/sleep_assessment/cases/case27/case27_02.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 239.0 to 240.0, B = minute 42.0 to 43.0, C = minute 36.0 to 37.0. For this file, X1, X3, and X4 are EMG channels. Compare the EMG activity strength across A, B, and C. Report the segment labels in descending order of EMG activity strength, and state whether any of the three segments is REM sleep. If REM sleep is present, identify which segment it is." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the descending EMG activity order across A/B/C and preserve the full report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"emg_activity_order_desc\" and \"full_text\".\n5. \"emg_activity_order_desc\" must be a JSON array with EXACTLY 3 elements or null.\n6. If present, the array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n8. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n\n### OUTPUT TEMPLATE\n{\"emg_activity_order_desc\": [\"B\", \"A\", \"C\"], \"full_text\": }", + "metrics": [ + { + "metric_id": "emg_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "emg_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "B", + "C", + "A" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "rem_presence_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. REM sleep is present among A/B/C, and the correct segment is A. The response is correct only if it states that REM sleep is present and identifies segment A as the REM segment. Mark the response incorrect if it says REM is present when it is absent, says REM is absent when it is present, identifies the wrong segment, identifies multiple segments, or gives an ambiguous conclusion about REM presence. Output only one valid JSON object: {\"status\": true} or {\"status\": false}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case27/case27_03.json b/sleep_assessment/cases/case27/case27_03.json new file mode 100644 index 0000000000000000000000000000000000000000..ddc38ca32f9be9cfe639440cc1e464fd92a132c0 --- /dev/null +++ b/sleep_assessment/cases/case27/case27_03.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 226.0 to 227.0, B = minute 169.5 to 170.5, C = minute 222.5 to 223.5. For this file, X1, X3, and X4 are EMG channels. Compare the EMG activity strength across A, B, and C. Report the segment labels in descending order of EMG activity strength, and state whether any of the three segments is REM sleep. If REM sleep is present, identify which segment it is." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the descending EMG activity order across A/B/C and preserve the full report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"emg_activity_order_desc\" and \"full_text\".\n5. \"emg_activity_order_desc\" must be a JSON array with EXACTLY 3 elements or null.\n6. If present, the array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n8. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n\n### OUTPUT TEMPLATE\n{\"emg_activity_order_desc\": [\"B\", \"A\", \"C\"], \"full_text\": }", + "metrics": [ + { + "metric_id": "emg_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "emg_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "C", + "A", + "B" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "rem_presence_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. REM sleep is present among A/B/C, and the correct segment is B. The response is correct only if it states that REM sleep is present and identifies segment B as the REM segment. Mark the response incorrect if it says REM is present when it is absent, says REM is absent when it is present, identifies the wrong segment, identifies multiple segments, or gives an ambiguous conclusion about REM presence. Output only one valid JSON object: {\"status\": true} or {\"status\": false}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case27/case27_04.json b/sleep_assessment/cases/case27/case27_04.json new file mode 100644 index 0000000000000000000000000000000000000000..69b60b90f9acd3490522cd9d6041853485453f60 --- /dev/null +++ b/sleep_assessment/cases/case27/case27_04.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "ISRUC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_04.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 258.0 to 259.0, B = minute 155.5 to 156.5, C = minute 3.0 to 4.0. For this file, X1, X3, and X4 are EMG channels. Compare the EMG activity strength across A, B, and C. Report the segment labels in descending order of EMG activity strength, and state whether any of the three segments is REM sleep. If REM sleep is present, identify which segment it is." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the descending EMG activity order across A/B/C and preserve the full report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"emg_activity_order_desc\" and \"full_text\".\n5. \"emg_activity_order_desc\" must be a JSON array with EXACTLY 3 elements or null.\n6. If present, the array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n8. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n\n### OUTPUT TEMPLATE\n{\"emg_activity_order_desc\": [\"B\", \"A\", \"C\"], \"full_text\": }", + "metrics": [ + { + "metric_id": "emg_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "emg_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "A", + "C", + "B" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "rem_presence_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. REM sleep is not present among A/B/C. The response is correct only if it states that REM sleep is absent or that none of the three segments is REM. Mark the response incorrect if it says REM is present when it is absent, says REM is absent when it is present, identifies the wrong segment, identifies multiple segments, or gives an ambiguous conclusion about REM presence. Output only one valid JSON object: {\"status\": true} or {\"status\": false}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case27/case27_05.json b/sleep_assessment/cases/case27/case27_05.json new file mode 100644 index 0000000000000000000000000000000000000000..35f0c0d8762d6f8a50405b430df086635bb4b972 --- /dev/null +++ b/sleep_assessment/cases/case27/case27_05.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 31.0 to 32.0, B = minute 28.0 to 29.0, C = minute 200.5 to 201.5. For this file, X1, X3, and X4 are EMG channels. Compare the EMG activity strength across A, B, and C. Report the segment labels in descending order of EMG activity strength, and state whether any of the three segments is REM sleep. If REM sleep is present, identify which segment it is." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the descending EMG activity order across A/B/C and preserve the full report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"emg_activity_order_desc\" and \"full_text\".\n5. \"emg_activity_order_desc\" must be a JSON array with EXACTLY 3 elements or null.\n6. If present, the array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n8. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n\n### OUTPUT TEMPLATE\n{\"emg_activity_order_desc\": [\"B\", \"A\", \"C\"], \"full_text\": }", + "metrics": [ + { + "metric_id": "emg_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "emg_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "B", + "A", + "C" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "rem_presence_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. REM sleep is not present among A/B/C. The response is correct only if it states that REM sleep is absent or that none of the three segments is REM. Mark the response incorrect if it says REM is present when it is absent, says REM is absent when it is present, identifies the wrong segment, identifies multiple segments, or gives an ambiguous conclusion about REM presence. Output only one valid JSON object: {\"status\": true} or {\"status\": false}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case27/case27_06.json b/sleep_assessment/cases/case27/case27_06.json new file mode 100644 index 0000000000000000000000000000000000000000..cbde80d1613bb06212a6b36e324e7a5ebff8769c --- /dev/null +++ b/sleep_assessment/cases/case27/case27_06.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "HMC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_01.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 189.5 to 190.5, B = minute 392.5 to 393.5, C = minute 73.0 to 74.0. Compare the EMG activity strength across A, B, and C. Report the segment labels in descending order of EMG activity strength, and state whether any of the three segments is REM sleep. If REM sleep is present, identify which segment it is." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the descending EMG activity order across A/B/C and preserve the full report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"emg_activity_order_desc\" and \"full_text\".\n5. \"emg_activity_order_desc\" must be a JSON array with EXACTLY 3 elements or null.\n6. If present, the array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n8. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n\n### OUTPUT TEMPLATE\n{\"emg_activity_order_desc\": [\"B\", \"A\", \"C\"], \"full_text\": }", + "metrics": [ + { + "metric_id": "emg_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "emg_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "A", + "B", + "C" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "rem_presence_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. REM sleep is not present among A/B/C. The response is correct only if it states that REM sleep is absent or that none of the three segments is REM. Mark the response incorrect if it says REM is present when it is absent, says REM is absent when it is present, identifies the wrong segment, identifies multiple segments, or gives an ambiguous conclusion about REM presence. Output only one valid JSON object: {\"status\": true} or {\"status\": false}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case27/case27_07.json b/sleep_assessment/cases/case27/case27_07.json new file mode 100644 index 0000000000000000000000000000000000000000..a077b0b0196ba87db8985650f3c659bed6d7e371 --- /dev/null +++ b/sleep_assessment/cases/case27/case27_07.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "HMC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_02.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 193.0 to 194.0, B = minute 91.5 to 92.5, C = minute 4.5 to 5.5. Compare the EMG activity strength across A, B, and C. Report the segment labels in descending order of EMG activity strength, and state whether any of the three segments is REM sleep. If REM sleep is present, identify which segment it is." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the descending EMG activity order across A/B/C and preserve the full report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"emg_activity_order_desc\" and \"full_text\".\n5. \"emg_activity_order_desc\" must be a JSON array with EXACTLY 3 elements or null.\n6. If present, the array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n8. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n\n### OUTPUT TEMPLATE\n{\"emg_activity_order_desc\": [\"B\", \"A\", \"C\"], \"full_text\": }", + "metrics": [ + { + "metric_id": "emg_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "emg_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "B", + "C", + "A" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "rem_presence_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. REM sleep is not present among A/B/C. The response is correct only if it states that REM sleep is absent or that none of the three segments is REM. Mark the response incorrect if it says REM is present when it is absent, says REM is absent when it is present, identifies the wrong segment, identifies multiple segments, or gives an ambiguous conclusion about REM presence. Output only one valid JSON object: {\"status\": true} or {\"status\": false}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case27/case27_08.json b/sleep_assessment/cases/case27/case27_08.json new file mode 100644 index 0000000000000000000000000000000000000000..8010e0ab0d0ff10389a89bf7df18561a9586a8c2 --- /dev/null +++ b/sleep_assessment/cases/case27/case27_08.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "HMC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_03.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 359.0 to 360.0, B = minute 294.0 to 295.0, C = minute 180.5 to 181.5. Compare the EMG activity strength across A, B, and C. Report the segment labels in descending order of EMG activity strength, and state whether any of the three segments is REM sleep. If REM sleep is present, identify which segment it is." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the descending EMG activity order across A/B/C and preserve the full report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"emg_activity_order_desc\" and \"full_text\".\n5. \"emg_activity_order_desc\" must be a JSON array with EXACTLY 3 elements or null.\n6. If present, the array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n8. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n\n### OUTPUT TEMPLATE\n{\"emg_activity_order_desc\": [\"B\", \"A\", \"C\"], \"full_text\": }", + "metrics": [ + { + "metric_id": "emg_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "emg_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "C", + "A", + "B" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "rem_presence_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. REM sleep is present among A/B/C, and the correct segment is B. The response is correct only if it states that REM sleep is present and identifies segment B as the REM segment. Mark the response incorrect if it says REM is present when it is absent, says REM is absent when it is present, identifies the wrong segment, identifies multiple segments, or gives an ambiguous conclusion about REM presence. Output only one valid JSON object: {\"status\": true} or {\"status\": false}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case27/case27_09.json b/sleep_assessment/cases/case27/case27_09.json new file mode 100644 index 0000000000000000000000000000000000000000..1a9c0fb23c5d9f0e7f741f50333bce454b748472 --- /dev/null +++ b/sleep_assessment/cases/case27/case27_09.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "HMC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_04.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 0.0 to 1.0, B = minute 454.0 to 455.0, C = minute 7.5 to 8.5. Compare the EMG activity strength across A, B, and C. Report the segment labels in descending order of EMG activity strength, and state whether any of the three segments is REM sleep. If REM sleep is present, identify which segment it is." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the descending EMG activity order across A/B/C and preserve the full report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"emg_activity_order_desc\" and \"full_text\".\n5. \"emg_activity_order_desc\" must be a JSON array with EXACTLY 3 elements or null.\n6. If present, the array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n8. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n\n### OUTPUT TEMPLATE\n{\"emg_activity_order_desc\": [\"B\", \"A\", \"C\"], \"full_text\": }", + "metrics": [ + { + "metric_id": "emg_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "emg_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "A", + "C", + "B" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "rem_presence_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. REM sleep is present among A/B/C, and the correct segment is B. The response is correct only if it states that REM sleep is present and identifies segment B as the REM segment. Mark the response incorrect if it says REM is present when it is absent, says REM is absent when it is present, identifies the wrong segment, identifies multiple segments, or gives an ambiguous conclusion about REM presence. Output only one valid JSON object: {\"status\": true} or {\"status\": false}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case27/case27_10.json b/sleep_assessment/cases/case27/case27_10.json new file mode 100644 index 0000000000000000000000000000000000000000..3a785bcbe0680be53191e8aefef69d54b6f224f3 --- /dev/null +++ b/sleep_assessment/cases/case27/case27_10.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "HMC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_05.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 332.5 to 333.5, B = minute 58.0 to 59.0, C = minute 396.0 to 397.0. Compare the EMG activity strength across A, B, and C. Report the segment labels in descending order of EMG activity strength, and state whether any of the three segments is REM sleep. If REM sleep is present, identify which segment it is." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the descending EMG activity order across A/B/C and preserve the full report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"emg_activity_order_desc\" and \"full_text\".\n5. \"emg_activity_order_desc\" must be a JSON array with EXACTLY 3 elements or null.\n6. If present, the array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n8. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n\n### OUTPUT TEMPLATE\n{\"emg_activity_order_desc\": [\"B\", \"A\", \"C\"], \"full_text\": }", + "metrics": [ + { + "metric_id": "emg_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "emg_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "B", + "A", + "C" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "rem_presence_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. REM sleep is present among A/B/C, and the correct segment is C. The response is correct only if it states that REM sleep is present and identifies segment C as the REM segment. Mark the response incorrect if it says REM is present when it is absent, says REM is absent when it is present, identifies the wrong segment, identifies multiple segments, or gives an ambiguous conclusion about REM presence. Output only one valid JSON object: {\"status\": true} or {\"status\": false}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case27/case27_11.json b/sleep_assessment/cases/case27/case27_11.json new file mode 100644 index 0000000000000000000000000000000000000000..f566e85454761ff60fcb74db0e76a1f06a037f90 --- /dev/null +++ b/sleep_assessment/cases/case27/case27_11.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 529.5 to 530.5, B = minute 172.0 to 173.0, C = minute 239.0 to 240.0. Compare the EMG activity strength across A, B, and C. Report the segment labels in descending order of EMG activity strength, and state whether any of the three segments is REM sleep. If REM sleep is present, identify which segment it is." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the descending EMG activity order across A/B/C and preserve the full report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"emg_activity_order_desc\" and \"full_text\".\n5. \"emg_activity_order_desc\" must be a JSON array with EXACTLY 3 elements or null.\n6. If present, the array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n8. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n\n### OUTPUT TEMPLATE\n{\"emg_activity_order_desc\": [\"B\", \"A\", \"C\"], \"full_text\": }", + "metrics": [ + { + "metric_id": "emg_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "emg_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "A", + "B", + "C" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "rem_presence_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. REM sleep is present among A/B/C, and the correct segment is C. The response is correct only if it states that REM sleep is present and identifies segment C as the REM segment. Mark the response incorrect if it says REM is present when it is absent, says REM is absent when it is present, identifies the wrong segment, identifies multiple segments, or gives an ambiguous conclusion about REM presence. Output only one valid JSON object: {\"status\": true} or {\"status\": false}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case27/case27_12.json b/sleep_assessment/cases/case27/case27_12.json new file mode 100644 index 0000000000000000000000000000000000000000..7cc7bdf56126a47e49b2e7daf1a07bb0d3705faa --- /dev/null +++ b/sleep_assessment/cases/case27/case27_12.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 537.5 to 538.5, B = minute 226.0 to 227.0, C = minute 161.0 to 162.0. Compare the EMG activity strength across A, B, and C. Report the segment labels in descending order of EMG activity strength, and state whether any of the three segments is REM sleep. If REM sleep is present, identify which segment it is." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the descending EMG activity order across A/B/C and preserve the full report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"emg_activity_order_desc\" and \"full_text\".\n5. \"emg_activity_order_desc\" must be a JSON array with EXACTLY 3 elements or null.\n6. If present, the array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n8. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n\n### OUTPUT TEMPLATE\n{\"emg_activity_order_desc\": [\"B\", \"A\", \"C\"], \"full_text\": }", + "metrics": [ + { + "metric_id": "emg_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "emg_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "B", + "C", + "A" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "rem_presence_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. REM sleep is not present among A/B/C. The response is correct only if it states that REM sleep is absent or that none of the three segments is REM. Mark the response incorrect if it says REM is present when it is absent, says REM is absent when it is present, identifies the wrong segment, identifies multiple segments, or gives an ambiguous conclusion about REM presence. Output only one valid JSON object: {\"status\": true} or {\"status\": false}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case27/case27_13.json b/sleep_assessment/cases/case27/case27_13.json new file mode 100644 index 0000000000000000000000000000000000000000..d46aa9ac14d499f4b581deb28a259b19ffdba7b2 --- /dev/null +++ b/sleep_assessment/cases/case27/case27_13.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 430.0 to 431.0, B = minute 445.0 to 446.0, C = minute 48.0 to 49.0. Compare the EMG activity strength across A, B, and C. Report the segment labels in descending order of EMG activity strength, and state whether any of the three segments is REM sleep. If REM sleep is present, identify which segment it is." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the descending EMG activity order across A/B/C and preserve the full report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"emg_activity_order_desc\" and \"full_text\".\n5. \"emg_activity_order_desc\" must be a JSON array with EXACTLY 3 elements or null.\n6. If present, the array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n8. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n\n### OUTPUT TEMPLATE\n{\"emg_activity_order_desc\": [\"B\", \"A\", \"C\"], \"full_text\": }", + "metrics": [ + { + "metric_id": "emg_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "emg_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "C", + "A", + "B" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "rem_presence_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. REM sleep is not present among A/B/C. The response is correct only if it states that REM sleep is absent or that none of the three segments is REM. Mark the response incorrect if it says REM is present when it is absent, says REM is absent when it is present, identifies the wrong segment, identifies multiple segments, or gives an ambiguous conclusion about REM presence. Output only one valid JSON object: {\"status\": true} or {\"status\": false}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case27/case27_14.json b/sleep_assessment/cases/case27/case27_14.json new file mode 100644 index 0000000000000000000000000000000000000000..75d21d99cf5b8084391314b9965e41fd8614ba09 --- /dev/null +++ b/sleep_assessment/cases/case27/case27_14.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 30.5 to 31.5, B = minute 153.5 to 154.5, C = minute 317.5 to 318.5. Compare the EMG activity strength across A, B, and C. Report the segment labels in descending order of EMG activity strength, and state whether any of the three segments is REM sleep. If REM sleep is present, identify which segment it is." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the descending EMG activity order across A/B/C and preserve the full report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"emg_activity_order_desc\" and \"full_text\".\n5. \"emg_activity_order_desc\" must be a JSON array with EXACTLY 3 elements or null.\n6. If present, the array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n8. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n\n### OUTPUT TEMPLATE\n{\"emg_activity_order_desc\": [\"B\", \"A\", \"C\"], \"full_text\": }", + "metrics": [ + { + "metric_id": "emg_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "emg_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "A", + "C", + "B" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "rem_presence_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. REM sleep is present among A/B/C, and the correct segment is B. The response is correct only if it states that REM sleep is present and identifies segment B as the REM segment. Mark the response incorrect if it says REM is present when it is absent, says REM is absent when it is present, identifies the wrong segment, identifies multiple segments, or gives an ambiguous conclusion about REM presence. Output only one valid JSON object: {\"status\": true} or {\"status\": false}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case27/case27_15.json b/sleep_assessment/cases/case27/case27_15.json new file mode 100644 index 0000000000000000000000000000000000000000..27b33bcd9d5e284035c832b7c837b8b6dc58d268 --- /dev/null +++ b/sleep_assessment/cases/case27/case27_15.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 353.5 to 354.5, B = minute 1.5 to 2.5, C = minute 306.5 to 307.5. Compare the EMG activity strength across A, B, and C. Report the segment labels in descending order of EMG activity strength, and state whether any of the three segments is REM sleep. If REM sleep is present, identify which segment it is." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the descending EMG activity order across A/B/C and preserve the full report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"emg_activity_order_desc\" and \"full_text\".\n5. \"emg_activity_order_desc\" must be a JSON array with EXACTLY 3 elements or null.\n6. If present, the array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n8. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n\n### OUTPUT TEMPLATE\n{\"emg_activity_order_desc\": [\"B\", \"A\", \"C\"], \"full_text\": }", + "metrics": [ + { + "metric_id": "emg_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "emg_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "B", + "A", + "C" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "rem_presence_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. REM sleep is present among A/B/C, and the correct segment is C. The response is correct only if it states that REM sleep is present and identifies segment C as the REM segment. Mark the response incorrect if it says REM is present when it is absent, says REM is absent when it is present, identifies the wrong segment, identifies multiple segments, or gives an ambiguous conclusion about REM presence. Output only one valid JSON object: {\"status\": true} or {\"status\": false}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case27/case27_16.json b/sleep_assessment/cases/case27/case27_16.json new file mode 100644 index 0000000000000000000000000000000000000000..cee5d9ba0bc89eb677b375c5fdc100b0e937b110 --- /dev/null +++ b/sleep_assessment/cases/case27/case27_16.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "MASSSS3_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_01.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 415.5 to 416.5, B = minute 450.0 to 451.0, C = minute 291.0 to 292.0. Compare the EMG activity strength across A, B, and C. Report the segment labels in descending order of EMG activity strength, and state whether any of the three segments is REM sleep. If REM sleep is present, identify which segment it is." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the descending EMG activity order across A/B/C and preserve the full report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"emg_activity_order_desc\" and \"full_text\".\n5. \"emg_activity_order_desc\" must be a JSON array with EXACTLY 3 elements or null.\n6. If present, the array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n8. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n\n### OUTPUT TEMPLATE\n{\"emg_activity_order_desc\": [\"B\", \"A\", \"C\"], \"full_text\": }", + "metrics": [ + { + "metric_id": "emg_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "emg_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "A", + "B", + "C" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "rem_presence_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. REM sleep is not present among A/B/C. The response is correct only if it states that REM sleep is absent or that none of the three segments is REM. Mark the response incorrect if it says REM is present when it is absent, says REM is absent when it is present, identifies the wrong segment, identifies multiple segments, or gives an ambiguous conclusion about REM presence. Output only one valid JSON object: {\"status\": true} or {\"status\": false}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case27/case27_17.json b/sleep_assessment/cases/case27/case27_17.json new file mode 100644 index 0000000000000000000000000000000000000000..543073ac98bd139193e15254f227ced2092cead5 --- /dev/null +++ b/sleep_assessment/cases/case27/case27_17.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "MASSSS3_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_02.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 477.0 to 478.0, B = minute 407.0 to 408.0, C = minute 0.5 to 1.5. Compare the EMG activity strength across A, B, and C. Report the segment labels in descending order of EMG activity strength, and state whether any of the three segments is REM sleep. If REM sleep is present, identify which segment it is." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the descending EMG activity order across A/B/C and preserve the full report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"emg_activity_order_desc\" and \"full_text\".\n5. \"emg_activity_order_desc\" must be a JSON array with EXACTLY 3 elements or null.\n6. If present, the array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n8. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n\n### OUTPUT TEMPLATE\n{\"emg_activity_order_desc\": [\"B\", \"A\", \"C\"], \"full_text\": }", + "metrics": [ + { + "metric_id": "emg_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "emg_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "B", + "C", + "A" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "rem_presence_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. REM sleep is present among A/B/C, and the correct segment is A. The response is correct only if it states that REM sleep is present and identifies segment A as the REM segment. Mark the response incorrect if it says REM is present when it is absent, says REM is absent when it is present, identifies the wrong segment, identifies multiple segments, or gives an ambiguous conclusion about REM presence. Output only one valid JSON object: {\"status\": true} or {\"status\": false}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case27/case27_18.json b/sleep_assessment/cases/case27/case27_18.json new file mode 100644 index 0000000000000000000000000000000000000000..841e770c9f1e9c9d111a7befacae9f4c6c9efca3 --- /dev/null +++ b/sleep_assessment/cases/case27/case27_18.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "MASSSS3_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_03.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 408.0 to 409.0, B = minute 365.0 to 366.0, C = minute 0.0 to 1.0. Compare the EMG activity strength across A, B, and C. Report the segment labels in descending order of EMG activity strength, and state whether any of the three segments is REM sleep. If REM sleep is present, identify which segment it is." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the descending EMG activity order across A/B/C and preserve the full report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"emg_activity_order_desc\" and \"full_text\".\n5. \"emg_activity_order_desc\" must be a JSON array with EXACTLY 3 elements or null.\n6. If present, the array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n8. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n\n### OUTPUT TEMPLATE\n{\"emg_activity_order_desc\": [\"B\", \"A\", \"C\"], \"full_text\": }", + "metrics": [ + { + "metric_id": "emg_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "emg_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "C", + "A", + "B" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "rem_presence_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. REM sleep is present among A/B/C, and the correct segment is B. The response is correct only if it states that REM sleep is present and identifies segment B as the REM segment. Mark the response incorrect if it says REM is present when it is absent, says REM is absent when it is present, identifies the wrong segment, identifies multiple segments, or gives an ambiguous conclusion about REM presence. Output only one valid JSON object: {\"status\": true} or {\"status\": false}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case27/case27_19.json b/sleep_assessment/cases/case27/case27_19.json new file mode 100644 index 0000000000000000000000000000000000000000..585cabbf1e17955e0ee26a322b3b060fc3cb0ac4 --- /dev/null +++ b/sleep_assessment/cases/case27/case27_19.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "MASSSS3_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_04.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 42.0 to 43.0, B = minute 406.5 to 407.5, C = minute 18.0 to 19.0. Compare the EMG activity strength across A, B, and C. Report the segment labels in descending order of EMG activity strength, and state whether any of the three segments is REM sleep. If REM sleep is present, identify which segment it is." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the descending EMG activity order across A/B/C and preserve the full report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"emg_activity_order_desc\" and \"full_text\".\n5. \"emg_activity_order_desc\" must be a JSON array with EXACTLY 3 elements or null.\n6. If present, the array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n8. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n\n### OUTPUT TEMPLATE\n{\"emg_activity_order_desc\": [\"B\", \"A\", \"C\"], \"full_text\": }", + "metrics": [ + { + "metric_id": "emg_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "emg_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "A", + "C", + "B" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "rem_presence_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. REM sleep is not present among A/B/C. The response is correct only if it states that REM sleep is absent or that none of the three segments is REM. Mark the response incorrect if it says REM is present when it is absent, says REM is absent when it is present, identifies the wrong segment, identifies multiple segments, or gives an ambiguous conclusion about REM presence. Output only one valid JSON object: {\"status\": true} or {\"status\": false}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case27/case27_20.json b/sleep_assessment/cases/case27/case27_20.json new file mode 100644 index 0000000000000000000000000000000000000000..b0ad4b30eb6fa385b39eb294f5cf3978c09c8b5f --- /dev/null +++ b/sleep_assessment/cases/case27/case27_20.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "MASSSS3_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_05.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 205.5 to 206.5, B = minute 392.0 to 393.0, C = minute 187.5 to 188.5. Compare the EMG activity strength across A, B, and C. Report the segment labels in descending order of EMG activity strength, and state whether any of the three segments is REM sleep. If REM sleep is present, identify which segment it is." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the descending EMG activity order across A/B/C and preserve the full report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"emg_activity_order_desc\" and \"full_text\".\n5. \"emg_activity_order_desc\" must be a JSON array with EXACTLY 3 elements or null.\n6. If present, the array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n8. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n\n### OUTPUT TEMPLATE\n{\"emg_activity_order_desc\": [\"B\", \"A\", \"C\"], \"full_text\": }", + "metrics": [ + { + "metric_id": "emg_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "emg_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "B", + "A", + "C" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "rem_presence_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. REM sleep is present among A/B/C, and the correct segment is C. The response is correct only if it states that REM sleep is present and identifies segment C as the REM segment. Mark the response incorrect if it says REM is present when it is absent, says REM is absent when it is present, identifies the wrong segment, identifies multiple segments, or gives an ambiguous conclusion about REM presence. Output only one valid JSON object: {\"status\": true} or {\"status\": false}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case27/case27_21.json b/sleep_assessment/cases/case27/case27_21.json new file mode 100644 index 0000000000000000000000000000000000000000..67619536540c9feaa9c57e2339c008cd06f98955 --- /dev/null +++ b/sleep_assessment/cases/case27/case27_21.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 0.0 to 1.0, B = minute 73.0 to 74.0, C = minute 237.5 to 238.5. Compare the EMG activity strength across A, B, and C. Report the segment labels in descending order of EMG activity strength, and state whether any of the three segments is REM sleep. If REM sleep is present, identify which segment it is." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the descending EMG activity order across A/B/C and preserve the full report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"emg_activity_order_desc\" and \"full_text\".\n5. \"emg_activity_order_desc\" must be a JSON array with EXACTLY 3 elements or null.\n6. If present, the array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n8. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n\n### OUTPUT TEMPLATE\n{\"emg_activity_order_desc\": [\"B\", \"A\", \"C\"], \"full_text\": }", + "metrics": [ + { + "metric_id": "emg_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "emg_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "A", + "B", + "C" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "rem_presence_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. REM sleep is present among A/B/C, and the correct segment is C. The response is correct only if it states that REM sleep is present and identifies segment C as the REM segment. Mark the response incorrect if it says REM is present when it is absent, says REM is absent when it is present, identifies the wrong segment, identifies multiple segments, or gives an ambiguous conclusion about REM presence. Output only one valid JSON object: {\"status\": true} or {\"status\": false}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case27/case27_22.json b/sleep_assessment/cases/case27/case27_22.json new file mode 100644 index 0000000000000000000000000000000000000000..5f6958c6781589f422080e1ebb1bfaa4f659c1fd --- /dev/null +++ b/sleep_assessment/cases/case27/case27_22.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 201.0 to 202.0, B = minute 212.5 to 213.5, C = minute 371.0 to 372.0. Compare the EMG activity strength across A, B, and C. Report the segment labels in descending order of EMG activity strength, and state whether any of the three segments is REM sleep. If REM sleep is present, identify which segment it is." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the descending EMG activity order across A/B/C and preserve the full report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"emg_activity_order_desc\" and \"full_text\".\n5. \"emg_activity_order_desc\" must be a JSON array with EXACTLY 3 elements or null.\n6. If present, the array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n8. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n\n### OUTPUT TEMPLATE\n{\"emg_activity_order_desc\": [\"B\", \"A\", \"C\"], \"full_text\": }", + "metrics": [ + { + "metric_id": "emg_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "emg_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "B", + "C", + "A" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "rem_presence_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. REM sleep is not present among A/B/C. The response is correct only if it states that REM sleep is absent or that none of the three segments is REM. Mark the response incorrect if it says REM is present when it is absent, says REM is absent when it is present, identifies the wrong segment, identifies multiple segments, or gives an ambiguous conclusion about REM presence. Output only one valid JSON object: {\"status\": true} or {\"status\": false}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case27/case27_23.json b/sleep_assessment/cases/case27/case27_23.json new file mode 100644 index 0000000000000000000000000000000000000000..2ee0d78f4e0b03788720650fbdc334ad666c88f1 --- /dev/null +++ b/sleep_assessment/cases/case27/case27_23.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "Physionet2018_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_03.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 448.0 to 449.0, B = minute 364.5 to 365.5, C = minute 137.0 to 138.0. Compare the EMG activity strength across A, B, and C. Report the segment labels in descending order of EMG activity strength, and state whether any of the three segments is REM sleep. If REM sleep is present, identify which segment it is." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the descending EMG activity order across A/B/C and preserve the full report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"emg_activity_order_desc\" and \"full_text\".\n5. \"emg_activity_order_desc\" must be a JSON array with EXACTLY 3 elements or null.\n6. If present, the array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n8. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n\n### OUTPUT TEMPLATE\n{\"emg_activity_order_desc\": [\"B\", \"A\", \"C\"], \"full_text\": }", + "metrics": [ + { + "metric_id": "emg_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "emg_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "C", + "A", + "B" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "rem_presence_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. REM sleep is not present among A/B/C. The response is correct only if it states that REM sleep is absent or that none of the three segments is REM. Mark the response incorrect if it says REM is present when it is absent, says REM is absent when it is present, identifies the wrong segment, identifies multiple segments, or gives an ambiguous conclusion about REM presence. Output only one valid JSON object: {\"status\": true} or {\"status\": false}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case27/case27_24.json b/sleep_assessment/cases/case27/case27_24.json new file mode 100644 index 0000000000000000000000000000000000000000..a0e07e227fd7796cbfd49701fe2873862936895b --- /dev/null +++ b/sleep_assessment/cases/case27/case27_24.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 309.0 to 310.0, B = minute 70.5 to 71.5, C = minute 293.5 to 294.5. Compare the EMG activity strength across A, B, and C. Report the segment labels in descending order of EMG activity strength, and state whether any of the three segments is REM sleep. If REM sleep is present, identify which segment it is." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the descending EMG activity order across A/B/C and preserve the full report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"emg_activity_order_desc\" and \"full_text\".\n5. \"emg_activity_order_desc\" must be a JSON array with EXACTLY 3 elements or null.\n6. If present, the array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n8. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n\n### OUTPUT TEMPLATE\n{\"emg_activity_order_desc\": [\"B\", \"A\", \"C\"], \"full_text\": }", + "metrics": [ + { + "metric_id": "emg_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "emg_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "A", + "C", + "B" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "rem_presence_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. REM sleep is present among A/B/C, and the correct segment is B. The response is correct only if it states that REM sleep is present and identifies segment B as the REM segment. Mark the response incorrect if it says REM is present when it is absent, says REM is absent when it is present, identifies the wrong segment, identifies multiple segments, or gives an ambiguous conclusion about REM presence. Output only one valid JSON object: {\"status\": true} or {\"status\": false}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case27/case27_25.json b/sleep_assessment/cases/case27/case27_25.json new file mode 100644 index 0000000000000000000000000000000000000000..002484fde57f6b3bac8800347dd64de481200764 --- /dev/null +++ b/sleep_assessment/cases/case27/case27_25.json @@ -0,0 +1,40 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "instruction": "Please analyze the provided sleep file and examine the following three 1-minute sleep segments: A = minute 296.5 to 297.5, B = minute 6.0 to 7.0, C = minute 397.5 to 398.5. Compare the EMG activity strength across A, B, and C. Report the segment labels in descending order of EMG activity strength, and state whether any of the three segments is REM sleep. If REM sleep is present, identify which segment it is." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep EMG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the descending EMG activity order across A/B/C and preserve the full report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text outside the JSON object.\n4. The keys must be EXACTLY \"emg_activity_order_desc\" and \"full_text\".\n5. \"emg_activity_order_desc\" must be a JSON array with EXACTLY 3 elements or null.\n6. If present, the array elements must be exactly \"A\", \"B\", and \"C\", without duplicates, in descending order.\n7. If the report gives an order like A>B>C or A, B, C, convert it to the JSON array form.\n8. \"full_text\" must be the complete original report text as a string. If no report text is available, use null.\n\n### OUTPUT TEMPLATE\n{\"emg_activity_order_desc\": [\"B\", \"A\", \"C\"], \"full_text\": }", + "metrics": [ + { + "metric_id": "emg_activity_order_accuracy", + "type": "sequence_match_check", + "target_key": "emg_activity_order_desc", + "weight": 60, + "params": { + "gt_value": [ + "B", + "A", + "C" + ], + "match_mode": "exact_order" + } + }, + { + "metric_id": "rem_presence_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 40, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. REM sleep is present among A/B/C, and the correct segment is C. The response is correct only if it states that REM sleep is present and identifies segment C as the REM segment. Mark the response incorrect if it says REM is present when it is absent, says REM is absent when it is present, identifies the wrong segment, identifies multiple segments, or gives an ambiguous conclusion about REM presence. Output only one valid JSON object: {\"status\": true} or {\"status\": false}. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case28/case28_01.json b/sleep_assessment/cases/case28/case28_01.json new file mode 100644 index 0000000000000000000000000000000000000000..4f9900e0bf8b98d734a4d2c4595234874357a747 --- /dev/null +++ b/sleep_assessment/cases/case28/case28_01.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "instruction": "Please use the provided sleep file to identify the ECG channel and estimate whole-night heart-rate statistics from the ECG signal. For this ISRUC file, X2 is the ECG channel. Report three final values: mean_HR_bpm, min_HR_bpm, and max_HR_bpm." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep ECG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final whole-night heart-rate statistics from the agent's report:\n1) mean_HR_bpm\n2) min_HR_bpm\n3) max_HR_bpm\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"mean_HR_bpm\", \"min_HR_bpm\", and \"max_HR_bpm\".\n5. All three values must be FLOAT bpm values or null. Do not include units.\n6. If the report contains units like bpm or beats/min, strip the units and keep only the number.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"mean_HR_bpm\": , \"min_HR_bpm\": , \"max_HR_bpm\": }", + "metrics": [ + { + "metric_id": "mean_hr_accuracy", + "type": "numeric_check", + "target_key": "mean_HR_bpm", + "weight": 34, + "params": { + "gt_value": 60.65368360694034, + "tolerance": 3.0 + } + }, + { + "metric_id": "min_hr_accuracy", + "type": "numeric_check", + "target_key": "min_HR_bpm", + "weight": 33, + "params": { + "gt_value": 51.06382978723404, + "tolerance": 5.0 + } + }, + { + "metric_id": "max_hr_accuracy", + "type": "numeric_check", + "target_key": "max_HR_bpm", + "weight": 33, + "params": { + "gt_value": 77.6707163803938, + "tolerance": 7.76707163803938 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case28/case28_02.json b/sleep_assessment/cases/case28/case28_02.json new file mode 100644 index 0000000000000000000000000000000000000000..68dab933e7285c6a3bc9a04bc645f91b8ce44b17 --- /dev/null +++ b/sleep_assessment/cases/case28/case28_02.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "instruction": "Please use the provided sleep file to identify the ECG channel and estimate whole-night heart-rate statistics from the ECG signal. For this ISRUC file, X2 is the ECG channel. Report three final values: mean_HR_bpm, min_HR_bpm, and max_HR_bpm." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep ECG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final whole-night heart-rate statistics from the agent's report:\n1) mean_HR_bpm\n2) min_HR_bpm\n3) max_HR_bpm\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"mean_HR_bpm\", \"min_HR_bpm\", and \"max_HR_bpm\".\n5. All three values must be FLOAT bpm values or null. Do not include units.\n6. If the report contains units like bpm or beats/min, strip the units and keep only the number.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"mean_HR_bpm\": , \"min_HR_bpm\": , \"max_HR_bpm\": }", + "metrics": [ + { + "metric_id": "mean_hr_accuracy", + "type": "numeric_check", + "target_key": "mean_HR_bpm", + "weight": 34, + "params": { + "gt_value": 64.83364781619402, + "tolerance": 3.0 + } + }, + { + "metric_id": "min_hr_accuracy", + "type": "numeric_check", + "target_key": "min_HR_bpm", + "weight": 33, + "params": { + "gt_value": 57.69230769230769, + "tolerance": 5.0 + } + }, + { + "metric_id": "max_hr_accuracy", + "type": "numeric_check", + "target_key": "max_HR_bpm", + "weight": 33, + "params": { + "gt_value": 76.43312101910827, + "tolerance": 7.643312101910827 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case28/case28_03.json b/sleep_assessment/cases/case28/case28_03.json new file mode 100644 index 0000000000000000000000000000000000000000..aee2b7e77f54d259787b4e0e1e06e9bb30138a55 --- /dev/null +++ b/sleep_assessment/cases/case28/case28_03.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "instruction": "Please use the provided sleep file to identify the ECG channel and estimate whole-night heart-rate statistics from the ECG signal. For this ISRUC file, X2 is the ECG channel. Report three final values: mean_HR_bpm, min_HR_bpm, and max_HR_bpm." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep ECG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final whole-night heart-rate statistics from the agent's report:\n1) mean_HR_bpm\n2) min_HR_bpm\n3) max_HR_bpm\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"mean_HR_bpm\", \"min_HR_bpm\", and \"max_HR_bpm\".\n5. All three values must be FLOAT bpm values or null. Do not include units.\n6. If the report contains units like bpm or beats/min, strip the units and keep only the number.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"mean_HR_bpm\": , \"min_HR_bpm\": , \"max_HR_bpm\": }", + "metrics": [ + { + "metric_id": "mean_hr_accuracy", + "type": "numeric_check", + "target_key": "mean_HR_bpm", + "weight": 34, + "params": { + "gt_value": 64.37798526448601, + "tolerance": 3.0 + } + }, + { + "metric_id": "min_hr_accuracy", + "type": "numeric_check", + "target_key": "min_HR_bpm", + "weight": 33, + "params": { + "gt_value": 51.724137931034484, + "tolerance": 5.0 + } + }, + { + "metric_id": "max_hr_accuracy", + "type": "numeric_check", + "target_key": "max_HR_bpm", + "weight": 33, + "params": { + "gt_value": 85.1063829787234, + "tolerance": 8.51063829787234 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case28/case28_04.json b/sleep_assessment/cases/case28/case28_04.json new file mode 100644 index 0000000000000000000000000000000000000000..2c918cdeac67f1af46614d9897610b71bd162a50 --- /dev/null +++ b/sleep_assessment/cases/case28/case28_04.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "ISRUC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_04.edf", + "instruction": "Please use the provided sleep file to identify the ECG channel and estimate whole-night heart-rate statistics from the ECG signal. For this ISRUC file, X2 is the ECG channel. Report three final values: mean_HR_bpm, min_HR_bpm, and max_HR_bpm." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep ECG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final whole-night heart-rate statistics from the agent's report:\n1) mean_HR_bpm\n2) min_HR_bpm\n3) max_HR_bpm\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"mean_HR_bpm\", \"min_HR_bpm\", and \"max_HR_bpm\".\n5. All three values must be FLOAT bpm values or null. Do not include units.\n6. If the report contains units like bpm or beats/min, strip the units and keep only the number.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"mean_HR_bpm\": , \"min_HR_bpm\": , \"max_HR_bpm\": }", + "metrics": [ + { + "metric_id": "mean_hr_accuracy", + "type": "numeric_check", + "target_key": "mean_HR_bpm", + "weight": 34, + "params": { + "gt_value": 49.89002583050498, + "tolerance": 3.0 + } + }, + { + "metric_id": "min_hr_accuracy", + "type": "numeric_check", + "target_key": "min_HR_bpm", + "weight": 33, + "params": { + "gt_value": 39.34426229508197, + "tolerance": 5.0 + } + }, + { + "metric_id": "max_hr_accuracy", + "type": "numeric_check", + "target_key": "max_HR_bpm", + "weight": 33, + "params": { + "gt_value": 76.43312101910827, + "tolerance": 7.643312101910827 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case28/case28_05.json b/sleep_assessment/cases/case28/case28_05.json new file mode 100644 index 0000000000000000000000000000000000000000..88acf40c575c2e26edd910494f48ebcacc9cd17d --- /dev/null +++ b/sleep_assessment/cases/case28/case28_05.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "instruction": "Please use the provided sleep file to identify the ECG channel and estimate whole-night heart-rate statistics from the ECG signal. For this ISRUC file, X2 is the ECG channel. Report three final values: mean_HR_bpm, min_HR_bpm, and max_HR_bpm." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep ECG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final whole-night heart-rate statistics from the agent's report:\n1) mean_HR_bpm\n2) min_HR_bpm\n3) max_HR_bpm\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"mean_HR_bpm\", \"min_HR_bpm\", and \"max_HR_bpm\".\n5. All three values must be FLOAT bpm values or null. Do not include units.\n6. If the report contains units like bpm or beats/min, strip the units and keep only the number.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"mean_HR_bpm\": , \"min_HR_bpm\": , \"max_HR_bpm\": }", + "metrics": [ + { + "metric_id": "mean_hr_accuracy", + "type": "numeric_check", + "target_key": "mean_HR_bpm", + "weight": 34, + "params": { + "gt_value": 50.72243613035606, + "tolerance": 3.0 + } + }, + { + "metric_id": "min_hr_accuracy", + "type": "numeric_check", + "target_key": "min_HR_bpm", + "weight": 33, + "params": { + "gt_value": 44.11764705882353, + "tolerance": 5.0 + } + }, + { + "metric_id": "max_hr_accuracy", + "type": "numeric_check", + "target_key": "max_HR_bpm", + "weight": 33, + "params": { + "gt_value": 81.08478290645117, + "tolerance": 8.108478290645117 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case28/case28_06.json b/sleep_assessment/cases/case28/case28_06.json new file mode 100644 index 0000000000000000000000000000000000000000..628a8fd7c71dd5b727316eb5d7e97570d935460d --- /dev/null +++ b/sleep_assessment/cases/case28/case28_06.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "HMC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_01.edf", + "instruction": "Please use the provided sleep file to identify the ECG channel and estimate whole-night heart-rate statistics from the ECG signal. Report three final values: mean_HR_bpm, min_HR_bpm, and max_HR_bpm." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep ECG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final whole-night heart-rate statistics from the agent's report:\n1) mean_HR_bpm\n2) min_HR_bpm\n3) max_HR_bpm\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"mean_HR_bpm\", \"min_HR_bpm\", and \"max_HR_bpm\".\n5. All three values must be FLOAT bpm values or null. Do not include units.\n6. If the report contains units like bpm or beats/min, strip the units and keep only the number.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"mean_HR_bpm\": , \"min_HR_bpm\": , \"max_HR_bpm\": }", + "metrics": [ + { + "metric_id": "mean_hr_accuracy", + "type": "numeric_check", + "target_key": "mean_HR_bpm", + "weight": 34, + "params": { + "gt_value": 88.42568986196196, + "tolerance": 3.0 + } + }, + { + "metric_id": "min_hr_accuracy", + "type": "numeric_check", + "target_key": "min_HR_bpm", + "weight": 33, + "params": { + "gt_value": 75.29411764705883, + "tolerance": 5.0 + } + }, + { + "metric_id": "max_hr_accuracy", + "type": "numeric_check", + "target_key": "max_HR_bpm", + "weight": 33, + "params": { + "gt_value": 131.28205128205127, + "tolerance": 13.128205128205128 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case28/case28_07.json b/sleep_assessment/cases/case28/case28_07.json new file mode 100644 index 0000000000000000000000000000000000000000..831940c217edb70bc876384a1402cee4908c62f0 --- /dev/null +++ b/sleep_assessment/cases/case28/case28_07.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "HMC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_02.edf", + "instruction": "Please use the provided sleep file to identify the ECG channel and estimate whole-night heart-rate statistics from the ECG signal. Report three final values: mean_HR_bpm, min_HR_bpm, and max_HR_bpm." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep ECG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final whole-night heart-rate statistics from the agent's report:\n1) mean_HR_bpm\n2) min_HR_bpm\n3) max_HR_bpm\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"mean_HR_bpm\", \"min_HR_bpm\", and \"max_HR_bpm\".\n5. All three values must be FLOAT bpm values or null. Do not include units.\n6. If the report contains units like bpm or beats/min, strip the units and keep only the number.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"mean_HR_bpm\": , \"min_HR_bpm\": , \"max_HR_bpm\": }", + "metrics": [ + { + "metric_id": "mean_hr_accuracy", + "type": "numeric_check", + "target_key": "mean_HR_bpm", + "weight": 34, + "params": { + "gt_value": 46.807686841605936, + "tolerance": 3.0 + } + }, + { + "metric_id": "min_hr_accuracy", + "type": "numeric_check", + "target_key": "min_HR_bpm", + "weight": 33, + "params": { + "gt_value": 41.6260162601626, + "tolerance": 5.0 + } + }, + { + "metric_id": "max_hr_accuracy", + "type": "numeric_check", + "target_key": "max_HR_bpm", + "weight": 33, + "params": { + "gt_value": 76.41791044776119, + "tolerance": 7.641791044776119 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case28/case28_08.json b/sleep_assessment/cases/case28/case28_08.json new file mode 100644 index 0000000000000000000000000000000000000000..c4656242d8387419a60e00db13fe896b3a464324 --- /dev/null +++ b/sleep_assessment/cases/case28/case28_08.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "HMC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_03.edf", + "instruction": "Please use the provided sleep file to identify the ECG channel and estimate whole-night heart-rate statistics from the ECG signal. Report three final values: mean_HR_bpm, min_HR_bpm, and max_HR_bpm." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep ECG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final whole-night heart-rate statistics from the agent's report:\n1) mean_HR_bpm\n2) min_HR_bpm\n3) max_HR_bpm\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"mean_HR_bpm\", \"min_HR_bpm\", and \"max_HR_bpm\".\n5. All three values must be FLOAT bpm values or null. Do not include units.\n6. If the report contains units like bpm or beats/min, strip the units and keep only the number.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"mean_HR_bpm\": , \"min_HR_bpm\": , \"max_HR_bpm\": }", + "metrics": [ + { + "metric_id": "mean_hr_accuracy", + "type": "numeric_check", + "target_key": "mean_HR_bpm", + "weight": 34, + "params": { + "gt_value": 58.02496343340899, + "tolerance": 3.0 + } + }, + { + "metric_id": "min_hr_accuracy", + "type": "numeric_check", + "target_key": "min_HR_bpm", + "weight": 33, + "params": { + "gt_value": 50.19607843137255, + "tolerance": 5.0 + } + }, + { + "metric_id": "max_hr_accuracy", + "type": "numeric_check", + "target_key": "max_HR_bpm", + "weight": 33, + "params": { + "gt_value": 100.3921568627451, + "tolerance": 10.03921568627451 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case28/case28_09.json b/sleep_assessment/cases/case28/case28_09.json new file mode 100644 index 0000000000000000000000000000000000000000..f67609e56dde2d94e369a25b4a77bab53f88e2ea --- /dev/null +++ b/sleep_assessment/cases/case28/case28_09.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "HMC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_04.edf", + "instruction": "Please use the provided sleep file to identify the ECG channel and estimate whole-night heart-rate statistics from the ECG signal. Report three final values: mean_HR_bpm, min_HR_bpm, and max_HR_bpm." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep ECG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final whole-night heart-rate statistics from the agent's report:\n1) mean_HR_bpm\n2) min_HR_bpm\n3) max_HR_bpm\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"mean_HR_bpm\", \"min_HR_bpm\", and \"max_HR_bpm\".\n5. All three values must be FLOAT bpm values or null. Do not include units.\n6. If the report contains units like bpm or beats/min, strip the units and keep only the number.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"mean_HR_bpm\": , \"min_HR_bpm\": , \"max_HR_bpm\": }", + "metrics": [ + { + "metric_id": "mean_hr_accuracy", + "type": "numeric_check", + "target_key": "mean_HR_bpm", + "weight": 34, + "params": { + "gt_value": 57.397448859366236, + "tolerance": 3.0 + } + }, + { + "metric_id": "min_hr_accuracy", + "type": "numeric_check", + "target_key": "min_HR_bpm", + "weight": 33, + "params": { + "gt_value": 49.38906752411575, + "tolerance": 5.0 + } + }, + { + "metric_id": "max_hr_accuracy", + "type": "numeric_check", + "target_key": "max_HR_bpm", + "weight": 33, + "params": { + "gt_value": 82.3598412972227, + "tolerance": 8.23598412972227 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case28/case28_10.json b/sleep_assessment/cases/case28/case28_10.json new file mode 100644 index 0000000000000000000000000000000000000000..bc7b70318e9bc7753c06dc801cfa7ccb2e7d8877 --- /dev/null +++ b/sleep_assessment/cases/case28/case28_10.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "HMC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "HMC" + }, + "agent_input": { + "data_path": "data/sleep/HMC_05.edf", + "instruction": "Please use the provided sleep file to identify the ECG channel and estimate whole-night heart-rate statistics from the ECG signal. Report three final values: mean_HR_bpm, min_HR_bpm, and max_HR_bpm." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep ECG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final whole-night heart-rate statistics from the agent's report:\n1) mean_HR_bpm\n2) min_HR_bpm\n3) max_HR_bpm\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"mean_HR_bpm\", \"min_HR_bpm\", and \"max_HR_bpm\".\n5. All three values must be FLOAT bpm values or null. Do not include units.\n6. If the report contains units like bpm or beats/min, strip the units and keep only the number.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"mean_HR_bpm\": , \"min_HR_bpm\": , \"max_HR_bpm\": }", + "metrics": [ + { + "metric_id": "mean_hr_accuracy", + "type": "numeric_check", + "target_key": "mean_HR_bpm", + "weight": 34, + "params": { + "gt_value": 62.33478589741015, + "tolerance": 3.0 + } + }, + { + "metric_id": "min_hr_accuracy", + "type": "numeric_check", + "target_key": "min_HR_bpm", + "weight": 33, + "params": { + "gt_value": 54.661921708185055, + "tolerance": 5.0 + } + }, + { + "metric_id": "max_hr_accuracy", + "type": "numeric_check", + "target_key": "max_HR_bpm", + "weight": 33, + "params": { + "gt_value": 101.72185430463576, + "tolerance": 10.172185430463577 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case28/case28_11.json b/sleep_assessment/cases/case28/case28_11.json new file mode 100644 index 0000000000000000000000000000000000000000..06aba6d79d3708925b694d34656e42774dd4baff --- /dev/null +++ b/sleep_assessment/cases/case28/case28_11.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "instruction": "Please use the provided sleep file to identify the ECG channel and estimate whole-night heart-rate statistics from the ECG signal. Report three final values: mean_HR_bpm, min_HR_bpm, and max_HR_bpm." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep ECG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final whole-night heart-rate statistics from the agent's report:\n1) mean_HR_bpm\n2) min_HR_bpm\n3) max_HR_bpm\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"mean_HR_bpm\", \"min_HR_bpm\", and \"max_HR_bpm\".\n5. All three values must be FLOAT bpm values or null. Do not include units.\n6. If the report contains units like bpm or beats/min, strip the units and keep only the number.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"mean_HR_bpm\": , \"min_HR_bpm\": , \"max_HR_bpm\": }", + "metrics": [ + { + "metric_id": "mean_hr_accuracy", + "type": "numeric_check", + "target_key": "mean_HR_bpm", + "weight": 34, + "params": { + "gt_value": 70.00431226388605, + "tolerance": 3.0 + } + }, + { + "metric_id": "min_hr_accuracy", + "type": "numeric_check", + "target_key": "min_HR_bpm", + "weight": 33, + "params": { + "gt_value": 59.28946381702288, + "tolerance": 5.0 + } + }, + { + "metric_id": "max_hr_accuracy", + "type": "numeric_check", + "target_key": "max_HR_bpm", + "weight": 33, + "params": { + "gt_value": 118.11755952380952, + "tolerance": 11.811755952380953 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case28/case28_12.json b/sleep_assessment/cases/case28/case28_12.json new file mode 100644 index 0000000000000000000000000000000000000000..5ab526f14b9e76090b6504e839b4cfbcbb783495 --- /dev/null +++ b/sleep_assessment/cases/case28/case28_12.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "instruction": "Please use the provided sleep file to identify the ECG channel and estimate whole-night heart-rate statistics from the ECG signal. Report three final values: mean_HR_bpm, min_HR_bpm, and max_HR_bpm." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep ECG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final whole-night heart-rate statistics from the agent's report:\n1) mean_HR_bpm\n2) min_HR_bpm\n3) max_HR_bpm\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"mean_HR_bpm\", \"min_HR_bpm\", and \"max_HR_bpm\".\n5. All three values must be FLOAT bpm values or null. Do not include units.\n6. If the report contains units like bpm or beats/min, strip the units and keep only the number.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"mean_HR_bpm\": , \"min_HR_bpm\": , \"max_HR_bpm\": }", + "metrics": [ + { + "metric_id": "mean_hr_accuracy", + "type": "numeric_check", + "target_key": "mean_HR_bpm", + "weight": 34, + "params": { + "gt_value": 106.28390622107275, + "tolerance": 3.0 + } + }, + { + "metric_id": "min_hr_accuracy", + "type": "numeric_check", + "target_key": "min_HR_bpm", + "weight": 33, + "params": { + "gt_value": 50.33557046979866, + "tolerance": 5.0 + } + }, + { + "metric_id": "max_hr_accuracy", + "type": "numeric_check", + "target_key": "max_HR_bpm", + "weight": 33, + "params": { + "gt_value": 192.30769230769232, + "tolerance": 19.230769230769234 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case28/case28_13.json b/sleep_assessment/cases/case28/case28_13.json new file mode 100644 index 0000000000000000000000000000000000000000..d73d56fa557edefa31404f4c4a5903cbf86acd94 --- /dev/null +++ b/sleep_assessment/cases/case28/case28_13.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "instruction": "Please use the provided sleep file to identify the ECG channel and estimate whole-night heart-rate statistics from the ECG signal. Report three final values: mean_HR_bpm, min_HR_bpm, and max_HR_bpm." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep ECG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final whole-night heart-rate statistics from the agent's report:\n1) mean_HR_bpm\n2) min_HR_bpm\n3) max_HR_bpm\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"mean_HR_bpm\", \"min_HR_bpm\", and \"max_HR_bpm\".\n5. All three values must be FLOAT bpm values or null. Do not include units.\n6. If the report contains units like bpm or beats/min, strip the units and keep only the number.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"mean_HR_bpm\": , \"min_HR_bpm\": , \"max_HR_bpm\": }", + "metrics": [ + { + "metric_id": "mean_hr_accuracy", + "type": "numeric_check", + "target_key": "mean_HR_bpm", + "weight": 34, + "params": { + "gt_value": 71.97590376954118, + "tolerance": 3.0 + } + }, + { + "metric_id": "min_hr_accuracy", + "type": "numeric_check", + "target_key": "min_HR_bpm", + "weight": 33, + "params": { + "gt_value": 60.97560975609756, + "tolerance": 5.0 + } + }, + { + "metric_id": "max_hr_accuracy", + "type": "numeric_check", + "target_key": "max_HR_bpm", + "weight": 33, + "params": { + "gt_value": 115.38461538461539, + "tolerance": 11.53846153846154 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case28/case28_14.json b/sleep_assessment/cases/case28/case28_14.json new file mode 100644 index 0000000000000000000000000000000000000000..2bd95b5060dfd10f1b52e864fe7d8f4c6ab7db84 --- /dev/null +++ b/sleep_assessment/cases/case28/case28_14.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "instruction": "Please use the provided sleep file to identify the ECG channel and estimate whole-night heart-rate statistics from the ECG signal. Report three final values: mean_HR_bpm, min_HR_bpm, and max_HR_bpm." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep ECG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final whole-night heart-rate statistics from the agent's report:\n1) mean_HR_bpm\n2) min_HR_bpm\n3) max_HR_bpm\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"mean_HR_bpm\", \"min_HR_bpm\", and \"max_HR_bpm\".\n5. All three values must be FLOAT bpm values or null. Do not include units.\n6. If the report contains units like bpm or beats/min, strip the units and keep only the number.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"mean_HR_bpm\": , \"min_HR_bpm\": , \"max_HR_bpm\": }", + "metrics": [ + { + "metric_id": "mean_hr_accuracy", + "type": "numeric_check", + "target_key": "mean_HR_bpm", + "weight": 34, + "params": { + "gt_value": 60.66056818445853, + "tolerance": 3.0 + } + }, + { + "metric_id": "min_hr_accuracy", + "type": "numeric_check", + "target_key": "min_HR_bpm", + "weight": 33, + "params": { + "gt_value": 52.81690140845071, + "tolerance": 5.0 + } + }, + { + "metric_id": "max_hr_accuracy", + "type": "numeric_check", + "target_key": "max_HR_bpm", + "weight": 33, + "params": { + "gt_value": 111.11720807726076, + "tolerance": 11.111720807726076 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case28/case28_15.json b/sleep_assessment/cases/case28/case28_15.json new file mode 100644 index 0000000000000000000000000000000000000000..864942296cd9a5c33741a7d12f2be51e48643296 --- /dev/null +++ b/sleep_assessment/cases/case28/case28_15.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "instruction": "Please use the provided sleep file to identify the ECG channel and estimate whole-night heart-rate statistics from the ECG signal. Report three final values: mean_HR_bpm, min_HR_bpm, and max_HR_bpm." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep ECG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final whole-night heart-rate statistics from the agent's report:\n1) mean_HR_bpm\n2) min_HR_bpm\n3) max_HR_bpm\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"mean_HR_bpm\", \"min_HR_bpm\", and \"max_HR_bpm\".\n5. All three values must be FLOAT bpm values or null. Do not include units.\n6. If the report contains units like bpm or beats/min, strip the units and keep only the number.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"mean_HR_bpm\": , \"min_HR_bpm\": , \"max_HR_bpm\": }", + "metrics": [ + { + "metric_id": "mean_hr_accuracy", + "type": "numeric_check", + "target_key": "mean_HR_bpm", + "weight": 34, + "params": { + "gt_value": 78.74478672433858, + "tolerance": 3.0 + } + }, + { + "metric_id": "min_hr_accuracy", + "type": "numeric_check", + "target_key": "min_HR_bpm", + "weight": 33, + "params": { + "gt_value": 67.56756756756756, + "tolerance": 5.0 + } + }, + { + "metric_id": "max_hr_accuracy", + "type": "numeric_check", + "target_key": "max_HR_bpm", + "weight": 33, + "params": { + "gt_value": 113.63636363636363, + "tolerance": 11.363636363636363 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case28/case28_16.json b/sleep_assessment/cases/case28/case28_16.json new file mode 100644 index 0000000000000000000000000000000000000000..8fe29207f2a9dc11d6202e044bbf9cd99faf6ae3 --- /dev/null +++ b/sleep_assessment/cases/case28/case28_16.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "MASSSS3_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_01.edf", + "instruction": "Please use the provided sleep file to identify the ECG channel and estimate whole-night heart-rate statistics from the ECG signal. Report three final values: mean_HR_bpm, min_HR_bpm, and max_HR_bpm." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep ECG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final whole-night heart-rate statistics from the agent's report:\n1) mean_HR_bpm\n2) min_HR_bpm\n3) max_HR_bpm\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"mean_HR_bpm\", \"min_HR_bpm\", and \"max_HR_bpm\".\n5. All three values must be FLOAT bpm values or null. Do not include units.\n6. If the report contains units like bpm or beats/min, strip the units and keep only the number.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"mean_HR_bpm\": , \"min_HR_bpm\": , \"max_HR_bpm\": }", + "metrics": [ + { + "metric_id": "mean_hr_accuracy", + "type": "numeric_check", + "target_key": "mean_HR_bpm", + "weight": 34, + "params": { + "gt_value": 58.403546855412756, + "tolerance": 3.0 + } + }, + { + "metric_id": "min_hr_accuracy", + "type": "numeric_check", + "target_key": "min_HR_bpm", + "weight": 33, + "params": { + "gt_value": 50.114325860637415, + "tolerance": 5.0 + } + }, + { + "metric_id": "max_hr_accuracy", + "type": "numeric_check", + "target_key": "max_HR_bpm", + "weight": 33, + "params": { + "gt_value": 74.02452619843925, + "tolerance": 7.402452619843925 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case28/case28_17.json b/sleep_assessment/cases/case28/case28_17.json new file mode 100644 index 0000000000000000000000000000000000000000..ef3d709bf7871c69127f5e3975debf3d60d453c5 --- /dev/null +++ b/sleep_assessment/cases/case28/case28_17.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "MASSSS3_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_02.edf", + "instruction": "Please use the provided sleep file to identify the ECG channel and estimate whole-night heart-rate statistics from the ECG signal. Report three final values: mean_HR_bpm, min_HR_bpm, and max_HR_bpm." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep ECG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final whole-night heart-rate statistics from the agent's report:\n1) mean_HR_bpm\n2) min_HR_bpm\n3) max_HR_bpm\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"mean_HR_bpm\", \"min_HR_bpm\", and \"max_HR_bpm\".\n5. All three values must be FLOAT bpm values or null. Do not include units.\n6. If the report contains units like bpm or beats/min, strip the units and keep only the number.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"mean_HR_bpm\": , \"min_HR_bpm\": , \"max_HR_bpm\": }", + "metrics": [ + { + "metric_id": "mean_hr_accuracy", + "type": "numeric_check", + "target_key": "mean_HR_bpm", + "weight": 34, + "params": { + "gt_value": 54.36498724436511, + "tolerance": 3.0 + } + }, + { + "metric_id": "min_hr_accuracy", + "type": "numeric_check", + "target_key": "min_HR_bpm", + "weight": 33, + "params": { + "gt_value": 45.714285714285715, + "tolerance": 5.0 + } + }, + { + "metric_id": "max_hr_accuracy", + "type": "numeric_check", + "target_key": "max_HR_bpm", + "weight": 33, + "params": { + "gt_value": 72.11267605633803, + "tolerance": 7.211267605633804 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case28/case28_18.json b/sleep_assessment/cases/case28/case28_18.json new file mode 100644 index 0000000000000000000000000000000000000000..ad3f39d696b26c0ca58554d03df99302d0993bd3 --- /dev/null +++ b/sleep_assessment/cases/case28/case28_18.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "MASSSS3_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_03.edf", + "instruction": "Please use the provided sleep file to identify the ECG channel and estimate whole-night heart-rate statistics from the ECG signal. Report three final values: mean_HR_bpm, min_HR_bpm, and max_HR_bpm." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep ECG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final whole-night heart-rate statistics from the agent's report:\n1) mean_HR_bpm\n2) min_HR_bpm\n3) max_HR_bpm\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"mean_HR_bpm\", \"min_HR_bpm\", and \"max_HR_bpm\".\n5. All three values must be FLOAT bpm values or null. Do not include units.\n6. If the report contains units like bpm or beats/min, strip the units and keep only the number.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"mean_HR_bpm\": , \"min_HR_bpm\": , \"max_HR_bpm\": }", + "metrics": [ + { + "metric_id": "mean_hr_accuracy", + "type": "numeric_check", + "target_key": "mean_HR_bpm", + "weight": 34, + "params": { + "gt_value": 68.07917612862003, + "tolerance": 3.0 + } + }, + { + "metric_id": "min_hr_accuracy", + "type": "numeric_check", + "target_key": "min_HR_bpm", + "weight": 33, + "params": { + "gt_value": 60.23529411764706, + "tolerance": 5.0 + } + }, + { + "metric_id": "max_hr_accuracy", + "type": "numeric_check", + "target_key": "max_HR_bpm", + "weight": 33, + "params": { + "gt_value": 90.3529411764706, + "tolerance": 9.03529411764706 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case28/case28_19.json b/sleep_assessment/cases/case28/case28_19.json new file mode 100644 index 0000000000000000000000000000000000000000..6aef3b5a5f1cf1cdb7a429a8f61b5dca321d524b --- /dev/null +++ b/sleep_assessment/cases/case28/case28_19.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "MASSSS3_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_04.edf", + "instruction": "Please use the provided sleep file to identify the ECG channel and estimate whole-night heart-rate statistics from the ECG signal. Report three final values: mean_HR_bpm, min_HR_bpm, and max_HR_bpm." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep ECG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final whole-night heart-rate statistics from the agent's report:\n1) mean_HR_bpm\n2) min_HR_bpm\n3) max_HR_bpm\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"mean_HR_bpm\", \"min_HR_bpm\", and \"max_HR_bpm\".\n5. All three values must be FLOAT bpm values or null. Do not include units.\n6. If the report contains units like bpm or beats/min, strip the units and keep only the number.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"mean_HR_bpm\": , \"min_HR_bpm\": , \"max_HR_bpm\": }", + "metrics": [ + { + "metric_id": "mean_hr_accuracy", + "type": "numeric_check", + "target_key": "mean_HR_bpm", + "weight": 34, + "params": { + "gt_value": 51.09511192507425, + "tolerance": 3.0 + } + }, + { + "metric_id": "min_hr_accuracy", + "type": "numeric_check", + "target_key": "min_HR_bpm", + "weight": 33, + "params": { + "gt_value": 45.309734513274336, + "tolerance": 5.0 + } + }, + { + "metric_id": "max_hr_accuracy", + "type": "numeric_check", + "target_key": "max_HR_bpm", + "weight": 33, + "params": { + "gt_value": 66.49350649350649, + "tolerance": 6.649350649350649 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case28/case28_20.json b/sleep_assessment/cases/case28/case28_20.json new file mode 100644 index 0000000000000000000000000000000000000000..a554050e381a4632e0e9cb6d2d06093e7e475b39 --- /dev/null +++ b/sleep_assessment/cases/case28/case28_20.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "MASSSS3_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "MASSSS3" + }, + "agent_input": { + "data_path": "data/sleep/MASSSS3_05.edf", + "instruction": "Please use the provided sleep file to identify the ECG channel and estimate whole-night heart-rate statistics from the ECG signal. Report three final values: mean_HR_bpm, min_HR_bpm, and max_HR_bpm." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep ECG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final whole-night heart-rate statistics from the agent's report:\n1) mean_HR_bpm\n2) min_HR_bpm\n3) max_HR_bpm\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"mean_HR_bpm\", \"min_HR_bpm\", and \"max_HR_bpm\".\n5. All three values must be FLOAT bpm values or null. Do not include units.\n6. If the report contains units like bpm or beats/min, strip the units and keep only the number.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"mean_HR_bpm\": , \"min_HR_bpm\": , \"max_HR_bpm\": }", + "metrics": [ + { + "metric_id": "mean_hr_accuracy", + "type": "numeric_check", + "target_key": "mean_HR_bpm", + "weight": 34, + "params": { + "gt_value": 61.374637268140745, + "tolerance": 3.0 + } + }, + { + "metric_id": "min_hr_accuracy", + "type": "numeric_check", + "target_key": "min_HR_bpm", + "weight": 33, + "params": { + "gt_value": 53.5191637630662, + "tolerance": 5.0 + } + }, + { + "metric_id": "max_hr_accuracy", + "type": "numeric_check", + "target_key": "max_HR_bpm", + "weight": 33, + "params": { + "gt_value": 84.86446886446886, + "tolerance": 8.486446886446886 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case28/case28_21.json b/sleep_assessment/cases/case28/case28_21.json new file mode 100644 index 0000000000000000000000000000000000000000..b43c2d58a0d53789312b4314b750de2756fcbe5d --- /dev/null +++ b/sleep_assessment/cases/case28/case28_21.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "instruction": "Please use the provided sleep file to identify the ECG channel and estimate whole-night heart-rate statistics from the ECG signal. Report three final values: mean_HR_bpm, min_HR_bpm, and max_HR_bpm." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep ECG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final whole-night heart-rate statistics from the agent's report:\n1) mean_HR_bpm\n2) min_HR_bpm\n3) max_HR_bpm\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"mean_HR_bpm\", \"min_HR_bpm\", and \"max_HR_bpm\".\n5. All three values must be FLOAT bpm values or null. Do not include units.\n6. If the report contains units like bpm or beats/min, strip the units and keep only the number.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"mean_HR_bpm\": , \"min_HR_bpm\": , \"max_HR_bpm\": }", + "metrics": [ + { + "metric_id": "mean_hr_accuracy", + "type": "numeric_check", + "target_key": "mean_HR_bpm", + "weight": 34, + "params": { + "gt_value": 63.89444633974605, + "tolerance": 3.0 + } + }, + { + "metric_id": "min_hr_accuracy", + "type": "numeric_check", + "target_key": "min_HR_bpm", + "weight": 33, + "params": { + "gt_value": 56.60377358490566, + "tolerance": 5.0 + } + }, + { + "metric_id": "max_hr_accuracy", + "type": "numeric_check", + "target_key": "max_HR_bpm", + "weight": 33, + "params": { + "gt_value": 77.1712158808933, + "tolerance": 7.717121588089331 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case28/case28_22.json b/sleep_assessment/cases/case28/case28_22.json new file mode 100644 index 0000000000000000000000000000000000000000..987cae5c40b40ca0e386f6f9e036b2e600f00b35 --- /dev/null +++ b/sleep_assessment/cases/case28/case28_22.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "instruction": "Please use the provided sleep file to identify the ECG channel and estimate whole-night heart-rate statistics from the ECG signal. Report three final values: mean_HR_bpm, min_HR_bpm, and max_HR_bpm." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep ECG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final whole-night heart-rate statistics from the agent's report:\n1) mean_HR_bpm\n2) min_HR_bpm\n3) max_HR_bpm\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"mean_HR_bpm\", \"min_HR_bpm\", and \"max_HR_bpm\".\n5. All three values must be FLOAT bpm values or null. Do not include units.\n6. If the report contains units like bpm or beats/min, strip the units and keep only the number.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"mean_HR_bpm\": , \"min_HR_bpm\": , \"max_HR_bpm\": }", + "metrics": [ + { + "metric_id": "mean_hr_accuracy", + "type": "numeric_check", + "target_key": "mean_HR_bpm", + "weight": 34, + "params": { + "gt_value": 46.92386963415311, + "tolerance": 3.0 + } + }, + { + "metric_id": "min_hr_accuracy", + "type": "numeric_check", + "target_key": "min_HR_bpm", + "weight": 33, + "params": { + "gt_value": 37.73584905660377, + "tolerance": 5.0 + } + }, + { + "metric_id": "max_hr_accuracy", + "type": "numeric_check", + "target_key": "max_HR_bpm", + "weight": 33, + "params": { + "gt_value": 72.29178007621121, + "tolerance": 7.229178007621122 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case28/case28_23.json b/sleep_assessment/cases/case28/case28_23.json new file mode 100644 index 0000000000000000000000000000000000000000..9aac6840b6bdc79777d9a08e024da22628dea552 --- /dev/null +++ b/sleep_assessment/cases/case28/case28_23.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "Physionet2018_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_03.edf", + "instruction": "Please use the provided sleep file to identify the ECG channel and estimate whole-night heart-rate statistics from the ECG signal. Report three final values: mean_HR_bpm, min_HR_bpm, and max_HR_bpm." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep ECG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final whole-night heart-rate statistics from the agent's report:\n1) mean_HR_bpm\n2) min_HR_bpm\n3) max_HR_bpm\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"mean_HR_bpm\", \"min_HR_bpm\", and \"max_HR_bpm\".\n5. All three values must be FLOAT bpm values or null. Do not include units.\n6. If the report contains units like bpm or beats/min, strip the units and keep only the number.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"mean_HR_bpm\": , \"min_HR_bpm\": , \"max_HR_bpm\": }", + "metrics": [ + { + "metric_id": "mean_hr_accuracy", + "type": "numeric_check", + "target_key": "mean_HR_bpm", + "weight": 34, + "params": { + "gt_value": 77.1206828253841, + "tolerance": 3.0 + } + }, + { + "metric_id": "min_hr_accuracy", + "type": "numeric_check", + "target_key": "min_HR_bpm", + "weight": 33, + "params": { + "gt_value": 71.00591715976331, + "tolerance": 5.0 + } + }, + { + "metric_id": "max_hr_accuracy", + "type": "numeric_check", + "target_key": "max_HR_bpm", + "weight": 33, + "params": { + "gt_value": 90.9090909090909, + "tolerance": 9.090909090909092 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case28/case28_24.json b/sleep_assessment/cases/case28/case28_24.json new file mode 100644 index 0000000000000000000000000000000000000000..0860369894b806d091bc8b599ecc5a124aeda884 --- /dev/null +++ b/sleep_assessment/cases/case28/case28_24.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "instruction": "Please use the provided sleep file to identify the ECG channel and estimate whole-night heart-rate statistics from the ECG signal. Report three final values: mean_HR_bpm, min_HR_bpm, and max_HR_bpm." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep ECG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final whole-night heart-rate statistics from the agent's report:\n1) mean_HR_bpm\n2) min_HR_bpm\n3) max_HR_bpm\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"mean_HR_bpm\", \"min_HR_bpm\", and \"max_HR_bpm\".\n5. All three values must be FLOAT bpm values or null. Do not include units.\n6. If the report contains units like bpm or beats/min, strip the units and keep only the number.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"mean_HR_bpm\": , \"min_HR_bpm\": , \"max_HR_bpm\": }", + "metrics": [ + { + "metric_id": "mean_hr_accuracy", + "type": "numeric_check", + "target_key": "mean_HR_bpm", + "weight": 34, + "params": { + "gt_value": 97.19501874017652, + "tolerance": 3.0 + } + }, + { + "metric_id": "min_hr_accuracy", + "type": "numeric_check", + "target_key": "min_HR_bpm", + "weight": 33, + "params": { + "gt_value": 89.55223880597015, + "tolerance": 5.0 + } + }, + { + "metric_id": "max_hr_accuracy", + "type": "numeric_check", + "target_key": "max_HR_bpm", + "weight": 33, + "params": { + "gt_value": 110.09174311926604, + "tolerance": 11.009174311926605 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case28/case28_25.json b/sleep_assessment/cases/case28/case28_25.json new file mode 100644 index 0000000000000000000000000000000000000000..957377e6acf10996a0499144b0bce02fd8d5b501 --- /dev/null +++ b/sleep_assessment/cases/case28/case28_25.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "instruction": "Please use the provided sleep file to identify the ECG channel and estimate whole-night heart-rate statistics from the ECG signal. Report three final values: mean_HR_bpm, min_HR_bpm, and max_HR_bpm." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep ECG analysis reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract three final whole-night heart-rate statistics from the agent's report:\n1) mean_HR_bpm\n2) min_HR_bpm\n3) max_HR_bpm\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"mean_HR_bpm\", \"min_HR_bpm\", and \"max_HR_bpm\".\n5. All three values must be FLOAT bpm values or null. Do not include units.\n6. If the report contains units like bpm or beats/min, strip the units and keep only the number.\n7. If one value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"mean_HR_bpm\": , \"min_HR_bpm\": , \"max_HR_bpm\": }", + "metrics": [ + { + "metric_id": "mean_hr_accuracy", + "type": "numeric_check", + "target_key": "mean_HR_bpm", + "weight": 34, + "params": { + "gt_value": 75.95668174580578, + "tolerance": 3.0 + } + }, + { + "metric_id": "min_hr_accuracy", + "type": "numeric_check", + "target_key": "min_HR_bpm", + "weight": 33, + "params": { + "gt_value": 67.0391061452514, + "tolerance": 5.0 + } + }, + { + "metric_id": "max_hr_accuracy", + "type": "numeric_check", + "target_key": "max_HR_bpm", + "weight": 33, + "params": { + "gt_value": 95.23809523809524, + "tolerance": 9.523809523809524 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_01.json b/sleep_assessment/cases/case29/case29_01.json new file mode 100644 index 0000000000000000000000000000000000000000..d9c063ae1dc90f63d85dbc41200154606f9c6029 --- /dev/null +++ b/sleep_assessment/cases/case29/case29_01.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_01_ISRUC_01", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 0.0 to 1.0, B = minute 28.0 to 29.0, C = minute 66.0 to 67.0, D = minute 94.0 to 95.0, E = minute 123.0 to 124.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_02.json b/sleep_assessment/cases/case29/case29_02.json new file mode 100644 index 0000000000000000000000000000000000000000..6467aee4f0841200d5809452615204edf2b6d729 --- /dev/null +++ b/sleep_assessment/cases/case29/case29_02.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_02_ISRUC_01", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 151.0 to 152.0, B = minute 188.0 to 189.0, C = minute 33.25 to 34.25, D = minute 234.0 to 235.0, E = minute 281.0 to 282.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_03.json b/sleep_assessment/cases/case29/case29_03.json new file mode 100644 index 0000000000000000000000000000000000000000..e8f4487935e47fbf14b3c74cbc4ad579dbfa9a9b --- /dev/null +++ b/sleep_assessment/cases/case29/case29_03.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_03_ISRUC_01", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 177.75 to 178.75, B = minute 309.0 to 310.0, C = minute 349.0 to 350.0, D = minute 377.0 to 378.0, E = minute 226.75 to 227.75. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "yes" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_04.json b/sleep_assessment/cases/case29/case29_04.json new file mode 100644 index 0000000000000000000000000000000000000000..e345703e937ef1151b4c903757acbcd4ae4aa188 --- /dev/null +++ b/sleep_assessment/cases/case29/case29_04.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_04_ISRUC_01", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 406.0 to 407.0, B = minute 246.75 to 247.75, C = minute 322.25 to 323.25, D = minute 438.75 to 439.75, E = minute 438.0 to 439.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_05.json b/sleep_assessment/cases/case29/case29_05.json new file mode 100644 index 0000000000000000000000000000000000000000..e60002e2e03b1edbc6ca136e1c0f9965e4b1a6bc --- /dev/null +++ b/sleep_assessment/cases/case29/case29_05.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_05_ISRUC_02", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 0.0 to 1.0, B = minute 31.0 to 32.0, C = minute 63.0 to 64.0, D = minute 106.0 to 107.0, E = minute 159.0 to 160.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_06.json b/sleep_assessment/cases/case29/case29_06.json new file mode 100644 index 0000000000000000000000000000000000000000..c3fbb88f716766df8c6b1898cbe4e419d14c59cf --- /dev/null +++ b/sleep_assessment/cases/case29/case29_06.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_06_ISRUC_02", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 194.0 to 195.0, B = minute 230.0 to 231.0, C = minute 87.25 to 88.25, D = minute 262.0 to 263.0, E = minute 295.0 to 296.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_07.json b/sleep_assessment/cases/case29/case29_07.json new file mode 100644 index 0000000000000000000000000000000000000000..b89795cc836a2c64250b2086c9c23e52626658a8 --- /dev/null +++ b/sleep_assessment/cases/case29/case29_07.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_07_ISRUC_02", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 123.75 to 124.75, B = minute 328.0 to 329.0, C = minute 369.0 to 370.0, D = minute 408.0 to 409.0, E = minute 147.75 to 148.75. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "yes" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_08.json b/sleep_assessment/cases/case29/case29_08.json new file mode 100644 index 0000000000000000000000000000000000000000..e533eb3fdb3c8f48e945c0d3fd32477e799d46f5 --- /dev/null +++ b/sleep_assessment/cases/case29/case29_08.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_08_ISRUC_02", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 442.0 to 443.0, B = minute 319.25 to 320.25, C = minute 379.75 to 380.75, D = minute 481.0 to 482.0, E = minute 479.0 to 480.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_09.json b/sleep_assessment/cases/case29/case29_09.json new file mode 100644 index 0000000000000000000000000000000000000000..96e7d1438d38e8f6f5ef99714f6a91d6d3a1c3d7 --- /dev/null +++ b/sleep_assessment/cases/case29/case29_09.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_09_ISRUC_03", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 0.0 to 1.0, B = minute 45.0 to 46.0, C = minute 82.0 to 83.0, D = minute 119.0 to 120.0, E = minute 149.0 to 150.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_10.json b/sleep_assessment/cases/case29/case29_10.json new file mode 100644 index 0000000000000000000000000000000000000000..54e31b5e5b7beff6abf29c18452390f9d51aba12 --- /dev/null +++ b/sleep_assessment/cases/case29/case29_10.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_10_ISRUC_03", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 187.0 to 188.0, B = minute 219.0 to 220.0, C = minute 5.25 to 6.25, D = minute 253.0 to 254.0, E = minute 287.0 to 288.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_11.json b/sleep_assessment/cases/case29/case29_11.json new file mode 100644 index 0000000000000000000000000000000000000000..955c88eb582b2e72833faa003da1cce80fc910fe --- /dev/null +++ b/sleep_assessment/cases/case29/case29_11.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_11_ISRUC_03", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 62.75 to 63.75, B = minute 326.0 to 327.0, C = minute 370.0 to 371.0, D = minute 405.0 to 406.0, E = minute 170.75 to 171.75. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "yes" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_12.json b/sleep_assessment/cases/case29/case29_12.json new file mode 100644 index 0000000000000000000000000000000000000000..2497e2c68a8ead43364972b725e71605922b62d1 --- /dev/null +++ b/sleep_assessment/cases/case29/case29_12.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_12_ISRUC_03", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 435.0 to 436.0, B = minute 273.75 to 274.75, C = minute 345.25 to 346.25, D = minute 470.5 to 471.5, E = minute 469.0 to 470.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_13.json b/sleep_assessment/cases/case29/case29_13.json new file mode 100644 index 0000000000000000000000000000000000000000..f5f69dfbeee62a0474b4e3a5abf042fd66da77af --- /dev/null +++ b/sleep_assessment/cases/case29/case29_13.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_13_ISRUC_04", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_04.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 0.0 to 1.0, B = minute 40.0 to 41.0, C = minute 77.0 to 78.0, D = minute 113.0 to 114.0, E = minute 149.0 to 150.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_14.json b/sleep_assessment/cases/case29/case29_14.json new file mode 100644 index 0000000000000000000000000000000000000000..7e6f0488bfbe9892c2a373f559324de82e326a70 --- /dev/null +++ b/sleep_assessment/cases/case29/case29_14.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_14_ISRUC_04", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_04.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 187.0 to 188.0, B = minute 224.0 to 225.0, C = minute 2.75 to 3.75, D = minute 260.0 to 261.0, E = minute 297.0 to 298.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_15.json b/sleep_assessment/cases/case29/case29_15.json new file mode 100644 index 0000000000000000000000000000000000000000..e27ea2f10b44e3c31f37fd54f59744576c2ab6d5 --- /dev/null +++ b/sleep_assessment/cases/case29/case29_15.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_15_ISRUC_04", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_04.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 16.25 to 17.25, B = minute 334.0 to 335.0, C = minute 372.0 to 373.0, D = minute 408.0 to 409.0, E = minute 45.25 to 46.25. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "yes" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_16.json b/sleep_assessment/cases/case29/case29_16.json new file mode 100644 index 0000000000000000000000000000000000000000..6eb7729acdd3fa39aff764c4d0f3d4e66744049d --- /dev/null +++ b/sleep_assessment/cases/case29/case29_16.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_16_ISRUC_04", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_04.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 444.0 to 445.0, B = minute 199.25 to 200.25, C = minute 298.25 to 299.25, D = minute 345.75 to 346.75, E = minute 480.0 to 481.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_17.json b/sleep_assessment/cases/case29/case29_17.json new file mode 100644 index 0000000000000000000000000000000000000000..1c94298eb241a3ca1975492326f9d830e78b4eef --- /dev/null +++ b/sleep_assessment/cases/case29/case29_17.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_17_ISRUC_05", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 0.0 to 1.0, B = minute 29.0 to 30.0, C = minute 58.0 to 59.0, D = minute 92.0 to 93.0, E = minute 122.0 to 123.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_18.json b/sleep_assessment/cases/case29/case29_18.json new file mode 100644 index 0000000000000000000000000000000000000000..eddda02cfcf77d113d61308392043c37f863c575 --- /dev/null +++ b/sleep_assessment/cases/case29/case29_18.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_18_ISRUC_05", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 156.0 to 157.0, B = minute 185.0 to 186.0, C = minute 75.25 to 76.25, D = minute 218.0 to 219.0, E = minute 252.0 to 253.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_19.json b/sleep_assessment/cases/case29/case29_19.json new file mode 100644 index 0000000000000000000000000000000000000000..9b09d7f6df17ee9b5b9191b14e409e03ea0167c9 --- /dev/null +++ b/sleep_assessment/cases/case29/case29_19.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_19_ISRUC_05", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 206.75 to 207.75, B = minute 286.0 to 287.0, C = minute 336.0 to 337.0, D = minute 377.0 to 378.0, E = minute 268.75 to 269.75. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "yes" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_20.json b/sleep_assessment/cases/case29/case29_20.json new file mode 100644 index 0000000000000000000000000000000000000000..361fb9716d54ec917e7b32f4ed3d8f77ea0ea486 --- /dev/null +++ b/sleep_assessment/cases/case29/case29_20.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_20_ISRUC_05", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 407.0 to 408.0, B = minute 308.75 to 309.75, C = minute 342.25 to 343.25, D = minute 403.25 to 404.25, E = minute 436.0 to 437.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_21.json b/sleep_assessment/cases/case29/case29_21.json new file mode 100644 index 0000000000000000000000000000000000000000..0d99e278c0da20667b5fe0c4f3519f3bff162227 --- /dev/null +++ b/sleep_assessment/cases/case29/case29_21.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_21_SHHS1_01", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 0.0 to 1.0, B = minute 35.0 to 36.0, C = minute 97.0 to 98.0, D = minute 157.0 to 158.0, E = minute 188.0 to 189.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_22.json b/sleep_assessment/cases/case29/case29_22.json new file mode 100644 index 0000000000000000000000000000000000000000..5d8152b86370f18419fc6aa507ad46399681ebd7 --- /dev/null +++ b/sleep_assessment/cases/case29/case29_22.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_22_SHHS1_01", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 216.0 to 217.0, B = minute 259.0 to 260.0, C = minute 15.79 to 16.79, D = minute 291.0 to 292.0, E = minute 343.0 to 344.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_23.json b/sleep_assessment/cases/case29/case29_23.json new file mode 100644 index 0000000000000000000000000000000000000000..2139f25b502bc9fb9eff97f1fe2311a2513c3ef7 --- /dev/null +++ b/sleep_assessment/cases/case29/case29_23.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_23_SHHS1_01", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 88.3 to 89.3, B = minute 388.0 to 389.0, C = minute 456.0 to 457.0, D = minute 484.0 to 485.0, E = minute 141.58 to 142.58. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "yes" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_24.json b/sleep_assessment/cases/case29/case29_24.json new file mode 100644 index 0000000000000000000000000000000000000000..e7b6cfb8603d1fe77c2114f472869a35e02718e3 --- /dev/null +++ b/sleep_assessment/cases/case29/case29_24.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_24_SHHS1_01", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 513.0 to 514.0, B = minute 313.53 to 314.53, C = minute 400.86 to 401.86, D = minute 454.28 to 455.28, E = minute 541.0 to 542.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_25.json b/sleep_assessment/cases/case29/case29_25.json new file mode 100644 index 0000000000000000000000000000000000000000..f789c8ad0892a0639f64ef3a3c6e1d6d7a19a02b --- /dev/null +++ b/sleep_assessment/cases/case29/case29_25.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_25_SHHS1_02", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 0.0 to 1.0, B = minute 41.0 to 42.0, C = minute 102.0 to 103.0, D = minute 147.0 to 148.0, E = minute 207.0 to 208.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_26.json b/sleep_assessment/cases/case29/case29_26.json new file mode 100644 index 0000000000000000000000000000000000000000..06cf43db1d36e38a0d36110dafe4daa9ca9517d5 --- /dev/null +++ b/sleep_assessment/cases/case29/case29_26.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_26_SHHS1_02", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 257.0 to 258.0, B = minute 292.0 to 293.0, C = minute 3.44 to 4.44, D = minute 328.0 to 329.0, E = minute 363.0 to 364.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_27.json b/sleep_assessment/cases/case29/case29_27.json new file mode 100644 index 0000000000000000000000000000000000000000..1674b4abdd06e555badd2d369db236d0c8d5899a --- /dev/null +++ b/sleep_assessment/cases/case29/case29_27.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_27_SHHS1_02", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 70.53 to 71.53, B = minute 398.0 to 399.0, C = minute 433.0 to 434.0, D = minute 468.0 to 469.0, E = minute 123.24 to 124.24. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "yes" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_28.json b/sleep_assessment/cases/case29/case29_28.json new file mode 100644 index 0000000000000000000000000000000000000000..7212951e29ba3c6317cfafb8dda7ddf1733441fb --- /dev/null +++ b/sleep_assessment/cases/case29/case29_28.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_28_SHHS1_02", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 503.0 to 504.0, B = minute 170.3 to 171.3, C = minute 200.78 to 201.78, D = minute 224.79 to 225.79, E = minute 538.0 to 539.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_29.json b/sleep_assessment/cases/case29/case29_29.json new file mode 100644 index 0000000000000000000000000000000000000000..411bc515088c7402a274a078ed6900653694c4e4 --- /dev/null +++ b/sleep_assessment/cases/case29/case29_29.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_29_SHHS1_03", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 0.0 to 1.0, B = minute 30.0 to 31.0, C = minute 60.0 to 61.0, D = minute 92.0 to 93.0, E = minute 137.0 to 138.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_30.json b/sleep_assessment/cases/case29/case29_30.json new file mode 100644 index 0000000000000000000000000000000000000000..b1b070cf3d6037665f4886570ae7d49d06d5e2de --- /dev/null +++ b/sleep_assessment/cases/case29/case29_30.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_30_SHHS1_03", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 179.0 to 180.0, B = minute 239.0 to 240.0, C = minute 83.56 to 84.56, D = minute 281.0 to 282.0, E = minute 324.0 to 325.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_31.json b/sleep_assessment/cases/case29/case29_31.json new file mode 100644 index 0000000000000000000000000000000000000000..2bd18184e879bf8a30c492add601cfe614f250ba --- /dev/null +++ b/sleep_assessment/cases/case29/case29_31.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_31_SHHS1_03", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 168.06 to 169.06, B = minute 355.0 to 356.0, C = minute 393.0 to 394.0, D = minute 453.0 to 454.0, E = minute 226.73 to 227.73. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "yes" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_32.json b/sleep_assessment/cases/case29/case29_32.json new file mode 100644 index 0000000000000000000000000000000000000000..bc1853c06fade2868ba3aa0e5c5a077a62743b0e --- /dev/null +++ b/sleep_assessment/cases/case29/case29_32.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_32_SHHS1_03", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 488.0 to 489.0, B = minute 302.65 to 303.65, C = minute 413.82 to 414.82, D = minute 514.6 to 515.6, E = minute 523.0 to 524.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_33.json b/sleep_assessment/cases/case29/case29_33.json new file mode 100644 index 0000000000000000000000000000000000000000..458762bc25337ca3d6a487de12356c2742f7b020 --- /dev/null +++ b/sleep_assessment/cases/case29/case29_33.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_33_SHHS1_04", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 0.0 to 1.0, B = minute 30.0 to 31.0, C = minute 63.0 to 64.0, D = minute 94.0 to 95.0, E = minute 126.0 to 127.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_34.json b/sleep_assessment/cases/case29/case29_34.json new file mode 100644 index 0000000000000000000000000000000000000000..0f90d538579830338df7f7c689b5669afddb8587 --- /dev/null +++ b/sleep_assessment/cases/case29/case29_34.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_34_SHHS1_04", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 158.0 to 159.0, B = minute 193.0 to 194.0, C = minute 42.13 to 43.13, D = minute 224.0 to 225.0, E = minute 259.0 to 260.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_35.json b/sleep_assessment/cases/case29/case29_35.json new file mode 100644 index 0000000000000000000000000000000000000000..4a92cbd2e50ec9b90d20f7e3e8b19944f8d2a3a5 --- /dev/null +++ b/sleep_assessment/cases/case29/case29_35.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_35_SHHS1_04", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 172.91 to 173.91, B = minute 295.0 to 296.0, C = minute 335.0 to 336.0, D = minute 365.0 to 366.0, E = minute 249.01 to 250.01. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "yes" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_36.json b/sleep_assessment/cases/case29/case29_36.json new file mode 100644 index 0000000000000000000000000000000000000000..f54cddc180454f34e2e2b5b9e63f48e1c4256b97 --- /dev/null +++ b/sleep_assessment/cases/case29/case29_36.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_36_SHHS1_04", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 406.0 to 407.0, B = minute 297.65 to 298.65, C = minute 365.86 to 366.86, D = minute 385.14 to 386.14, E = minute 436.0 to 437.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_37.json b/sleep_assessment/cases/case29/case29_37.json new file mode 100644 index 0000000000000000000000000000000000000000..011a4a9a28b6a884c67bf5c26029a84ebe5f0eb6 --- /dev/null +++ b/sleep_assessment/cases/case29/case29_37.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_37_SHHS1_05", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 0.0 to 1.0, B = minute 33.0 to 34.0, C = minute 73.0 to 74.0, D = minute 108.0 to 109.0, E = minute 142.0 to 143.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_38.json b/sleep_assessment/cases/case29/case29_38.json new file mode 100644 index 0000000000000000000000000000000000000000..6bb0c0456254539693e3a27d9c0b63fdd4186b86 --- /dev/null +++ b/sleep_assessment/cases/case29/case29_38.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_38_SHHS1_05", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 194.0 to 195.0, B = minute 236.0 to 237.0, C = minute 10.81 to 11.81, D = minute 283.0 to 284.0, E = minute 336.0 to 337.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_39.json b/sleep_assessment/cases/case29/case29_39.json new file mode 100644 index 0000000000000000000000000000000000000000..d5a462e9b13172c7c6c06a9cc382d1ae6930bd15 --- /dev/null +++ b/sleep_assessment/cases/case29/case29_39.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_39_SHHS1_05", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 159.44 to 160.44, B = minute 369.0 to 370.0, C = minute 416.0 to 417.0, D = minute 469.0 to 470.0, E = minute 251.2 to 252.2. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "yes" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case29/case29_40.json b/sleep_assessment/cases/case29/case29_40.json new file mode 100644 index 0000000000000000000000000000000000000000..ef8bebc423ba0ca4ae434fcb37e100bd443f3cd0 --- /dev/null +++ b/sleep_assessment/cases/case29/case29_40.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "case29_40_SHHS1_05", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "instruction": "Please analyze the provided sleep file and determine whether arousal occurs in each of the following five 60-second sleep segments: A = minute 510.0 to 511.0, B = minute 322.55 to 323.55, C = minute 436.31 to 437.31, D = minute 479.46 to 480.46, E = minute 541.0 to 542.0. Report which segment labels contain arousal and which segment labels do not contain arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract whether each segment A, B, C, D, and E is reported as containing arousal.\nIf the report lists arousal segments and non-arousal segments as groups, convert the groups into one yes/no value for each segment.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"yes\", \"no\", or null.\n6. Use \"yes\" if the segment is reported as containing arousal.\n7. Use \"no\" if the segment is reported as not containing arousal.\n8. If a segment's status is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"yes\", \"B\": \"no\", \"C\": \"yes\", \"D\": \"no\", \"E\": \"no\"}", + "metrics": [ + { + "metric_id": "segment_A_arousal_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_arousal_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_C_arousal_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_D_arousal_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_E_arousal_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_01.json b/sleep_assessment/cases/case30/case30_01.json new file mode 100644 index 0000000000000000000000000000000000000000..9ad52c4cc88c8973784b1121e01032611439faf3 --- /dev/null +++ b/sleep_assessment/cases/case30/case30_01.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 182.75-183.75 minutes; B: 109.75-110.75 minutes; C: 22.5-23.5 minutes; D: 168.25-169.25 minutes; E: 81.5-82.5 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: X5 Snore is the snore signal; X6 Flow-1 is pressure-based airflow; DC3 Flow-2 is the second airflow signal; X7 and X8 Abdominal effort are abdominal respiratory-effort channels." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_02.json b/sleep_assessment/cases/case30/case30_02.json new file mode 100644 index 0000000000000000000000000000000000000000..3b7f1917b1970193c265ccbe331678f176d77e03 --- /dev/null +++ b/sleep_assessment/cases/case30/case30_02.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 225.25-226.25 minutes; B: 194.25-195.25 minutes; C: 141-142 minutes; D: 329.25-330.25 minutes; E: 295.5-296.5 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: X5 Snore is the snore signal; X6 Flow-1 is pressure-based airflow; DC3 Flow-2 is the second airflow signal; X7 and X8 Abdominal effort are abdominal respiratory-effort channels." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_03.json b/sleep_assessment/cases/case30/case30_03.json new file mode 100644 index 0000000000000000000000000000000000000000..376bb5052b787da2b8b13bc8742fbcd422db33da --- /dev/null +++ b/sleep_assessment/cases/case30/case30_03.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 250.75-251.75 minutes; B: 344.25-345.25 minutes; C: 372.5-373.5 minutes; D: 417-418 minutes; E: 351.25-352.25 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: X5 Snore is the snore signal; X6 Flow-1 is pressure-based airflow; DC3 Flow-2 is the second airflow signal; X7 and X8 Abdominal effort are abdominal respiratory-effort channels." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "apnea" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_04.json b/sleep_assessment/cases/case30/case30_04.json new file mode 100644 index 0000000000000000000000000000000000000000..abbe6ee4714613e7e46b24b233a800666a57991f --- /dev/null +++ b/sleep_assessment/cases/case30/case30_04.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 166.25-167.25 minutes; B: 123.75-124.75 minutes; C: 30.5-31.5 minutes; D: 131.75-132.75 minutes; E: 100-101 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: X5 Snore is the snore signal; X6 Flow-1 is pressure-based airflow; DC3 Flow-2 is the second airflow signal; X7 and X8 Abdominal effort are abdominal respiratory-effort channels." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_05.json b/sleep_assessment/cases/case30/case30_05.json new file mode 100644 index 0000000000000000000000000000000000000000..31595b25ed6f2e09b109e0d61625e14ff21449de --- /dev/null +++ b/sleep_assessment/cases/case30/case30_05.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 146.75-147.75 minutes; B: 151.25-152.25 minutes; C: 188-189 minutes; D: 335.25-336.25 minutes; E: 252.5-253.5 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: X5 Snore is the snore signal; X6 Flow-1 is pressure-based airflow; DC3 Flow-2 is the second airflow signal; X7 and X8 Abdominal effort are abdominal respiratory-effort channels." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_06.json b/sleep_assessment/cases/case30/case30_06.json new file mode 100644 index 0000000000000000000000000000000000000000..86f9cc485b52fbb4a18b870aa51b00a49631632e --- /dev/null +++ b/sleep_assessment/cases/case30/case30_06.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 446.75-447.75 minutes; B: 480.25-481.25 minutes; C: 316-317 minutes; D: 381.5-382.5 minutes; E: 445-446 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: X5 Snore is the snore signal; X6 Flow-1 is pressure-based airflow; DC3 Flow-2 is the second airflow signal; X7 and X8 Abdominal effort are abdominal respiratory-effort channels." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_07.json b/sleep_assessment/cases/case30/case30_07.json new file mode 100644 index 0000000000000000000000000000000000000000..105628d16c6687daf763e99b08b41fd731e76593 --- /dev/null +++ b/sleep_assessment/cases/case30/case30_07.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 168.25-169.25 minutes; B: 259.25-260.25 minutes; C: 37-38 minutes; D: 263.75-264.75 minutes; E: 113.5-114.5 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: X5 Snore is the snore signal; X6 Flow-1 is pressure-based airflow; DC3 Flow-2 is the second airflow signal; X7 and X8 Abdominal effort are abdominal respiratory-effort channels." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_08.json b/sleep_assessment/cases/case30/case30_08.json new file mode 100644 index 0000000000000000000000000000000000000000..d80840863843c055e5091e49b8eb340b88c0fd0a --- /dev/null +++ b/sleep_assessment/cases/case30/case30_08.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 372.75-373.75 minutes; B: 264.75-265.75 minutes; C: 191-192 minutes; D: 339.25-340.25 minutes; E: 273.5-274.5 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: X5 Snore is the snore signal; X6 Flow-1 is pressure-based airflow; DC3 Flow-2 is the second airflow signal; X7 and X8 Abdominal effort are abdominal respiratory-effort channels." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_09.json b/sleep_assessment/cases/case30/case30_09.json new file mode 100644 index 0000000000000000000000000000000000000000..9ad47c3984a0073fb921df7e2ec0fff36a00f45a --- /dev/null +++ b/sleep_assessment/cases/case30/case30_09.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 445.75-446.75 minutes; B: 362.75-363.75 minutes; C: 352-353 minutes; D: 394.75-395.75 minutes; E: 432-433 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: X5 Snore is the snore signal; X6 Flow-1 is pressure-based airflow; DC3 Flow-2 is the second airflow signal; X7 and X8 Abdominal effort are abdominal respiratory-effort channels." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_10.json b/sleep_assessment/cases/case30/case30_10.json new file mode 100644 index 0000000000000000000000000000000000000000..73c98b5f42422cf62fe3f97c60adda249a15b804 --- /dev/null +++ b/sleep_assessment/cases/case30/case30_10.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "ISRUC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_04.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 194.25-195.25 minutes; B: 288.25-289.25 minutes; C: 33-34 minutes; D: 342.25-343.25 minutes; E: 100.5-101.5 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: X5 Snore is the snore signal; X6 Flow-1 is pressure-based airflow; DC3 Flow-2 is the second airflow signal; X7 and X8 Abdominal effort are abdominal respiratory-effort channels." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_11.json b/sleep_assessment/cases/case30/case30_11.json new file mode 100644 index 0000000000000000000000000000000000000000..da47ed06834be5d0d2720ef8b6794ba43e0537cb --- /dev/null +++ b/sleep_assessment/cases/case30/case30_11.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "ISRUC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_04.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 345.25-346.25 minutes; B: 348.25-349.25 minutes; C: 168-169 minutes; D: 349.25-350.25 minutes; E: 236.5-237.5 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: X5 Snore is the snore signal; X6 Flow-1 is pressure-based airflow; DC3 Flow-2 is the second airflow signal; X7 and X8 Abdominal effort are abdominal respiratory-effort channels." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_12.json b/sleep_assessment/cases/case30/case30_12.json new file mode 100644 index 0000000000000000000000000000000000000000..a0ef6184fa2384473586cafe7a4ae5ac34d14c95 --- /dev/null +++ b/sleep_assessment/cases/case30/case30_12.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "ISRUC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_04.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 300.75-301.75 minutes; B: 294.75-295.75 minutes; C: 309.5-310.5 minutes; D: 382.5-383.5 minutes; E: 448-449 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: X5 Snore is the snore signal; X6 Flow-1 is pressure-based airflow; DC3 Flow-2 is the second airflow signal; X7 and X8 Abdominal effort are abdominal respiratory-effort channels." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_13.json b/sleep_assessment/cases/case30/case30_13.json new file mode 100644 index 0000000000000000000000000000000000000000..38b66221799c8c281e9cf40f29980e7b21376ab1 --- /dev/null +++ b/sleep_assessment/cases/case30/case30_13.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 355.75-356.75 minutes; B: 129.75-130.75 minutes; C: 31-32 minutes; D: 267.25-268.25 minutes; E: 97.5-98.5 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: X5 Snore is the snore signal; X6 Flow-1 is pressure-based airflow; DC3 Flow-2 is the second airflow signal; X7 and X8 Abdominal effort are abdominal respiratory-effort channels." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_14.json b/sleep_assessment/cases/case30/case30_14.json new file mode 100644 index 0000000000000000000000000000000000000000..95b11a4e67b514467d1741bfd299c9a67fc7c758 --- /dev/null +++ b/sleep_assessment/cases/case30/case30_14.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 361.75-362.75 minutes; B: 310.75-311.75 minutes; C: 164.5-165.5 minutes; D: 323.25-324.25 minutes; E: 231-232 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: X5 Snore is the snore signal; X6 Flow-1 is pressure-based airflow; DC3 Flow-2 is the second airflow signal; X7 and X8 Abdominal effort are abdominal respiratory-effort channels." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_15.json b/sleep_assessment/cases/case30/case30_15.json new file mode 100644 index 0000000000000000000000000000000000000000..885f33018b8c2a001954068477b957df115c2663 --- /dev/null +++ b/sleep_assessment/cases/case30/case30_15.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 348.25-349.25 minutes; B: 352.75-353.75 minutes; C: 302.5-303.5 minutes; D: 382.75-383.75 minutes; E: 406-407 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: X5 Snore is the snore signal; X6 Flow-1 is pressure-based airflow; DC3 Flow-2 is the second airflow signal; X7 and X8 Abdominal effort are abdominal respiratory-effort channels." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_16.json b/sleep_assessment/cases/case30/case30_16.json new file mode 100644 index 0000000000000000000000000000000000000000..8dc0ba8aabcf397020d2433855c4b906c5662857 --- /dev/null +++ b/sleep_assessment/cases/case30/case30_16.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 231.8-232.8 minutes; B: 80.51-81.51 minutes; C: 24.5-25.5 minutes; D: 142.92-143.92 minutes; E: 77.5-78.5 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: THOR RES is thoracic respiratory effort; ABDO RES is abdominal respiratory effort; NEW AIR is nasal/oral airflow; SaO2 is oxygen saturation." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_17.json b/sleep_assessment/cases/case30/case30_17.json new file mode 100644 index 0000000000000000000000000000000000000000..0e64996b4113837afa827ebbc3b76af2b06f0d7d --- /dev/null +++ b/sleep_assessment/cases/case30/case30_17.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 236.41-237.41 minutes; B: 311.25-312.25 minutes; C: 156.5-157.5 minutes; D: 350.75-351.75 minutes; E: 208-209 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: THOR RES is thoracic respiratory effort; ABDO RES is abdominal respiratory effort; NEW AIR is nasal/oral airflow; SaO2 is oxygen saturation." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_18.json b/sleep_assessment/cases/case30/case30_18.json new file mode 100644 index 0000000000000000000000000000000000000000..43a572a2be43f932f38ebfc6c983df8745478869 --- /dev/null +++ b/sleep_assessment/cases/case30/case30_18.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 412.35-413.35 minutes; B: 437.86-438.86 minutes; C: 276.5-277.5 minutes; D: 339.5-340.5 minutes; E: 403-404 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: THOR RES is thoracic respiratory effort; ABDO RES is abdominal respiratory effort; NEW AIR is nasal/oral airflow; SaO2 is oxygen saturation." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_19.json b/sleep_assessment/cases/case30/case30_19.json new file mode 100644 index 0000000000000000000000000000000000000000..d51e8415493b1666c4053adde1766532347e01aa --- /dev/null +++ b/sleep_assessment/cases/case30/case30_19.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 74.95-75.95 minutes; B: 47.25-48.25 minutes; C: 59.5-60.5 minutes; D: 112.14-113.14 minutes; E: 86.5-87.5 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: THOR RES is thoracic respiratory effort; ABDO RES is abdominal respiratory effort; NEW AIR is nasal/oral airflow; SaO2 is oxygen saturation." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_20.json b/sleep_assessment/cases/case30/case30_20.json new file mode 100644 index 0000000000000000000000000000000000000000..70036560eb259955ae6227f92b722a123508077e --- /dev/null +++ b/sleep_assessment/cases/case30/case30_20.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 75-76 minutes; B: 151.87-152.87 minutes; C: 118.5-119.5 minutes; D: 172.5-173.5 minutes; E: 157-158 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: THOR RES is thoracic respiratory effort; ABDO RES is abdominal respiratory effort; NEW AIR is nasal/oral airflow; SaO2 is oxygen saturation." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_21.json b/sleep_assessment/cases/case30/case30_21.json new file mode 100644 index 0000000000000000000000000000000000000000..886270b163627c7d96802aa79a1cbd184fcca84b --- /dev/null +++ b/sleep_assessment/cases/case30/case30_21.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 76.8-77.8 minutes; B: 197.02-198.02 minutes; C: 186-187 minutes; D: 223-224 minutes; E: 200.43-201.43 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: THOR RES is thoracic respiratory effort; ABDO RES is abdominal respiratory effort; NEW AIR is nasal/oral airflow; SaO2 is oxygen saturation." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "apnea" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_22.json b/sleep_assessment/cases/case30/case30_22.json new file mode 100644 index 0000000000000000000000000000000000000000..94c337fe26338ed0d1bfe971d9af69a7631e4732 --- /dev/null +++ b/sleep_assessment/cases/case30/case30_22.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 225.24-226.24 minutes; B: 144.84-145.84 minutes; C: 32-33 minutes; D: 227.25-228.25 minutes; E: 97-98 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: THOR RES is thoracic respiratory effort; ABDO RES is abdominal respiratory effort; NEW AIR is nasal/oral airflow; SaO2 is oxygen saturation." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_23.json b/sleep_assessment/cases/case30/case30_23.json new file mode 100644 index 0000000000000000000000000000000000000000..9e2e16f01a199bbecee9211deb9054f6b57b6fd8 --- /dev/null +++ b/sleep_assessment/cases/case30/case30_23.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 254.78-255.78 minutes; B: 248.53-249.53 minutes; C: 172-173 minutes; D: 266.28-267.28 minutes; E: 294-295 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: THOR RES is thoracic respiratory effort; ABDO RES is abdominal respiratory effort; NEW AIR is nasal/oral airflow; SaO2 is oxygen saturation." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_24.json b/sleep_assessment/cases/case30/case30_24.json new file mode 100644 index 0000000000000000000000000000000000000000..7614d6facefec6fbb5b622368f3f5b021f4d24c8 --- /dev/null +++ b/sleep_assessment/cases/case30/case30_24.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 259.73-260.73 minutes; B: 344.73-345.73 minutes; C: 386-387 minutes; D: 455.17-456.17 minutes; E: 472-473 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: THOR RES is thoracic respiratory effort; ABDO RES is abdominal respiratory effort; NEW AIR is nasal/oral airflow; SaO2 is oxygen saturation." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_25.json b/sleep_assessment/cases/case30/case30_25.json new file mode 100644 index 0000000000000000000000000000000000000000..f1eefebd1be208574906e1e933f81ea32b4fd681 --- /dev/null +++ b/sleep_assessment/cases/case30/case30_25.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 299.34-300.34 minutes; B: 27.94-28.94 minutes; C: 48-49 minutes; D: 107.57-108.57 minutes; E: 126.5-127.5 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: THOR RES is thoracic respiratory effort; ABDO RES is abdominal respiratory effort; NEW AIR is nasal/oral airflow; SaO2 is oxygen saturation." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_26.json b/sleep_assessment/cases/case30/case30_26.json new file mode 100644 index 0000000000000000000000000000000000000000..f91d7acb48a9a742dd45d2da0936584dc677a2a0 --- /dev/null +++ b/sleep_assessment/cases/case30/case30_26.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 375.5-376.5 minutes; B: 175.84-176.84 minutes; C: 210-211 minutes; D: 281.85-282.85 minutes; E: 314.5-315.5 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: THOR RES is thoracic respiratory effort; ABDO RES is abdominal respiratory effort; NEW AIR is nasal/oral airflow; SaO2 is oxygen saturation." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_27.json b/sleep_assessment/cases/case30/case30_27.json new file mode 100644 index 0000000000000000000000000000000000000000..e64814384d7fff35f843910375f0042445d7df7c --- /dev/null +++ b/sleep_assessment/cases/case30/case30_27.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 423.86-424.86 minutes; B: 321.09-322.09 minutes; C: 401.5-402.5 minutes; D: 377.91-378.91 minutes; E: 294.04-295.04 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: THOR RES is thoracic respiratory effort; ABDO RES is abdominal respiratory effort; NEW AIR is nasal/oral airflow; SaO2 is oxygen saturation." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "apnea" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_28.json b/sleep_assessment/cases/case30/case30_28.json new file mode 100644 index 0000000000000000000000000000000000000000..89e89cc917ceb1ef7beea5d977510d7be3f7c35d --- /dev/null +++ b/sleep_assessment/cases/case30/case30_28.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 146.53-147.53 minutes; B: 26.8-27.8 minutes; C: 29.5-30.5 minutes; D: 81.78-82.78 minutes; E: 98.5-99.5 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: THOR RES is thoracic respiratory effort; ABDO RES is abdominal respiratory effort; NEW AIR is nasal/oral airflow; SaO2 is oxygen saturation." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_29.json b/sleep_assessment/cases/case30/case30_29.json new file mode 100644 index 0000000000000000000000000000000000000000..bfe24d42b1f82684e5316ef97030c09a29c5c5c2 --- /dev/null +++ b/sleep_assessment/cases/case30/case30_29.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 156.25-157.25 minutes; B: 306.9-307.9 minutes; C: 171.5-172.5 minutes; D: 389.21-390.21 minutes; E: 225.5-226.5 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: THOR RES is thoracic respiratory effort; ABDO RES is abdominal respiratory effort; NEW AIR is nasal/oral airflow; SaO2 is oxygen saturation." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_30.json b/sleep_assessment/cases/case30/case30_30.json new file mode 100644 index 0000000000000000000000000000000000000000..f49f71a8674a14e9e2b66e7c2404966dfc3389b8 --- /dev/null +++ b/sleep_assessment/cases/case30/case30_30.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 416.52-417.52 minutes; B: 471.73-472.73 minutes; C: 278-279 minutes; D: 363.5-364.5 minutes; E: 434-435 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: THOR RES is thoracic respiratory effort; ABDO RES is abdominal respiratory effort; NEW AIR is nasal/oral airflow; SaO2 is oxygen saturation." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_31.json b/sleep_assessment/cases/case30/case30_31.json new file mode 100644 index 0000000000000000000000000000000000000000..dfcc099a5e1f6cef6659e94ae939b8ca322d20a8 --- /dev/null +++ b/sleep_assessment/cases/case30/case30_31.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 14.76-15.76 minutes; B: 66.2-67.2 minutes; C: 36.5-37.5 minutes; D: 116-117 minutes; E: 96-97 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: ABD is abdominal respiratory effort; CHEST is thoracic respiratory effort; AIRFLOW is the airflow signal; SaO2 is oxygen saturation." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_32.json b/sleep_assessment/cases/case30/case30_32.json new file mode 100644 index 0000000000000000000000000000000000000000..c0f70e4b09a578ebd3496ccaec3788948b2c6793 --- /dev/null +++ b/sleep_assessment/cases/case30/case30_32.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 132-133 minutes; B: 165.65-166.65 minutes; C: 164-165 minutes; D: 166.96-167.96 minutes; E: 236-237 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: ABD is abdominal respiratory effort; CHEST is thoracic respiratory effort; AIRFLOW is the airflow signal; SaO2 is oxygen saturation." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_33.json b/sleep_assessment/cases/case30/case30_33.json new file mode 100644 index 0000000000000000000000000000000000000000..c20bdd443fb005977a3dad362454c9d6caf938dd --- /dev/null +++ b/sleep_assessment/cases/case30/case30_33.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 144.55-145.55 minutes; B: 204.25-205.25 minutes; C: 295.5-296.5 minutes; D: 347-348 minutes; E: 273.78-274.78 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: ABD is abdominal respiratory effort; CHEST is thoracic respiratory effort; AIRFLOW is the airflow signal; SaO2 is oxygen saturation." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "apnea" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_34.json b/sleep_assessment/cases/case30/case30_34.json new file mode 100644 index 0000000000000000000000000000000000000000..012a9b2ae0e4835824751590c59c224ff40232f0 --- /dev/null +++ b/sleep_assessment/cases/case30/case30_34.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 26.14-27.14 minutes; B: 48.8-49.8 minutes; C: 57-58 minutes; D: 153.75-154.75 minutes; E: 107-108 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: ABD is abdominal respiratory effort; CHEST is thoracic respiratory effort; AIRFLOW is the airflow signal; SaO2 is oxygen saturation." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_35.json b/sleep_assessment/cases/case30/case30_35.json new file mode 100644 index 0000000000000000000000000000000000000000..b3c0882782f5fc9c32da6a25f2ef745926ba0de7 --- /dev/null +++ b/sleep_assessment/cases/case30/case30_35.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 135.77-136.77 minutes; B: 208.75-209.75 minutes; C: 201-202 minutes; D: 256.41-257.41 minutes; E: 280-281 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: ABD is abdominal respiratory effort; CHEST is thoracic respiratory effort; AIRFLOW is the airflow signal; SaO2 is oxygen saturation." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_36.json b/sleep_assessment/cases/case30/case30_36.json new file mode 100644 index 0000000000000000000000000000000000000000..35539df11139ea1b8c28934ca4ef14054aaf2a66 --- /dev/null +++ b/sleep_assessment/cases/case30/case30_36.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 233.59-234.59 minutes; B: 330.25-331.25 minutes; C: 334-335 minutes; D: 260.42-261.42 minutes; E: 357.25-358.25 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: ABD is abdominal respiratory effort; CHEST is thoracic respiratory effort; AIRFLOW is the airflow signal; SaO2 is oxygen saturation." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "apnea" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_37.json b/sleep_assessment/cases/case30/case30_37.json new file mode 100644 index 0000000000000000000000000000000000000000..0aa9ce42898bb7be9d95f586eb7271ea487ea50d --- /dev/null +++ b/sleep_assessment/cases/case30/case30_37.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 47.75-48.75 minutes; B: 87.75-88.75 minutes; C: 97.5-98.5 minutes; D: 128.76-129.76 minutes; E: 244.5-245.5 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: ABD is abdominal respiratory effort; CHEST is thoracic respiratory effort; AIRFLOW is the airflow signal; SaO2 is oxygen saturation." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_38.json b/sleep_assessment/cases/case30/case30_38.json new file mode 100644 index 0000000000000000000000000000000000000000..a636ef142f68f6622b14fa5e61a5ec39bd16620f --- /dev/null +++ b/sleep_assessment/cases/case30/case30_38.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 121.99-122.99 minutes; B: 229.66-230.66 minutes; C: 363.5-364.5 minutes; D: 298.51-299.51 minutes; E: 257.85-258.85 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: ABD is abdominal respiratory effort; CHEST is thoracic respiratory effort; AIRFLOW is the airflow signal; SaO2 is oxygen saturation." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "apnea" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_39.json b/sleep_assessment/cases/case30/case30_39.json new file mode 100644 index 0000000000000000000000000000000000000000..aacbb1f1ca43eb57e1e43ea3996d24682fef8a09 --- /dev/null +++ b/sleep_assessment/cases/case30/case30_39.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 60.79-61.79 minutes; B: 37.42-38.42 minutes; C: 138.5-139.5 minutes; D: 140-141 minutes; E: 247-248 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: ABD is abdominal respiratory effort; CHEST is thoracic respiratory effort; AIRFLOW is the airflow signal; SaO2 is oxygen saturation." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "null" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case30/case30_40.json b/sleep_assessment/cases/case30/case30_40.json new file mode 100644 index 0000000000000000000000000000000000000000..b854b3eccf72de326bf19b5e04f8de29c4c98601 --- /dev/null +++ b/sleep_assessment/cases/case30/case30_40.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "instruction": "Given the sleep file, determine whether each of the following sleep segments contains an apnea event or a hypopnea event: A: 127.85-128.85 minutes; B: 279.22-280.22 minutes; C: 332.5-333.5 minutes; D: 380-381 minutes; E: 205.32-206.32 minutes. Report the corresponding event for segments A-E in order as apnea, hypopnea, or null.\n\nEach segment contains 60 seconds of sleep data. Prioritize airflow, respiratory-effort, and oxygen-saturation related channels when making the classification, and list the final A-E conclusions clearly.\n\nRespiratory channel notes: ABD is abdominal respiratory effort; CHEST is thoracic respiratory effort; AIRFLOW is the airflow signal; SaO2 is oxygen saturation." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the final event classification for each of five segments A, B, C, D, and E.\nNormalize the values as follows:\n- apnea: apnea, respiratory apnea, obstructive apnea, central apnea, mixed apnea\n- hypopnea: hypopnea, respiratory hypopnea\n- null: Null, null, none, no event, no apnea or hypopnea, no target respiratory event\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or extra fields.\n4. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n5. Each value must be exactly one of: \"apnea\", \"hypopnea\", or \"null\".\n6. If a segment's classification is missing or unclear, return the string \"null\" for that segment.\n7. Do not infer labels from explanations alone if no final A-E classification is stated clearly; return the string \"null\" for any unclear segment.\n\n### OUTPUT TEMPLATE\n{\"A\": \"apnea\", \"B\": \"hypopnea\", \"C\": \"null\", \"D\": \"apnea\", \"E\": \"null\"}", + "metrics": [ + { + "metric_id": "segment_A_event_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "apnea" + } + }, + { + "metric_id": "segment_B_event_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "hypopnea" + } + }, + { + "metric_id": "segment_C_event_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_D_event_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "null" + } + }, + { + "metric_id": "segment_E_event_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "apnea" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case31/case31_01.json b/sleep_assessment/cases/case31/case31_01.json new file mode 100644 index 0000000000000000000000000000000000000000..d3279499f821b579d0185ed11ab15abc21b44275 --- /dev/null +++ b/sleep_assessment/cases/case31/case31_01.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Arousal Index (ARI). Count the number of arousal events across the full recording from the sleep signal. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds, converted to hours. ARI is defined as arousal count divided by TST in hours, reported as arousal events per hour of total sleep time. Please clearly report the final ARI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the whole-night Arousal Index (ARI) from the agent's report. ARI should be reported as arousal events per hour of total sleep time.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"ari\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"ari\": null}.\n\n### OUTPUT TEMPLATE\n{\"ari\": }", + "metrics": [ + { + "metric_id": "arousal_index_accuracy", + "type": "numeric_check", + "target_key": "ari", + "weight": 100, + "params": { + "gt_value": 15.779220779220777, + "tolerance": 3.1558441558441555 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case31/case31_02.json b/sleep_assessment/cases/case31/case31_02.json new file mode 100644 index 0000000000000000000000000000000000000000..f8e50c6090bf46f96b2b466425bd103036d8c8a4 --- /dev/null +++ b/sleep_assessment/cases/case31/case31_02.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "label_path": "data/sleep/ISRUC_02.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Arousal Index (ARI). Count the number of arousal events across the full recording from the sleep signal. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds, converted to hours. ARI is defined as arousal count divided by TST in hours, reported as arousal events per hour of total sleep time. Please clearly report the final ARI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the whole-night Arousal Index (ARI) from the agent's report. ARI should be reported as arousal events per hour of total sleep time.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"ari\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"ari\": null}.\n\n### OUTPUT TEMPLATE\n{\"ari\": }", + "metrics": [ + { + "metric_id": "arousal_index_accuracy", + "type": "numeric_check", + "target_key": "ari", + "weight": 100, + "params": { + "gt_value": 13.518776077885954, + "tolerance": 2.703755215577191 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case31/case31_03.json b/sleep_assessment/cases/case31/case31_03.json new file mode 100644 index 0000000000000000000000000000000000000000..03a8498f7d46024ca9dcf91c37c9e5c34fdd2467 --- /dev/null +++ b/sleep_assessment/cases/case31/case31_03.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "label_path": "data/sleep/ISRUC_03.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Arousal Index (ARI). Count the number of arousal events across the full recording from the sleep signal. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds, converted to hours. ARI is defined as arousal count divided by TST in hours, reported as arousal events per hour of total sleep time. Please clearly report the final ARI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the whole-night Arousal Index (ARI) from the agent's report. ARI should be reported as arousal events per hour of total sleep time.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"ari\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"ari\": null}.\n\n### OUTPUT TEMPLATE\n{\"ari\": }", + "metrics": [ + { + "metric_id": "arousal_index_accuracy", + "type": "numeric_check", + "target_key": "ari", + "weight": 100, + "params": { + "gt_value": 14.500616522811344, + "tolerance": 2.900123304562269 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case31/case31_04.json b/sleep_assessment/cases/case31/case31_04.json new file mode 100644 index 0000000000000000000000000000000000000000..04de9491e3faf004d137c55bb157b4ca498945e3 --- /dev/null +++ b/sleep_assessment/cases/case31/case31_04.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "ISRUC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_04.edf", + "label_path": "data/sleep/ISRUC_04.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Arousal Index (ARI). Count the number of arousal events across the full recording from the sleep signal. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds, converted to hours. ARI is defined as arousal count divided by TST in hours, reported as arousal events per hour of total sleep time. Please clearly report the final ARI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the whole-night Arousal Index (ARI) from the agent's report. ARI should be reported as arousal events per hour of total sleep time.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"ari\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"ari\": null}.\n\n### OUTPUT TEMPLATE\n{\"ari\": }", + "metrics": [ + { + "metric_id": "arousal_index_accuracy", + "type": "numeric_check", + "target_key": "ari", + "weight": 100, + "params": { + "gt_value": 1.5401069518716577, + "tolerance": 0.30802139037433157 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case31/case31_05.json b/sleep_assessment/cases/case31/case31_05.json new file mode 100644 index 0000000000000000000000000000000000000000..280f4ac4d1f4359021d72acf896edcfc915ef5b0 --- /dev/null +++ b/sleep_assessment/cases/case31/case31_05.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "label_path": "data/sleep/ISRUC_05.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Arousal Index (ARI). Count the number of arousal events across the full recording from the sleep signal. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds, converted to hours. ARI is defined as arousal count divided by TST in hours, reported as arousal events per hour of total sleep time. Please clearly report the final ARI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the whole-night Arousal Index (ARI) from the agent's report. ARI should be reported as arousal events per hour of total sleep time.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"ari\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"ari\": null}.\n\n### OUTPUT TEMPLATE\n{\"ari\": }", + "metrics": [ + { + "metric_id": "arousal_index_accuracy", + "type": "numeric_check", + "target_key": "ari", + "weight": 100, + "params": { + "gt_value": 14.507772020725389, + "tolerance": 2.901554404145078 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case31/case31_06.json b/sleep_assessment/cases/case31/case31_06.json new file mode 100644 index 0000000000000000000000000000000000000000..3ae3d5cbc011d0e2436a5ca112121f22b203b8f6 --- /dev/null +++ b/sleep_assessment/cases/case31/case31_06.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "label_path": "data/sleep/SHHS1_01.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Arousal Index (ARI). Count the number of arousal events across the full recording from the sleep signal. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds, converted to hours. ARI is defined as arousal count divided by TST in hours, reported as arousal events per hour of total sleep time. Please clearly report the final ARI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the whole-night Arousal Index (ARI) from the agent's report. ARI should be reported as arousal events per hour of total sleep time.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"ari\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"ari\": null}.\n\n### OUTPUT TEMPLATE\n{\"ari\": }", + "metrics": [ + { + "metric_id": "arousal_index_accuracy", + "type": "numeric_check", + "target_key": "ari", + "weight": 100, + "params": { + "gt_value": 29.241011984021302, + "tolerance": 5.848202396804261 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case31/case31_07.json b/sleep_assessment/cases/case31/case31_07.json new file mode 100644 index 0000000000000000000000000000000000000000..dbbd088b33192d7cd079639a138bed53a3304b8c --- /dev/null +++ b/sleep_assessment/cases/case31/case31_07.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "label_path": "data/sleep/SHHS1_02.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Arousal Index (ARI). Count the number of arousal events across the full recording from the sleep signal. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds, converted to hours. ARI is defined as arousal count divided by TST in hours, reported as arousal events per hour of total sleep time. Please clearly report the final ARI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the whole-night Arousal Index (ARI) from the agent's report. ARI should be reported as arousal events per hour of total sleep time.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"ari\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"ari\": null}.\n\n### OUTPUT TEMPLATE\n{\"ari\": }", + "metrics": [ + { + "metric_id": "arousal_index_accuracy", + "type": "numeric_check", + "target_key": "ari", + "weight": 100, + "params": { + "gt_value": 30.65934065934066, + "tolerance": 6.131868131868132 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case31/case31_08.json b/sleep_assessment/cases/case31/case31_08.json new file mode 100644 index 0000000000000000000000000000000000000000..6cc381340e0a03a2f57197c3e537c577ba1e27af --- /dev/null +++ b/sleep_assessment/cases/case31/case31_08.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "label_path": "data/sleep/SHHS1_03.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Arousal Index (ARI). Count the number of arousal events across the full recording from the sleep signal. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds, converted to hours. ARI is defined as arousal count divided by TST in hours, reported as arousal events per hour of total sleep time. Please clearly report the final ARI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the whole-night Arousal Index (ARI) from the agent's report. ARI should be reported as arousal events per hour of total sleep time.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"ari\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"ari\": null}.\n\n### OUTPUT TEMPLATE\n{\"ari\": }", + "metrics": [ + { + "metric_id": "arousal_index_accuracy", + "type": "numeric_check", + "target_key": "ari", + "weight": 100, + "params": { + "gt_value": 27.280334728033473, + "tolerance": 5.456066945606695 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case31/case31_09.json b/sleep_assessment/cases/case31/case31_09.json new file mode 100644 index 0000000000000000000000000000000000000000..6b9c8189877bc8e37caa92f920232744b544e683 --- /dev/null +++ b/sleep_assessment/cases/case31/case31_09.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "label_path": "data/sleep/SHHS1_04.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Arousal Index (ARI). Count the number of arousal events across the full recording from the sleep signal. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds, converted to hours. ARI is defined as arousal count divided by TST in hours, reported as arousal events per hour of total sleep time. Please clearly report the final ARI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the whole-night Arousal Index (ARI) from the agent's report. ARI should be reported as arousal events per hour of total sleep time.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"ari\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"ari\": null}.\n\n### OUTPUT TEMPLATE\n{\"ari\": }", + "metrics": [ + { + "metric_id": "arousal_index_accuracy", + "type": "numeric_check", + "target_key": "ari", + "weight": 100, + "params": { + "gt_value": 8.970099667774086, + "tolerance": 1.7940199335548173 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case31/case31_10.json b/sleep_assessment/cases/case31/case31_10.json new file mode 100644 index 0000000000000000000000000000000000000000..e7d3b27a8097444d2931bcfa9c85745b36105869 --- /dev/null +++ b/sleep_assessment/cases/case31/case31_10.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "label_path": "data/sleep/SHHS1_05.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Arousal Index (ARI). Count the number of arousal events across the full recording from the sleep signal. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds, converted to hours. ARI is defined as arousal count divided by TST in hours, reported as arousal events per hour of total sleep time. Please clearly report the final ARI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the whole-night Arousal Index (ARI) from the agent's report. ARI should be reported as arousal events per hour of total sleep time.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"ari\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"ari\": null}.\n\n### OUTPUT TEMPLATE\n{\"ari\": }", + "metrics": [ + { + "metric_id": "arousal_index_accuracy", + "type": "numeric_check", + "target_key": "ari", + "weight": 100, + "params": { + "gt_value": 20.756756756756754, + "tolerance": 4.151351351351351 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case32/case32_01.json b/sleep_assessment/cases/case32/case32_01.json new file mode 100644 index 0000000000000000000000000000000000000000..01ea0465431565480c0f344ea0031155bc5b0f8b --- /dev/null +++ b/sleep_assessment/cases/case32/case32_01.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Apnea-Hypopnea Index (AHI). Count all apnea events and all hypopnea events across the full recording from the sleep signal. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds, converted to hours. AHI is defined as (apnea_count + hypopnea_count) divided by TST in hours, reported as events per hour of total sleep time. Respiratory channels for ISRUC: X5 = snore signal; X6 = Flow-1, pressure-based airflow; DC3 = Flow-2, second airflow channel; X7 and X8 = abdominal respiratory-effort channels. Please clearly report the final AHI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the whole-night Apnea-Hypopnea Index (AHI) from the agent's report. AHI should be reported as apnea plus hypopnea events per hour of total sleep time.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"ahi\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"ahi\": null}.\n\n### OUTPUT TEMPLATE\n{\"ahi\": }", + "metrics": [ + { + "metric_id": "apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "ahi", + "weight": 100, + "params": { + "gt_value": 32.14285714285714, + "tolerance": 6.428571428571428 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case32/case32_02.json b/sleep_assessment/cases/case32/case32_02.json new file mode 100644 index 0000000000000000000000000000000000000000..2470b1fb27b448bc3b82d5461200110839cbb60a --- /dev/null +++ b/sleep_assessment/cases/case32/case32_02.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "label_path": "data/sleep/ISRUC_02.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Apnea-Hypopnea Index (AHI). Count all apnea events and all hypopnea events across the full recording from the sleep signal. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds, converted to hours. AHI is defined as (apnea_count + hypopnea_count) divided by TST in hours, reported as events per hour of total sleep time. Respiratory channels for ISRUC: X5 = snore signal; X6 = Flow-1, pressure-based airflow; DC3 = Flow-2, second airflow channel; X7 and X8 = abdominal respiratory-effort channels. Please clearly report the final AHI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the whole-night Apnea-Hypopnea Index (AHI) from the agent's report. AHI should be reported as apnea plus hypopnea events per hour of total sleep time.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"ahi\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"ahi\": null}.\n\n### OUTPUT TEMPLATE\n{\"ahi\": }", + "metrics": [ + { + "metric_id": "apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "ahi", + "weight": 100, + "params": { + "gt_value": 5.674547983310154, + "tolerance": 1.1349095966620308 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case32/case32_03.json b/sleep_assessment/cases/case32/case32_03.json new file mode 100644 index 0000000000000000000000000000000000000000..21b85bd1efe25c825a3d1d4e8191aa50904c8221 --- /dev/null +++ b/sleep_assessment/cases/case32/case32_03.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "label_path": "data/sleep/ISRUC_03.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Apnea-Hypopnea Index (AHI). Count all apnea events and all hypopnea events across the full recording from the sleep signal. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds, converted to hours. AHI is defined as (apnea_count + hypopnea_count) divided by TST in hours, reported as events per hour of total sleep time. Respiratory channels for ISRUC: X5 = snore signal; X6 = Flow-1, pressure-based airflow; DC3 = Flow-2, second airflow channel; X7 and X8 = abdominal respiratory-effort channels. Please clearly report the final AHI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the whole-night Apnea-Hypopnea Index (AHI) from the agent's report. AHI should be reported as apnea plus hypopnea events per hour of total sleep time.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"ahi\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"ahi\": null}.\n\n### OUTPUT TEMPLATE\n{\"ahi\": }", + "metrics": [ + { + "metric_id": "apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "ahi", + "weight": 100, + "params": { + "gt_value": 1.7755856966707768, + "tolerance": 0.3551171393341554 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case32/case32_04.json b/sleep_assessment/cases/case32/case32_04.json new file mode 100644 index 0000000000000000000000000000000000000000..5d6c467b7ff9f505ddccaeb9a33e7369c84653b6 --- /dev/null +++ b/sleep_assessment/cases/case32/case32_04.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "ISRUC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_04.edf", + "label_path": "data/sleep/ISRUC_04.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Apnea-Hypopnea Index (AHI). Count all apnea events and all hypopnea events across the full recording from the sleep signal. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds, converted to hours. AHI is defined as (apnea_count + hypopnea_count) divided by TST in hours, reported as events per hour of total sleep time. Respiratory channels for ISRUC: X5 = snore signal; X6 = Flow-1, pressure-based airflow; DC3 = Flow-2, second airflow channel; X7 and X8 = abdominal respiratory-effort channels. Please clearly report the final AHI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the whole-night Apnea-Hypopnea Index (AHI) from the agent's report. AHI should be reported as apnea plus hypopnea events per hour of total sleep time.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"ahi\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"ahi\": null}.\n\n### OUTPUT TEMPLATE\n{\"ahi\": }", + "metrics": [ + { + "metric_id": "apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "ahi", + "weight": 100, + "params": { + "gt_value": 1.1550802139037433, + "tolerance": 0.23101604278074866 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case32/case32_05.json b/sleep_assessment/cases/case32/case32_05.json new file mode 100644 index 0000000000000000000000000000000000000000..cc1c0f9c205197301cd640a7a5b012cd362b768f --- /dev/null +++ b/sleep_assessment/cases/case32/case32_05.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "label_path": "data/sleep/ISRUC_05.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Apnea-Hypopnea Index (AHI). Count all apnea events and all hypopnea events across the full recording from the sleep signal. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds, converted to hours. AHI is defined as (apnea_count + hypopnea_count) divided by TST in hours, reported as events per hour of total sleep time. Respiratory channels for ISRUC: X5 = snore signal; X6 = Flow-1, pressure-based airflow; DC3 = Flow-2, second airflow channel; X7 and X8 = abdominal respiratory-effort channels. Please clearly report the final AHI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the whole-night Apnea-Hypopnea Index (AHI) from the agent's report. AHI should be reported as apnea plus hypopnea events per hour of total sleep time.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"ahi\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"ahi\": null}.\n\n### OUTPUT TEMPLATE\n{\"ahi\": }", + "metrics": [ + { + "metric_id": "apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "ahi", + "weight": 100, + "params": { + "gt_value": 8.911917098445596, + "tolerance": 1.7823834196891193 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case32/case32_06.json b/sleep_assessment/cases/case32/case32_06.json new file mode 100644 index 0000000000000000000000000000000000000000..6c77e721e92a99d3185abf2cbdc2501d431a82a5 --- /dev/null +++ b/sleep_assessment/cases/case32/case32_06.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "label_path": "data/sleep/SHHS1_01.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Apnea-Hypopnea Index (AHI). Count all apnea events and all hypopnea events across the full recording from the sleep signal. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds, converted to hours. AHI is defined as (apnea_count + hypopnea_count) divided by TST in hours, reported as events per hour of total sleep time. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry. Please clearly report the final AHI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the whole-night Apnea-Hypopnea Index (AHI) from the agent's report. AHI should be reported as apnea plus hypopnea events per hour of total sleep time.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"ahi\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"ahi\": null}.\n\n### OUTPUT TEMPLATE\n{\"ahi\": }", + "metrics": [ + { + "metric_id": "apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "ahi", + "weight": 100, + "params": { + "gt_value": 13.901464713715045, + "tolerance": 2.780292942743009 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case32/case32_07.json b/sleep_assessment/cases/case32/case32_07.json new file mode 100644 index 0000000000000000000000000000000000000000..754ae374c7a91fd875306543886636b80f5c3426 --- /dev/null +++ b/sleep_assessment/cases/case32/case32_07.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "label_path": "data/sleep/SHHS1_02.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Apnea-Hypopnea Index (AHI). Count all apnea events and all hypopnea events across the full recording from the sleep signal. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds, converted to hours. AHI is defined as (apnea_count + hypopnea_count) divided by TST in hours, reported as events per hour of total sleep time. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry. Please clearly report the final AHI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the whole-night Apnea-Hypopnea Index (AHI) from the agent's report. AHI should be reported as apnea plus hypopnea events per hour of total sleep time.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"ahi\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"ahi\": null}.\n\n### OUTPUT TEMPLATE\n{\"ahi\": }", + "metrics": [ + { + "metric_id": "apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "ahi", + "weight": 100, + "params": { + "gt_value": 58.02197802197802, + "tolerance": 11.604395604395606 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case32/case32_08.json b/sleep_assessment/cases/case32/case32_08.json new file mode 100644 index 0000000000000000000000000000000000000000..426e73f1885a7465a54963ce00d7f0ce590a12cf --- /dev/null +++ b/sleep_assessment/cases/case32/case32_08.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "label_path": "data/sleep/SHHS1_03.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Apnea-Hypopnea Index (AHI). Count all apnea events and all hypopnea events across the full recording from the sleep signal. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds, converted to hours. AHI is defined as (apnea_count + hypopnea_count) divided by TST in hours, reported as events per hour of total sleep time. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry. Please clearly report the final AHI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the whole-night Apnea-Hypopnea Index (AHI) from the agent's report. AHI should be reported as apnea plus hypopnea events per hour of total sleep time.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"ahi\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"ahi\": null}.\n\n### OUTPUT TEMPLATE\n{\"ahi\": }", + "metrics": [ + { + "metric_id": "apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "ahi", + "weight": 100, + "params": { + "gt_value": 18.410041841004187, + "tolerance": 3.6820083682008375 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case32/case32_09.json b/sleep_assessment/cases/case32/case32_09.json new file mode 100644 index 0000000000000000000000000000000000000000..30013975c9b9a6cd789d15d62253c1343256eafa --- /dev/null +++ b/sleep_assessment/cases/case32/case32_09.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "label_path": "data/sleep/SHHS1_04.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Apnea-Hypopnea Index (AHI). Count all apnea events and all hypopnea events across the full recording from the sleep signal. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds, converted to hours. AHI is defined as (apnea_count + hypopnea_count) divided by TST in hours, reported as events per hour of total sleep time. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry. Please clearly report the final AHI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the whole-night Apnea-Hypopnea Index (AHI) from the agent's report. AHI should be reported as apnea plus hypopnea events per hour of total sleep time.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"ahi\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"ahi\": null}.\n\n### OUTPUT TEMPLATE\n{\"ahi\": }", + "metrics": [ + { + "metric_id": "apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "ahi", + "weight": 100, + "params": { + "gt_value": 25.714285714285715, + "tolerance": 5.142857142857143 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case32/case32_10.json b/sleep_assessment/cases/case32/case32_10.json new file mode 100644 index 0000000000000000000000000000000000000000..81259d1ede59130f43aac614bafaa4f54f1e356c --- /dev/null +++ b/sleep_assessment/cases/case32/case32_10.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "label_path": "data/sleep/SHHS1_05.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Apnea-Hypopnea Index (AHI). Count all apnea events and all hypopnea events across the full recording from the sleep signal. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds, converted to hours. AHI is defined as (apnea_count + hypopnea_count) divided by TST in hours, reported as events per hour of total sleep time. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry. Please clearly report the final AHI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the whole-night Apnea-Hypopnea Index (AHI) from the agent's report. AHI should be reported as apnea plus hypopnea events per hour of total sleep time.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"ahi\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"ahi\": null}.\n\n### OUTPUT TEMPLATE\n{\"ahi\": }", + "metrics": [ + { + "metric_id": "apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "ahi", + "weight": 100, + "params": { + "gt_value": 15.72972972972973, + "tolerance": 3.1459459459459462 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case32/case32_11.json b/sleep_assessment/cases/case32/case32_11.json new file mode 100644 index 0000000000000000000000000000000000000000..5f0a43b63eb09ec268f2fe668bd91f00b1071db1 --- /dev/null +++ b/sleep_assessment/cases/case32/case32_11.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "label_path": "data/sleep/Physionet2018_01.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Apnea-Hypopnea Index (AHI). Count all apnea events and all hypopnea events across the full recording from the sleep signal. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds, converted to hours. AHI is defined as (apnea_count + hypopnea_count) divided by TST in hours, reported as events per hour of total sleep time. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry. Please clearly report the final AHI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the whole-night Apnea-Hypopnea Index (AHI) from the agent's report. AHI should be reported as apnea plus hypopnea events per hour of total sleep time.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"ahi\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"ahi\": null}.\n\n### OUTPUT TEMPLATE\n{\"ahi\": }", + "metrics": [ + { + "metric_id": "apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "ahi", + "weight": 100, + "params": { + "gt_value": 10.794979079497908, + "tolerance": 2.1589958158995817 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case32/case32_12.json b/sleep_assessment/cases/case32/case32_12.json new file mode 100644 index 0000000000000000000000000000000000000000..2192eb80b6d1e57a712b941d9840b6e966cc0d7e --- /dev/null +++ b/sleep_assessment/cases/case32/case32_12.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "label_path": "data/sleep/Physionet2018_02.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Apnea-Hypopnea Index (AHI). Count all apnea events and all hypopnea events across the full recording from the sleep signal. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds, converted to hours. AHI is defined as (apnea_count + hypopnea_count) divided by TST in hours, reported as events per hour of total sleep time. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry. Please clearly report the final AHI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the whole-night Apnea-Hypopnea Index (AHI) from the agent's report. AHI should be reported as apnea plus hypopnea events per hour of total sleep time.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"ahi\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"ahi\": null}.\n\n### OUTPUT TEMPLATE\n{\"ahi\": }", + "metrics": [ + { + "metric_id": "apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "ahi", + "weight": 100, + "params": { + "gt_value": 39.91304347826087, + "tolerance": 7.982608695652174 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case32/case32_13.json b/sleep_assessment/cases/case32/case32_13.json new file mode 100644 index 0000000000000000000000000000000000000000..b9e60222899d2adcd33be965a07b35a73c0a918f --- /dev/null +++ b/sleep_assessment/cases/case32/case32_13.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "Physionet2018_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_03.edf", + "label_path": "data/sleep/Physionet2018_03.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Apnea-Hypopnea Index (AHI). Count all apnea events and all hypopnea events across the full recording from the sleep signal. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds, converted to hours. AHI is defined as (apnea_count + hypopnea_count) divided by TST in hours, reported as events per hour of total sleep time. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry. Please clearly report the final AHI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the whole-night Apnea-Hypopnea Index (AHI) from the agent's report. AHI should be reported as apnea plus hypopnea events per hour of total sleep time.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"ahi\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"ahi\": null}.\n\n### OUTPUT TEMPLATE\n{\"ahi\": }", + "metrics": [ + { + "metric_id": "apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "ahi", + "weight": 100, + "params": { + "gt_value": 0.7964601769911505, + "tolerance": 0.15929203539823011 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case32/case32_14.json b/sleep_assessment/cases/case32/case32_14.json new file mode 100644 index 0000000000000000000000000000000000000000..492d276b74eb28c6bd026be34512d4faff6a7ed1 --- /dev/null +++ b/sleep_assessment/cases/case32/case32_14.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "label_path": "data/sleep/Physionet2018_04.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Apnea-Hypopnea Index (AHI). Count all apnea events and all hypopnea events across the full recording from the sleep signal. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds, converted to hours. AHI is defined as (apnea_count + hypopnea_count) divided by TST in hours, reported as events per hour of total sleep time. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry. Please clearly report the final AHI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the whole-night Apnea-Hypopnea Index (AHI) from the agent's report. AHI should be reported as apnea plus hypopnea events per hour of total sleep time.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"ahi\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"ahi\": null}.\n\n### OUTPUT TEMPLATE\n{\"ahi\": }", + "metrics": [ + { + "metric_id": "apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "ahi", + "weight": 100, + "params": { + "gt_value": 29.0956749672346, + "tolerance": 5.81913499344692 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case32/case32_15.json b/sleep_assessment/cases/case32/case32_15.json new file mode 100644 index 0000000000000000000000000000000000000000..ed2a0dc5dadcdd6e49fda5e32bb7f1480abfe956 --- /dev/null +++ b/sleep_assessment/cases/case32/case32_15.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "label_path": "data/sleep/Physionet2018_05.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Apnea-Hypopnea Index (AHI). Count all apnea events and all hypopnea events across the full recording from the sleep signal. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds, converted to hours. AHI is defined as (apnea_count + hypopnea_count) divided by TST in hours, reported as events per hour of total sleep time. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry. Please clearly report the final AHI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the whole-night Apnea-Hypopnea Index (AHI) from the agent's report. AHI should be reported as apnea plus hypopnea events per hour of total sleep time.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"ahi\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"ahi\": null}.\n\n### OUTPUT TEMPLATE\n{\"ahi\": }", + "metrics": [ + { + "metric_id": "apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "ahi", + "weight": 100, + "params": { + "gt_value": 32.666666666666664, + "tolerance": 6.533333333333333 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case33/case33_01.json b/sleep_assessment/cases/case33/case33_01.json new file mode 100644 index 0000000000000000000000000000000000000000..44ff81a8d8ac01208f7b2b8881fdb896a8652024 --- /dev/null +++ b/sleep_assessment/cases/case33/case33_01.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "label_path": "data/sleep/Physionet2018_01.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Respiratory Disturbance Index (RDI). Count all apnea events, hypopnea events, and respiratory-effort-related arousal (RERA) events across the full recording from the sleep signals. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds and converted to hours. RDI is defined as (apnea_count + hypopnea_count + RERA_count) divided by TST in hours, reported as events per hour. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry. Please clearly report the final RDI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the whole-night Respiratory Disturbance Index (RDI).\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The key must be EXACTLY \"rdi\".\n4. The value must be a FLOAT or null.\n5. If the value is missing or unclear, return {\"rdi\": null}.\n\n### OUTPUT TEMPLATE\n{\"rdi\": }", + "metrics": [ + { + "metric_id": "respiratory_disturbance_index_accuracy", + "type": "numeric_check", + "target_key": "rdi", + "weight": 100, + "params": { + "gt_value": 17.573221757322177, + "tolerance": 3.5146443514644354 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case33/case33_02.json b/sleep_assessment/cases/case33/case33_02.json new file mode 100644 index 0000000000000000000000000000000000000000..6138d7cd588abfa06dbcf4f9286f72e07c531516 --- /dev/null +++ b/sleep_assessment/cases/case33/case33_02.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "label_path": "data/sleep/Physionet2018_02.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Respiratory Disturbance Index (RDI). Count all apnea events, hypopnea events, and respiratory-effort-related arousal (RERA) events across the full recording from the sleep signals. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds and converted to hours. RDI is defined as (apnea_count + hypopnea_count + RERA_count) divided by TST in hours, reported as events per hour. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry. Please clearly report the final RDI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the whole-night Respiratory Disturbance Index (RDI).\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The key must be EXACTLY \"rdi\".\n4. The value must be a FLOAT or null.\n5. If the value is missing or unclear, return {\"rdi\": null}.\n\n### OUTPUT TEMPLATE\n{\"rdi\": }", + "metrics": [ + { + "metric_id": "respiratory_disturbance_index_accuracy", + "type": "numeric_check", + "target_key": "rdi", + "weight": 100, + "params": { + "gt_value": 51.391304347826086, + "tolerance": 10.278260869565218 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case33/case33_03.json b/sleep_assessment/cases/case33/case33_03.json new file mode 100644 index 0000000000000000000000000000000000000000..6ba7721117e20a49bd59889ae2daf6642e86e09d --- /dev/null +++ b/sleep_assessment/cases/case33/case33_03.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "Physionet2018_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_03.edf", + "label_path": "data/sleep/Physionet2018_03.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Respiratory Disturbance Index (RDI). Count all apnea events, hypopnea events, and respiratory-effort-related arousal (RERA) events across the full recording from the sleep signals. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds and converted to hours. RDI is defined as (apnea_count + hypopnea_count + RERA_count) divided by TST in hours, reported as events per hour. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry. Please clearly report the final RDI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the whole-night Respiratory Disturbance Index (RDI).\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The key must be EXACTLY \"rdi\".\n4. The value must be a FLOAT or null.\n5. If the value is missing or unclear, return {\"rdi\": null}.\n\n### OUTPUT TEMPLATE\n{\"rdi\": }", + "metrics": [ + { + "metric_id": "respiratory_disturbance_index_accuracy", + "type": "numeric_check", + "target_key": "rdi", + "weight": 100, + "params": { + "gt_value": 1.0619469026548674, + "tolerance": 0.21238938053097348 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case33/case33_04.json b/sleep_assessment/cases/case33/case33_04.json new file mode 100644 index 0000000000000000000000000000000000000000..108080a128ad1943275a5fd7e76cccc1883a6f42 --- /dev/null +++ b/sleep_assessment/cases/case33/case33_04.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "label_path": "data/sleep/Physionet2018_04.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Respiratory Disturbance Index (RDI). Count all apnea events, hypopnea events, and respiratory-effort-related arousal (RERA) events across the full recording from the sleep signals. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds and converted to hours. RDI is defined as (apnea_count + hypopnea_count + RERA_count) divided by TST in hours, reported as events per hour. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry. Please clearly report the final RDI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the whole-night Respiratory Disturbance Index (RDI).\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The key must be EXACTLY \"rdi\".\n4. The value must be a FLOAT or null.\n5. If the value is missing or unclear, return {\"rdi\": null}.\n\n### OUTPUT TEMPLATE\n{\"rdi\": }", + "metrics": [ + { + "metric_id": "respiratory_disturbance_index_accuracy", + "type": "numeric_check", + "target_key": "rdi", + "weight": 100, + "params": { + "gt_value": 32.870249017038006, + "tolerance": 6.574049803407601 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case33/case33_05.json b/sleep_assessment/cases/case33/case33_05.json new file mode 100644 index 0000000000000000000000000000000000000000..9ac5098562a008302fe398b4a1eec3708850f908 --- /dev/null +++ b/sleep_assessment/cases/case33/case33_05.json @@ -0,0 +1,28 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "label_path": "data/sleep/Physionet2018_05.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate the whole-night Respiratory Disturbance Index (RDI). Count all apnea events, hypopnea events, and respiratory-effort-related arousal (RERA) events across the full recording from the sleep signals. Compute total sleep time (TST) from the label file as the number of non-Wake epochs (N1, N2, N3, or R/REM) multiplied by 30 seconds and converted to hours. RDI is defined as (apnea_count + hypopnea_count + RERA_count) divided by TST in hours, reported as events per hour. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry. Please clearly report the final RDI value." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the whole-night Respiratory Disturbance Index (RDI).\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The key must be EXACTLY \"rdi\".\n4. The value must be a FLOAT or null.\n5. If the value is missing or unclear, return {\"rdi\": null}.\n\n### OUTPUT TEMPLATE\n{\"rdi\": }", + "metrics": [ + { + "metric_id": "respiratory_disturbance_index_accuracy", + "type": "numeric_check", + "target_key": "rdi", + "weight": 100, + "params": { + "gt_value": 37.86666666666667, + "tolerance": 7.573333333333334 + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case33/case33_06.json b/sleep_assessment/cases/case33/case33_06.json new file mode 100644 index 0000000000000000000000000000000000000000..a595226e714f9d9e8e079ec9edcaaf3dcd1dde33 --- /dev/null +++ b/sleep_assessment/cases/case33/case33_06.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "instruction": "Given the sleep file, determine whether a respiratory-effort-related arousal (RERA) event is present in each of the following five 60-second sleep segments: A: 66.92-67.92 minutes; B: 10.01-11.01 minutes; C: 77.7-78.7 minutes; D: 151.59-152.59 minutes; E: 86.5-87.5 minutes. Report yes or no for segments A-E in order. Distinguish RERA from apnea, hypopnea, and segments without a target respiratory event. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract whether RERA is present in segments A, B, C, D, and E.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n4. Each value must be exactly \"yes\", \"no\", or null.\n5. Normalize affirmative RERA conclusions to \"yes\" and absent/no-RERA conclusions to \"no\".\n6. If a segment is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"yes\",\"B\":\"no\",\"C\":\"no\",\"D\":\"yes\",\"E\":\"no\"}", + "metrics": [ + { + "metric_id": "segment_A_rera_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 25, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_B_rera_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 17, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_rera_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 17, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_D_rera_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 25, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_E_rera_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 16, + "params": { + "gt_value": "no" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case33/case33_07.json b/sleep_assessment/cases/case33/case33_07.json new file mode 100644 index 0000000000000000000000000000000000000000..991f95c6bdc1bb0461fd069c96b1e2cc7b9447cf --- /dev/null +++ b/sleep_assessment/cases/case33/case33_07.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "instruction": "Given the sleep file, determine whether a respiratory-effort-related arousal (RERA) event is present in each of the following five 60-second sleep segments: A: 144.55-145.55 minutes; B: 115.08-116.08 minutes; C: 166.89-167.89 minutes; D: 188.5-189.5 minutes; E: 238.42-239.42 minutes. Report yes or no for segments A-E in order. Distinguish RERA from apnea, hypopnea, and segments without a target respiratory event. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract whether RERA is present in segments A, B, C, D, and E.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n4. Each value must be exactly \"yes\", \"no\", or null.\n5. Normalize affirmative RERA conclusions to \"yes\" and absent/no-RERA conclusions to \"no\".\n6. If a segment is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"yes\",\"B\":\"no\",\"C\":\"no\",\"D\":\"yes\",\"E\":\"no\"}", + "metrics": [ + { + "metric_id": "segment_A_rera_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 17, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_rera_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 25, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_C_rera_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 17, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_D_rera_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 16, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_rera_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 25, + "params": { + "gt_value": "yes" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case33/case33_08.json b/sleep_assessment/cases/case33/case33_08.json new file mode 100644 index 0000000000000000000000000000000000000000..d03b5a094e479684dfe3fb1b703c5a9580f76412 --- /dev/null +++ b/sleep_assessment/cases/case33/case33_08.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "instruction": "Given the sleep file, determine whether a respiratory-effort-related arousal (RERA) event is present in each of the following five 60-second sleep segments: A: 127.68-128.68 minutes; B: 23.94-24.94 minutes; C: 250.04-251.04 minutes; D: 130.66-131.66 minutes; E: 200-201 minutes. Report yes or no for segments A-E in order. Distinguish RERA from apnea, hypopnea, and segments without a target respiratory event. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract whether RERA is present in segments A, B, C, D, and E.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n4. Each value must be exactly \"yes\", \"no\", or null.\n5. Normalize affirmative RERA conclusions to \"yes\" and absent/no-RERA conclusions to \"no\".\n6. If a segment is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"yes\",\"B\":\"no\",\"C\":\"no\",\"D\":\"yes\",\"E\":\"no\"}", + "metrics": [ + { + "metric_id": "segment_A_rera_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 25, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_B_rera_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 17, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_rera_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 25, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_D_rera_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 17, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_rera_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 16, + "params": { + "gt_value": "no" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case33/case33_09.json b/sleep_assessment/cases/case33/case33_09.json new file mode 100644 index 0000000000000000000000000000000000000000..29ed90a8104733db417d7e9f61c573bd1241056d --- /dev/null +++ b/sleep_assessment/cases/case33/case33_09.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "instruction": "Given the sleep file, determine whether a respiratory-effort-related arousal (RERA) event is present in each of the following five 60-second sleep segments: A: 320.3-321.3 minutes; B: 251.74-252.74 minutes; C: 291.5-292.5 minutes; D: 305.98-306.98 minutes; E: 367.31-368.31 minutes. Report yes or no for segments A-E in order. Distinguish RERA from apnea, hypopnea, and segments without a target respiratory event. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract whether RERA is present in segments A, B, C, D, and E.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n4. Each value must be exactly \"yes\", \"no\", or null.\n5. Normalize affirmative RERA conclusions to \"yes\" and absent/no-RERA conclusions to \"no\".\n6. If a segment is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"yes\",\"B\":\"no\",\"C\":\"no\",\"D\":\"yes\",\"E\":\"no\"}", + "metrics": [ + { + "metric_id": "segment_A_rera_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 17, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_rera_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 17, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_rera_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 16, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_D_rera_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 25, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_E_rera_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 25, + "params": { + "gt_value": "yes" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case33/case33_10.json b/sleep_assessment/cases/case33/case33_10.json new file mode 100644 index 0000000000000000000000000000000000000000..a64a9539c8048f485837e629f79f7a8da869bdc1 --- /dev/null +++ b/sleep_assessment/cases/case33/case33_10.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "Physionet2018_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_03.edf", + "instruction": "Given the sleep file, determine whether a respiratory-effort-related arousal (RERA) event is present in each of the following five 60-second sleep segments: A: 16.52-17.52 minutes; B: 23.43-24.43 minutes; C: 132.98-133.98 minutes; D: 82.5-83.5 minutes; E: 270.5-271.5 minutes. Report yes or no for segments A-E in order. Distinguish RERA from apnea, hypopnea, and segments without a target respiratory event. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract whether RERA is present in segments A, B, C, D, and E.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n4. Each value must be exactly \"yes\", \"no\", or null.\n5. Normalize affirmative RERA conclusions to \"yes\" and absent/no-RERA conclusions to \"no\".\n6. If a segment is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"yes\",\"B\":\"no\",\"C\":\"no\",\"D\":\"yes\",\"E\":\"no\"}", + "metrics": [ + { + "metric_id": "segment_A_rera_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_rera_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_rera_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_D_rera_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_rera_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case33/case33_11.json b/sleep_assessment/cases/case33/case33_11.json new file mode 100644 index 0000000000000000000000000000000000000000..0fb54c880f94af5327eeb5a53736d8845774e8d2 --- /dev/null +++ b/sleep_assessment/cases/case33/case33_11.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "Physionet2018_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_03.edf", + "instruction": "Given the sleep file, determine whether a respiratory-effort-related arousal (RERA) event is present in each of the following five 60-second sleep segments: A: 26.95-27.95 minutes; B: 4-5 minutes; C: 59-60 minutes; D: 294.5-295.5 minutes; E: 397.5-398.5 minutes. Report yes or no for segments A-E in order. Distinguish RERA from apnea, hypopnea, and segments without a target respiratory event. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract whether RERA is present in segments A, B, C, D, and E.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n4. Each value must be exactly \"yes\", \"no\", or null.\n5. Normalize affirmative RERA conclusions to \"yes\" and absent/no-RERA conclusions to \"no\".\n6. If a segment is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"yes\",\"B\":\"no\",\"C\":\"no\",\"D\":\"yes\",\"E\":\"no\"}", + "metrics": [ + { + "metric_id": "segment_A_rera_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_rera_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_rera_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_D_rera_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 20, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_rera_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 20, + "params": { + "gt_value": "no" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case33/case33_12.json b/sleep_assessment/cases/case33/case33_12.json new file mode 100644 index 0000000000000000000000000000000000000000..8d71e111661b08e947ae23941669b4715db57c5b --- /dev/null +++ b/sleep_assessment/cases/case33/case33_12.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "instruction": "Given the sleep file, determine whether a respiratory-effort-related arousal (RERA) event is present in each of the following five 60-second sleep segments: A: 33.16-34.16 minutes; B: 85.14-86.14 minutes; C: 55.52-56.52 minutes; D: 165.98-166.98 minutes; E: 193.5-194.5 minutes. Report yes or no for segments A-E in order. Distinguish RERA from apnea, hypopnea, and segments without a target respiratory event. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract whether RERA is present in segments A, B, C, D, and E.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n4. Each value must be exactly \"yes\", \"no\", or null.\n5. Normalize affirmative RERA conclusions to \"yes\" and absent/no-RERA conclusions to \"no\".\n6. If a segment is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"yes\",\"B\":\"no\",\"C\":\"no\",\"D\":\"yes\",\"E\":\"no\"}", + "metrics": [ + { + "metric_id": "segment_A_rera_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 17, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_rera_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 25, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_C_rera_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 17, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_D_rera_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 25, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_E_rera_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 16, + "params": { + "gt_value": "no" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case33/case33_13.json b/sleep_assessment/cases/case33/case33_13.json new file mode 100644 index 0000000000000000000000000000000000000000..c2054eb674066ca7855ca198fa18ba723a695ccd --- /dev/null +++ b/sleep_assessment/cases/case33/case33_13.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "instruction": "Given the sleep file, determine whether a respiratory-effort-related arousal (RERA) event is present in each of the following five 60-second sleep segments: A: 108.17-109.17 minutes; B: 121.99-122.99 minutes; C: 181-182 minutes; D: 359.5-360.5 minutes; E: 370.41-371.41 minutes. Report yes or no for segments A-E in order. Distinguish RERA from apnea, hypopnea, and segments without a target respiratory event. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract whether RERA is present in segments A, B, C, D, and E.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n4. Each value must be exactly \"yes\", \"no\", or null.\n5. Normalize affirmative RERA conclusions to \"yes\" and absent/no-RERA conclusions to \"no\".\n6. If a segment is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"yes\",\"B\":\"no\",\"C\":\"no\",\"D\":\"yes\",\"E\":\"no\"}", + "metrics": [ + { + "metric_id": "segment_A_rera_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 25, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_B_rera_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 17, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_rera_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 17, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_D_rera_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 16, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_rera_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 25, + "params": { + "gt_value": "yes" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case33/case33_14.json b/sleep_assessment/cases/case33/case33_14.json new file mode 100644 index 0000000000000000000000000000000000000000..e74e6b0eb28d85487754cb83dc1a4b61b7bf5bc7 --- /dev/null +++ b/sleep_assessment/cases/case33/case33_14.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "instruction": "Given the sleep file, determine whether a respiratory-effort-related arousal (RERA) event is present in each of the following five 60-second sleep segments: A: 60.79-61.79 minutes; B: 17.04-18.04 minutes; C: 50.8-51.8 minutes; D: 159.49-160.49 minutes; E: 243.5-244.5 minutes. Report yes or no for segments A-E in order. Distinguish RERA from apnea, hypopnea, and segments without a target respiratory event. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract whether RERA is present in segments A, B, C, D, and E.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n4. Each value must be exactly \"yes\", \"no\", or null.\n5. Normalize affirmative RERA conclusions to \"yes\" and absent/no-RERA conclusions to \"no\".\n6. If a segment is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"yes\",\"B\":\"no\",\"C\":\"no\",\"D\":\"yes\",\"E\":\"no\"}", + "metrics": [ + { + "metric_id": "segment_A_rera_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 17, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_B_rera_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 17, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_C_rera_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 25, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_D_rera_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 25, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_E_rera_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 16, + "params": { + "gt_value": "no" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case33/case33_15.json b/sleep_assessment/cases/case33/case33_15.json new file mode 100644 index 0000000000000000000000000000000000000000..649b65f7bbac6e500ba4e3c00487cd38399ad1ab --- /dev/null +++ b/sleep_assessment/cases/case33/case33_15.json @@ -0,0 +1,62 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "instruction": "Given the sleep file, determine whether a respiratory-effort-related arousal (RERA) event is present in each of the following five 60-second sleep segments: A: 178.75-179.75 minutes; B: 330.02-331.02 minutes; C: 183.95-184.95 minutes; D: 173.98-174.98 minutes; E: 381.5-382.5 minutes. Report yes or no for segments A-E in order. Distinguish RERA from apnea, hypopnea, and segments without a target respiratory event. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract whether RERA is present in segments A, B, C, D, and E.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\", \"B\", \"C\", \"D\", and \"E\".\n4. Each value must be exactly \"yes\", \"no\", or null.\n5. Normalize affirmative RERA conclusions to \"yes\" and absent/no-RERA conclusions to \"no\".\n6. If a segment is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"yes\",\"B\":\"no\",\"C\":\"no\",\"D\":\"yes\",\"E\":\"no\"}", + "metrics": [ + { + "metric_id": "segment_A_rera_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 25, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_B_rera_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 25, + "params": { + "gt_value": "yes" + } + }, + { + "metric_id": "segment_C_rera_accuracy", + "type": "categorical_check", + "target_key": "C", + "weight": 17, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_D_rera_accuracy", + "type": "categorical_check", + "target_key": "D", + "weight": 17, + "params": { + "gt_value": "no" + } + }, + { + "metric_id": "segment_E_rera_accuracy", + "type": "categorical_check", + "target_key": "E", + "weight": 16, + "params": { + "gt_value": "no" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case34/case34_01.json b/sleep_assessment/cases/case34/case34_01.json new file mode 100644 index 0000000000000000000000000000000000000000..48b7b598ad843868e36bc4a86f6f41a8c9bd738b --- /dev/null +++ b/sleep_assessment/cases/case34/case34_01.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "label_path": "data/sleep/Physionet2018_01.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to identify the apnea event with the longest duration across the whole recording. Report its duration in seconds, its start time in minutes from the beginning of the recording, and the sleep stage at the event start. The sleep-stage output must be one of W, N1, N2, N3, or R. Respiratory channels: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the duration in seconds, start time in minutes, and sleep stage of the longest apnea event.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"longest_apnea_duration_sec\", \"longest_apnea_start_min\", and \"sleep_stage\".\n4. The two numeric values must be FLOATS or null.\n5. \"sleep_stage\" must be exactly W, N1, N2, N3, R, or null.\n6. Normalize REM to R.\n7. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_apnea_duration_sec\": , \"longest_apnea_start_min\": , \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "longest_apnea_duration_accuracy", + "type": "numeric_check", + "target_key": "longest_apnea_duration_sec", + "weight": 40, + "params": { + "gt_value": 40.70000000000073, + "tolerance": 1.0 + } + }, + { + "metric_id": "longest_apnea_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_apnea_start_min", + "weight": 40, + "params": { + "gt_value": 224.34666666666666, + "tolerance": 1.0 + } + }, + { + "metric_id": "longest_apnea_sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 20, + "params": { + "gt_value": "N2" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case34/case34_02.json b/sleep_assessment/cases/case34/case34_02.json new file mode 100644 index 0000000000000000000000000000000000000000..bc0b401c6a9d5904a57e8af84845792c46ed08e2 --- /dev/null +++ b/sleep_assessment/cases/case34/case34_02.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "label_path": "data/sleep/Physionet2018_02.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to identify the apnea event with the longest duration across the whole recording. Report its duration in seconds, its start time in minutes from the beginning of the recording, and the sleep stage at the event start. The sleep-stage output must be one of W, N1, N2, N3, or R. Respiratory channels: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the duration in seconds, start time in minutes, and sleep stage of the longest apnea event.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"longest_apnea_duration_sec\", \"longest_apnea_start_min\", and \"sleep_stage\".\n4. The two numeric values must be FLOATS or null.\n5. \"sleep_stage\" must be exactly W, N1, N2, N3, R, or null.\n6. Normalize REM to R.\n7. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_apnea_duration_sec\": , \"longest_apnea_start_min\": , \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "longest_apnea_duration_accuracy", + "type": "numeric_check", + "target_key": "longest_apnea_duration_sec", + "weight": 40, + "params": { + "gt_value": 26.100000000000364, + "tolerance": 1.0 + } + }, + { + "metric_id": "longest_apnea_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_apnea_start_min", + "weight": 40, + "params": { + "gt_value": 250.91833333333335, + "tolerance": 1.0 + } + }, + { + "metric_id": "longest_apnea_sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 20, + "params": { + "gt_value": "N2" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case34/case34_03.json b/sleep_assessment/cases/case34/case34_03.json new file mode 100644 index 0000000000000000000000000000000000000000..df536a98b980c5227bbe312e112c10482ac4b6e5 --- /dev/null +++ b/sleep_assessment/cases/case34/case34_03.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "label_path": "data/sleep/Physionet2018_04.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to identify the apnea event with the longest duration across the whole recording. Report its duration in seconds, its start time in minutes from the beginning of the recording, and the sleep stage at the event start. The sleep-stage output must be one of W, N1, N2, N3, or R. Respiratory channels: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the duration in seconds, start time in minutes, and sleep stage of the longest apnea event.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"longest_apnea_duration_sec\", \"longest_apnea_start_min\", and \"sleep_stage\".\n4. The two numeric values must be FLOATS or null.\n5. \"sleep_stage\" must be exactly W, N1, N2, N3, R, or null.\n6. Normalize REM to R.\n7. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_apnea_duration_sec\": , \"longest_apnea_start_min\": , \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "longest_apnea_duration_accuracy", + "type": "numeric_check", + "target_key": "longest_apnea_duration_sec", + "weight": 40, + "params": { + "gt_value": 24.5, + "tolerance": 1.0 + } + }, + { + "metric_id": "longest_apnea_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_apnea_start_min", + "weight": 40, + "params": { + "gt_value": 70.6, + "tolerance": 1.0 + } + }, + { + "metric_id": "longest_apnea_sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 20, + "params": { + "gt_value": "R" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case34/case34_04.json b/sleep_assessment/cases/case34/case34_04.json new file mode 100644 index 0000000000000000000000000000000000000000..ec66a85219b975fee97b6692c20a5120e4650c9a --- /dev/null +++ b/sleep_assessment/cases/case34/case34_04.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "label_path": "data/sleep/Physionet2018_05.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to identify the apnea event with the longest duration across the whole recording. Report its duration in seconds, its start time in minutes from the beginning of the recording, and the sleep stage at the event start. The sleep-stage output must be one of W, N1, N2, N3, or R. Respiratory channels: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the duration in seconds, start time in minutes, and sleep stage of the longest apnea event.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"longest_apnea_duration_sec\", \"longest_apnea_start_min\", and \"sleep_stage\".\n4. The two numeric values must be FLOATS or null.\n5. \"sleep_stage\" must be exactly W, N1, N2, N3, R, or null.\n6. Normalize REM to R.\n7. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_apnea_duration_sec\": , \"longest_apnea_start_min\": , \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "longest_apnea_duration_accuracy", + "type": "numeric_check", + "target_key": "longest_apnea_duration_sec", + "weight": 40, + "params": { + "gt_value": 30.900000000000546, + "tolerance": 1.0 + } + }, + { + "metric_id": "longest_apnea_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_apnea_start_min", + "weight": 40, + "params": { + "gt_value": 114.90816666666666, + "tolerance": 1.0 + } + }, + { + "metric_id": "longest_apnea_sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 20, + "params": { + "gt_value": "N2" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case34/case34_05.json b/sleep_assessment/cases/case34/case34_05.json new file mode 100644 index 0000000000000000000000000000000000000000..df2aa9410f57853f8141bcfae7cebd710f848115 --- /dev/null +++ b/sleep_assessment/cases/case34/case34_05.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "label_path": "data/sleep/SHHS1_01.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to identify the apnea event with the longest duration across the whole recording. Report its duration in seconds, its start time in minutes from the beginning of the recording, and the sleep stage at the event start. The sleep-stage output must be one of W, N1, N2, N3, or R. Respiratory channels: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the duration in seconds, start time in minutes, and sleep stage of the longest apnea event.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"longest_apnea_duration_sec\", \"longest_apnea_start_min\", and \"sleep_stage\".\n4. The two numeric values must be FLOATS or null.\n5. \"sleep_stage\" must be exactly W, N1, N2, N3, R, or null.\n6. Normalize REM to R.\n7. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_apnea_duration_sec\": , \"longest_apnea_start_min\": , \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "longest_apnea_duration_accuracy", + "type": "numeric_check", + "target_key": "longest_apnea_duration_sec", + "weight": 40, + "params": { + "gt_value": 39.399999999999636, + "tolerance": 1.0 + } + }, + { + "metric_id": "longest_apnea_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_apnea_start_min", + "weight": 40, + "params": { + "gt_value": 135.77333333333334, + "tolerance": 1.0 + } + }, + { + "metric_id": "longest_apnea_sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 20, + "params": { + "gt_value": "R" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case34/case34_06.json b/sleep_assessment/cases/case34/case34_06.json new file mode 100644 index 0000000000000000000000000000000000000000..8848f1689bdd4cd280fbbdd1e99a6821793a6087 --- /dev/null +++ b/sleep_assessment/cases/case34/case34_06.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "label_path": "data/sleep/SHHS1_02.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to identify the apnea event with the longest duration across the whole recording. Report its duration in seconds, its start time in minutes from the beginning of the recording, and the sleep stage at the event start. The sleep-stage output must be one of W, N1, N2, N3, or R. Respiratory channels: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the duration in seconds, start time in minutes, and sleep stage of the longest apnea event.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"longest_apnea_duration_sec\", \"longest_apnea_start_min\", and \"sleep_stage\".\n4. The two numeric values must be FLOATS or null.\n5. \"sleep_stage\" must be exactly W, N1, N2, N3, R, or null.\n6. Normalize REM to R.\n7. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_apnea_duration_sec\": , \"longest_apnea_start_min\": , \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "longest_apnea_duration_accuracy", + "type": "numeric_check", + "target_key": "longest_apnea_duration_sec", + "weight": 40, + "params": { + "gt_value": 58.70000000000073, + "tolerance": 1.0 + } + }, + { + "metric_id": "longest_apnea_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_apnea_start_min", + "weight": 40, + "params": { + "gt_value": 145.27333333333334, + "tolerance": 1.0 + } + }, + { + "metric_id": "longest_apnea_sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 20, + "params": { + "gt_value": "R" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case34/case34_07.json b/sleep_assessment/cases/case34/case34_07.json new file mode 100644 index 0000000000000000000000000000000000000000..8552d3b5315c9d8702c94bfc4b8b031ddf2894f8 --- /dev/null +++ b/sleep_assessment/cases/case34/case34_07.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "label_path": "data/sleep/SHHS1_03.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to identify the apnea event with the longest duration across the whole recording. Report its duration in seconds, its start time in minutes from the beginning of the recording, and the sleep stage at the event start. The sleep-stage output must be one of W, N1, N2, N3, or R. Respiratory channels: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the duration in seconds, start time in minutes, and sleep stage of the longest apnea event.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"longest_apnea_duration_sec\", \"longest_apnea_start_min\", and \"sleep_stage\".\n4. The two numeric values must be FLOATS or null.\n5. \"sleep_stage\" must be exactly W, N1, N2, N3, R, or null.\n6. Normalize REM to R.\n7. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_apnea_duration_sec\": , \"longest_apnea_start_min\": , \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "longest_apnea_duration_accuracy", + "type": "numeric_check", + "target_key": "longest_apnea_duration_sec", + "weight": 40, + "params": { + "gt_value": 31.299999999999272, + "tolerance": 1.0 + } + }, + { + "metric_id": "longest_apnea_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_apnea_start_min", + "weight": 40, + "params": { + "gt_value": 231.87833333333336, + "tolerance": 1.0 + } + }, + { + "metric_id": "longest_apnea_sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 20, + "params": { + "gt_value": "N2" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case34/case34_08.json b/sleep_assessment/cases/case34/case34_08.json new file mode 100644 index 0000000000000000000000000000000000000000..5ed2d2b775188ef046b66619d7c9088e573ac3a5 --- /dev/null +++ b/sleep_assessment/cases/case34/case34_08.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "label_path": "data/sleep/SHHS1_04.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to identify the apnea event with the longest duration across the whole recording. Report its duration in seconds, its start time in minutes from the beginning of the recording, and the sleep stage at the event start. The sleep-stage output must be one of W, N1, N2, N3, or R. Respiratory channels: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the duration in seconds, start time in minutes, and sleep stage of the longest apnea event.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"longest_apnea_duration_sec\", \"longest_apnea_start_min\", and \"sleep_stage\".\n4. The two numeric values must be FLOATS or null.\n5. \"sleep_stage\" must be exactly W, N1, N2, N3, R, or null.\n6. Normalize REM to R.\n7. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_apnea_duration_sec\": , \"longest_apnea_start_min\": , \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "longest_apnea_duration_accuracy", + "type": "numeric_check", + "target_key": "longest_apnea_duration_sec", + "weight": 40, + "params": { + "gt_value": 38.0, + "tolerance": 1.0 + } + }, + { + "metric_id": "longest_apnea_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_apnea_start_min", + "weight": 40, + "params": { + "gt_value": 384.98333333333335, + "tolerance": 1.0 + } + }, + { + "metric_id": "longest_apnea_sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 20, + "params": { + "gt_value": "R" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case34/case34_09.json b/sleep_assessment/cases/case34/case34_09.json new file mode 100644 index 0000000000000000000000000000000000000000..d572745cd0d18632f89e20e9855aa363790da9c2 --- /dev/null +++ b/sleep_assessment/cases/case34/case34_09.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "label_path": "data/sleep/SHHS1_05.npy", + "instruction": "Please use the provided sleep EEG/PSG file and paired sleep-stage label file to identify the apnea event with the longest duration across the whole recording. Report its duration in seconds, its start time in minutes from the beginning of the recording, and the sleep stage at the event start. The sleep-stage output must be one of W, N1, N2, N3, or R. Respiratory channels: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the duration in seconds, start time in minutes, and sleep stage of the longest apnea event.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"longest_apnea_duration_sec\", \"longest_apnea_start_min\", and \"sleep_stage\".\n4. The two numeric values must be FLOATS or null.\n5. \"sleep_stage\" must be exactly W, N1, N2, N3, R, or null.\n6. Normalize REM to R.\n7. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"longest_apnea_duration_sec\": , \"longest_apnea_start_min\": , \"sleep_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|null>}", + "metrics": [ + { + "metric_id": "longest_apnea_duration_accuracy", + "type": "numeric_check", + "target_key": "longest_apnea_duration_sec", + "weight": 40, + "params": { + "gt_value": 23.399999999999636, + "tolerance": 1.0 + } + }, + { + "metric_id": "longest_apnea_start_time_accuracy", + "type": "numeric_check", + "target_key": "longest_apnea_start_min", + "weight": 40, + "params": { + "gt_value": 146.82999999999998, + "tolerance": 1.0 + } + }, + { + "metric_id": "longest_apnea_sleep_stage_accuracy", + "type": "categorical_check", + "target_key": "sleep_stage", + "weight": 20, + "params": { + "gt_value": "R" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_01.json b/sleep_assessment/cases/case35/case35_01.json new file mode 100644 index 0000000000000000000000000000000000000000..c0e78fa293d73e0263cb04666bcf6a5d757f38c3 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_01.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 50.75-51.75 minutes; B: 137.25-138.25 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_02.json b/sleep_assessment/cases/case35/case35_02.json new file mode 100644 index 0000000000000000000000000000000000000000..531ade8db4a6417b5b412f30a677e9ac0bbe434c --- /dev/null +++ b/sleep_assessment/cases/case35/case35_02.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 261-262 minutes; B: 53.75-54.75 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "MA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "CA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_03.json b/sleep_assessment/cases/case35/case35_03.json new file mode 100644 index 0000000000000000000000000000000000000000..16cdcaa8572da90b9f2da09144edc073baf6e368 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_03.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 167.25-168.25 minutes; B: 263.5-264.5 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "OA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "MA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_04.json b/sleep_assessment/cases/case35/case35_04.json new file mode 100644 index 0000000000000000000000000000000000000000..8d4c41be0ba2f07b18ce35567b1f0ca6a133c988 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_04.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 210.75-211.75 minutes; B: 172.75-173.75 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_05.json b/sleep_assessment/cases/case35/case35_05.json new file mode 100644 index 0000000000000000000000000000000000000000..527ae8b3a1645f4ac7ae12a2a786aad2a1cfc02f --- /dev/null +++ b/sleep_assessment/cases/case35/case35_05.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 264-265 minutes; B: 213.75-214.75 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "MA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "CA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_06.json b/sleep_assessment/cases/case35/case35_06.json new file mode 100644 index 0000000000000000000000000000000000000000..f19053ca0c205f85c6feb3544896e062ccc7ffc9 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_06.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 182.75-183.75 minutes; B: 265.5-266.5 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "OA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "MA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_07.json b/sleep_assessment/cases/case35/case35_07.json new file mode 100644 index 0000000000000000000000000000000000000000..de968352b5b959bc9608f2aa807bd857b10ca10e --- /dev/null +++ b/sleep_assessment/cases/case35/case35_07.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 225.25-226.25 minutes; B: 192.25-193.25 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_08.json b/sleep_assessment/cases/case35/case35_08.json new file mode 100644 index 0000000000000000000000000000000000000000..d3ba676d0ab2c1d499450d2092017ce32676d6d4 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_08.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 267-268 minutes; B: 236.75-237.75 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "MA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "CA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_09.json b/sleep_assessment/cases/case35/case35_09.json new file mode 100644 index 0000000000000000000000000000000000000000..4004e660612bb683996c02aa64664fc24144154a --- /dev/null +++ b/sleep_assessment/cases/case35/case35_09.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 198.75-199.75 minutes; B: 271-272 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "OA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "MA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_10.json b/sleep_assessment/cases/case35/case35_10.json new file mode 100644 index 0000000000000000000000000000000000000000..141434999e98d46d788de2fe6bb12263f337f45e --- /dev/null +++ b/sleep_assessment/cases/case35/case35_10.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 240.75-241.75 minutes; B: 346.75-347.75 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_11.json b/sleep_assessment/cases/case35/case35_11.json new file mode 100644 index 0000000000000000000000000000000000000000..8f34c815ed5d7cb3950b83141c8a8393e6455c04 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_11.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "label_path": "data/sleep/ISRUC_03.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 168.25-169.25 minutes; B: 372.75-373.75 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "CA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_12.json b/sleep_assessment/cases/case35/case35_12.json new file mode 100644 index 0000000000000000000000000000000000000000..8cdf1a1f26218b812e02a0d2b59c70554f358ad5 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_12.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "label_path": "data/sleep/ISRUC_05.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 355.75-356.75 minutes; B: 361.75-362.75 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "OA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_13.json b/sleep_assessment/cases/case35/case35_13.json new file mode 100644 index 0000000000000000000000000000000000000000..825f5f5aa516a28e2e34e20cd8105479386d3997 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_13.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "label_path": "data/sleep/Physionet2018_01.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 4.0053-5.0053 minutes; B: 108.0262-109.0262 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_14.json b/sleep_assessment/cases/case35/case35_14.json new file mode 100644 index 0000000000000000000000000000000000000000..abc02e86d1b93bb9adf5e09e319c398d21bf5895 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_14.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "label_path": "data/sleep/Physionet2018_01.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 8.0337-9.0337 minutes; B: 125.9338-126.9338 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_15.json b/sleep_assessment/cases/case35/case35_15.json new file mode 100644 index 0000000000000000000000000000000000000000..e251a15d3c490ab6d72f62d105f83245d27c12fb --- /dev/null +++ b/sleep_assessment/cases/case35/case35_15.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "label_path": "data/sleep/Physionet2018_01.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 8.883-9.883 minutes; B: 134.8967-135.8967 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_16.json b/sleep_assessment/cases/case35/case35_16.json new file mode 100644 index 0000000000000000000000000000000000000000..080f4547e2f3e379054d30381fd972e6b4d2b023 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_16.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "label_path": "data/sleep/Physionet2018_01.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 18.4128-19.4128 minutes; B: 136.317-137.317 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_17.json b/sleep_assessment/cases/case35/case35_17.json new file mode 100644 index 0000000000000000000000000000000000000000..4e7b3184a28db97866ae89b91401f60f8ebd9317 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_17.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "label_path": "data/sleep/Physionet2018_01.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 21.4165-22.4165 minutes; B: 137.6913-138.6913 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_18.json b/sleep_assessment/cases/case35/case35_18.json new file mode 100644 index 0000000000000000000000000000000000000000..844049783e54ec88418694a0df61b1c0efcec256 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_18.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "label_path": "data/sleep/Physionet2018_01.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 130.2753-131.2753 minutes; B: 138.5528-139.5528 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_19.json b/sleep_assessment/cases/case35/case35_19.json new file mode 100644 index 0000000000000000000000000000000000000000..e1c9b817262ecef86477c8503e6326bc733c8183 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_19.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "label_path": "data/sleep/Physionet2018_01.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 208.4252-209.4252 minutes; B: 222.8902-223.8902 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_20.json b/sleep_assessment/cases/case35/case35_20.json new file mode 100644 index 0000000000000000000000000000000000000000..5eb9a8376e6ff57c1d96786b79caf113987d31e6 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_20.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "label_path": "data/sleep/Physionet2018_01.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 266.2477-267.2477 minutes; B: 231.1058-232.1058 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_21.json b/sleep_assessment/cases/case35/case35_21.json new file mode 100644 index 0000000000000000000000000000000000000000..38d13dd7c8222cf62b0d1c58c3e9494c4c9ecfa2 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_21.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "label_path": "data/sleep/Physionet2018_02.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 158.2463-159.2463 minutes; B: 13.1352-14.1352 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_22.json b/sleep_assessment/cases/case35/case35_22.json new file mode 100644 index 0000000000000000000000000000000000000000..609da6b9b25f1e72c10b0d897ef8ed959f92a775 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_22.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "label_path": "data/sleep/Physionet2018_02.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 220.9055-221.9055 minutes; B: 17.0533-18.0533 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_23.json b/sleep_assessment/cases/case35/case35_23.json new file mode 100644 index 0000000000000000000000000000000000000000..3d618091363bf3f285f639d9fce291aae4518946 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_23.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "label_path": "data/sleep/Physionet2018_02.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 224.4773-225.4773 minutes; B: 31.8415-32.8415 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_24.json b/sleep_assessment/cases/case35/case35_24.json new file mode 100644 index 0000000000000000000000000000000000000000..3e4cf625eb9a1007aab0f9c68a23051abe80df02 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_24.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "label_path": "data/sleep/Physionet2018_02.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 225.3435-226.3435 minutes; B: 41.4067-42.4067 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_25.json b/sleep_assessment/cases/case35/case35_25.json new file mode 100644 index 0000000000000000000000000000000000000000..0293c4a7c8e89b34a14bb44f84a70b115a9e2ba1 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_25.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "label_path": "data/sleep/Physionet2018_02.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 226.4082-227.4082 minutes; B: 117.7177-118.7177 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_26.json b/sleep_assessment/cases/case35/case35_26.json new file mode 100644 index 0000000000000000000000000000000000000000..a9e1e84d93581532ac64e7ef4c4334d6287aa821 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_26.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "label_path": "data/sleep/Physionet2018_02.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 237.0553-238.0553 minutes; B: 147.5847-148.5847 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_27.json b/sleep_assessment/cases/case35/case35_27.json new file mode 100644 index 0000000000000000000000000000000000000000..52145be00ae1a7c4e2bf73c0a7643d0214af135c --- /dev/null +++ b/sleep_assessment/cases/case35/case35_27.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "label_path": "data/sleep/Physionet2018_02.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 237.7278-238.7278 minutes; B: 154.952-155.952 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_28.json b/sleep_assessment/cases/case35/case35_28.json new file mode 100644 index 0000000000000000000000000000000000000000..7c6a35f56ab9dcc9f940a26cd91e37ee0515bc46 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_28.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "label_path": "data/sleep/Physionet2018_02.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 189.9237-190.9237 minutes; B: 243.536-244.536 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "OA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_29.json b/sleep_assessment/cases/case35/case35_29.json new file mode 100644 index 0000000000000000000000000000000000000000..783d09461e494f538241d3ff8d2951048d1bafd3 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_29.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "label_path": "data/sleep/Physionet2018_02.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 311.2957-312.2957 minutes; B: 348.2958-349.2958 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "OA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_30.json b/sleep_assessment/cases/case35/case35_30.json new file mode 100644 index 0000000000000000000000000000000000000000..01c12976026771339ffca4066739803cedc0b6c7 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_30.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "label_path": "data/sleep/Physionet2018_04.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 8.8767-9.8767 minutes; B: 3.1475-4.1475 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_31.json b/sleep_assessment/cases/case35/case35_31.json new file mode 100644 index 0000000000000000000000000000000000000000..216fbe945e9b29b1b2dcf2cd8b839789a454ca71 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_31.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "label_path": "data/sleep/Physionet2018_04.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 13.4398-14.4398 minutes; B: 4.93-5.93 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_32.json b/sleep_assessment/cases/case35/case35_32.json new file mode 100644 index 0000000000000000000000000000000000000000..1680d12d4f63e7044c13dba927bcbb207caf0417 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_32.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "label_path": "data/sleep/Physionet2018_04.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 19.1613-20.1613 minutes; B: 18.3727-19.3727 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_33.json b/sleep_assessment/cases/case35/case35_33.json new file mode 100644 index 0000000000000000000000000000000000000000..6607600989bf42cee4319a35cb0aab1e30a9ccb6 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_33.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "label_path": "data/sleep/Physionet2018_04.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 20.998-21.998 minutes; B: 21.7163-22.7163 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_34.json b/sleep_assessment/cases/case35/case35_34.json new file mode 100644 index 0000000000000000000000000000000000000000..1b037a7828afcc67637e7df75e40f52f44288e30 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_34.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "label_path": "data/sleep/Physionet2018_04.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 23.7058-24.7058 minutes; B: 33.8285-34.8285 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_35.json b/sleep_assessment/cases/case35/case35_35.json new file mode 100644 index 0000000000000000000000000000000000000000..9a82fb1d11e4b4239b7927d7a452bca28f3c54ed --- /dev/null +++ b/sleep_assessment/cases/case35/case35_35.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "label_path": "data/sleep/Physionet2018_04.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 24.3633-25.3633 minutes; B: 70.3042-71.3042 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_36.json b/sleep_assessment/cases/case35/case35_36.json new file mode 100644 index 0000000000000000000000000000000000000000..96dd4d6eaa381fc3bd7a98177b3d7b2942bb647f --- /dev/null +++ b/sleep_assessment/cases/case35/case35_36.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "label_path": "data/sleep/Physionet2018_04.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 65.6665-66.6665 minutes; B: 107.9857-108.9857 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_37.json b/sleep_assessment/cases/case35/case35_37.json new file mode 100644 index 0000000000000000000000000000000000000000..cfaeeff8f056d283afbece68880b6df9b4f5165c --- /dev/null +++ b/sleep_assessment/cases/case35/case35_37.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "label_path": "data/sleep/Physionet2018_04.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 146.6173-147.6173 minutes; B: 207.053-208.053 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_38.json b/sleep_assessment/cases/case35/case35_38.json new file mode 100644 index 0000000000000000000000000000000000000000..9ec9c34bf1d132b5b6a78a8f2b48330f3f611b12 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_38.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "label_path": "data/sleep/Physionet2018_04.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 243.7145-244.7145 minutes; B: 363.5872-364.5872 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_39.json b/sleep_assessment/cases/case35/case35_39.json new file mode 100644 index 0000000000000000000000000000000000000000..8d21edcca344dbaf1c9a05d5dc5c4ba74ad1329b --- /dev/null +++ b/sleep_assessment/cases/case35/case35_39.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "label_path": "data/sleep/Physionet2018_05.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 176.952-177.952 minutes; B: 1.4788-2.4788 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_40.json b/sleep_assessment/cases/case35/case35_40.json new file mode 100644 index 0000000000000000000000000000000000000000..0fa14b9c63c95fedf48a29899f68aa3af9029d0a --- /dev/null +++ b/sleep_assessment/cases/case35/case35_40.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "label_path": "data/sleep/Physionet2018_05.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 271.4716-272.4716 minutes; B: 19.8838-20.8838 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_41.json b/sleep_assessment/cases/case35/case35_41.json new file mode 100644 index 0000000000000000000000000000000000000000..6c79afd26e970fa5adb192598c5f0557fba96037 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_41.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "label_path": "data/sleep/Physionet2018_05.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 377.8767-378.8767 minutes; B: 53.7885-54.7885 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "CA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_42.json b/sleep_assessment/cases/case35/case35_42.json new file mode 100644 index 0000000000000000000000000000000000000000..c1cac8cb27a20eefa4c96b52748c72201fbabc9d --- /dev/null +++ b/sleep_assessment/cases/case35/case35_42.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "label_path": "data/sleep/Physionet2018_05.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 67.2233-68.2233 minutes; B: 80.6448-81.6448 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "OA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_43.json b/sleep_assessment/cases/case35/case35_43.json new file mode 100644 index 0000000000000000000000000000000000000000..92417f34b18382c70dc7b6ab9f77d41619131f50 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_43.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "label_path": "data/sleep/Physionet2018_05.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 95.7438-96.7438 minutes; B: 108.6482-109.6482 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "OA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_44.json b/sleep_assessment/cases/case35/case35_44.json new file mode 100644 index 0000000000000000000000000000000000000000..cfa4e900589bd6d7a76b1b594e0f515e58afbcaa --- /dev/null +++ b/sleep_assessment/cases/case35/case35_44.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "label_path": "data/sleep/Physionet2018_05.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 116.554-117.554 minutes; B: 120.8507-121.8507 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "OA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_45.json b/sleep_assessment/cases/case35/case35_45.json new file mode 100644 index 0000000000000000000000000000000000000000..13024e1f2d24edc7a3469cb672d25d3c65df9a4f --- /dev/null +++ b/sleep_assessment/cases/case35/case35_45.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "label_path": "data/sleep/Physionet2018_05.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 127.9563-128.9563 minutes; B: 135.8412-136.8412 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "OA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case35/case35_46.json b/sleep_assessment/cases/case35/case35_46.json new file mode 100644 index 0000000000000000000000000000000000000000..691f2082533826809b98f01dd5c008ee81b50787 --- /dev/null +++ b/sleep_assessment/cases/case35/case35_46.json @@ -0,0 +1,36 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "label_path": "data/sleep/Physionet2018_05.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, determine the specific apnea subtype occurring in each 60-second segment: A: 159.0997-160.0997 minutes; B: 198.173-199.173 minutes. Return one subtype for each segment using only CA, OA, or MA, where CA = Central apnea, OA = Obstructive apnea, and MA = Mixed apnea. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final apnea subtype for segments A and B.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, or extra fields.\n3. The keys must be EXACTLY \"A\" and \"B\".\n4. Each value must be exactly \"CA\", \"OA\", \"MA\", or null.\n5. Normalize Central apnea to CA, Obstructive apnea to OA, and Mixed apnea to MA.\n6. If a segment subtype is missing or unclear, return null for that segment.\n\n### OUTPUT TEMPLATE\n{\"A\":\"CA\",\"B\":\"OA\"}", + "metrics": [ + { + "metric_id": "apnea_type_A_accuracy", + "type": "categorical_check", + "target_key": "A", + "weight": 50, + "params": { + "gt_value": "OA" + } + }, + { + "metric_id": "apnea_type_B_accuracy", + "type": "categorical_check", + "target_key": "B", + "weight": 50, + "params": { + "gt_value": "OA" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case36/case36_01.json b/sleep_assessment/cases/case36/case36_01.json new file mode 100644 index 0000000000000000000000000000000000000000..14123efc814f09a36ffad4e4061e71cdc6d2789c --- /dev/null +++ b/sleep_assessment/cases/case36/case36_01.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate stage-specific Apnea-Hypopnea Indices for the whole recording. Assign each apnea or hypopnea event to the sleep stage at the event start. REM-AHI is the number of apnea plus hypopnea events starting during R/REM divided by REM sleep time in hours. NREM-AHI is the number starting during N1, N2, or N3 divided by NREM sleep time in hours. Events starting during Wake are excluded from both indices. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort. Report the final REM-AHI and NREM-AHI as events per hour." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final REM-specific AHI and NREM-specific AHI.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"rem_ahi\" and \"nrem_ahi\".\n4. Each value must be a FLOAT or null.\n5. Normalize REM AHI to rem_ahi and NREM AHI to nrem_ahi.\n6. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"rem_ahi\": , \"nrem_ahi\": }", + "metrics": [ + { + "metric_id": "rem_apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "rem_ahi", + "weight": 50, + "params": { + "gt_value": 41.69491525423729, + "tolerance": 4.169491525423729 + } + }, + { + "metric_id": "nrem_apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "nrem_ahi", + "weight": 50, + "params": { + "gt_value": 29.63855421686747, + "tolerance": 2.963855421686747 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case36/case36_02.json b/sleep_assessment/cases/case36/case36_02.json new file mode 100644 index 0000000000000000000000000000000000000000..75e14780dcbacf2002547826a414288818c9b2da --- /dev/null +++ b/sleep_assessment/cases/case36/case36_02.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "label_path": "data/sleep/ISRUC_02.npy", + "instruction": "Use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate stage-specific Apnea-Hypopnea Indices for the whole recording. Assign each apnea or hypopnea event to the sleep stage at the event start. REM-AHI is the number of apnea plus hypopnea events starting during R/REM divided by REM sleep time in hours. NREM-AHI is the number starting during N1, N2, or N3 divided by NREM sleep time in hours. Events starting during Wake are excluded from both indices. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort. Report the final REM-AHI and NREM-AHI as events per hour." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final REM-specific AHI and NREM-specific AHI.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"rem_ahi\" and \"nrem_ahi\".\n4. Each value must be a FLOAT or null.\n5. Normalize REM AHI to rem_ahi and NREM AHI to nrem_ahi.\n6. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"rem_ahi\": , \"nrem_ahi\": }", + "metrics": [ + { + "metric_id": "rem_apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "rem_ahi", + "weight": 50, + "params": { + "gt_value": 2.3529411764705883, + "tolerance": 0.23529411764705885 + } + }, + { + "metric_id": "nrem_apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "nrem_ahi", + "weight": 50, + "params": { + "gt_value": 6.0291734197730955, + "tolerance": 0.6029173419773096 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case36/case36_03.json b/sleep_assessment/cases/case36/case36_03.json new file mode 100644 index 0000000000000000000000000000000000000000..3dcb564a8b04247d4e73cb9a01a0a4bfda2ad094 --- /dev/null +++ b/sleep_assessment/cases/case36/case36_03.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "label_path": "data/sleep/ISRUC_03.npy", + "instruction": "Use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate stage-specific Apnea-Hypopnea Indices for the whole recording. Assign each apnea or hypopnea event to the sleep stage at the event start. REM-AHI is the number of apnea plus hypopnea events starting during R/REM divided by REM sleep time in hours. NREM-AHI is the number starting during N1, N2, or N3 divided by NREM sleep time in hours. Events starting during Wake are excluded from both indices. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort. Report the final REM-AHI and NREM-AHI as events per hour." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final REM-specific AHI and NREM-specific AHI.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"rem_ahi\" and \"nrem_ahi\".\n4. Each value must be a FLOAT or null.\n5. Normalize REM AHI to rem_ahi and NREM AHI to nrem_ahi.\n6. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"rem_ahi\": , \"nrem_ahi\": }", + "metrics": [ + { + "metric_id": "rem_apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "rem_ahi", + "weight": 50, + "params": { + "gt_value": 2.643171806167401, + "tolerance": 0.2643171806167401 + } + }, + { + "metric_id": "nrem_apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "nrem_ahi", + "weight": 50, + "params": { + "gt_value": 1.4383561643835618, + "tolerance": 0.14383561643835618 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case36/case36_04.json b/sleep_assessment/cases/case36/case36_04.json new file mode 100644 index 0000000000000000000000000000000000000000..dca099be2a446d0154c4ba2d797190c4713f852b --- /dev/null +++ b/sleep_assessment/cases/case36/case36_04.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "ISRUC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_04.edf", + "label_path": "data/sleep/ISRUC_04.npy", + "instruction": "Use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate stage-specific Apnea-Hypopnea Indices for the whole recording. Assign each apnea or hypopnea event to the sleep stage at the event start. REM-AHI is the number of apnea plus hypopnea events starting during R/REM divided by REM sleep time in hours. NREM-AHI is the number starting during N1, N2, or N3 divided by NREM sleep time in hours. Events starting during Wake are excluded from both indices. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort. Report the final REM-AHI and NREM-AHI as events per hour." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final REM-specific AHI and NREM-specific AHI.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"rem_ahi\" and \"nrem_ahi\".\n4. Each value must be a FLOAT or null.\n5. Normalize REM AHI to rem_ahi and NREM AHI to nrem_ahi.\n6. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"rem_ahi\": , \"nrem_ahi\": }", + "metrics": [ + { + "metric_id": "rem_apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "rem_ahi", + "weight": 50, + "params": { + "gt_value": 2.608695652173913, + "tolerance": 0.2608695652173913 + } + }, + { + "metric_id": "nrem_apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "nrem_ahi", + "weight": 50, + "params": { + "gt_value": 0.6808510638297872, + "tolerance": 0.06808510638297872 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case36/case36_05.json b/sleep_assessment/cases/case36/case36_05.json new file mode 100644 index 0000000000000000000000000000000000000000..171837b50c617c608ef73f804f16befd5165fb80 --- /dev/null +++ b/sleep_assessment/cases/case36/case36_05.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "label_path": "data/sleep/ISRUC_05.npy", + "instruction": "Use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate stage-specific Apnea-Hypopnea Indices for the whole recording. Assign each apnea or hypopnea event to the sleep stage at the event start. REM-AHI is the number of apnea plus hypopnea events starting during R/REM divided by REM sleep time in hours. NREM-AHI is the number starting during N1, N2, or N3 divided by NREM sleep time in hours. Events starting during Wake are excluded from both indices. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort. Report the final REM-AHI and NREM-AHI as events per hour." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final REM-specific AHI and NREM-specific AHI.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"rem_ahi\" and \"nrem_ahi\".\n4. Each value must be a FLOAT or null.\n5. Normalize REM AHI to rem_ahi and NREM AHI to nrem_ahi.\n6. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"rem_ahi\": , \"nrem_ahi\": }", + "metrics": [ + { + "metric_id": "rem_apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "rem_ahi", + "weight": 50, + "params": { + "gt_value": 37.142857142857146, + "tolerance": 3.714285714285715 + } + }, + { + "metric_id": "nrem_apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "nrem_ahi", + "weight": 50, + "params": { + "gt_value": 6.480446927374302, + "tolerance": 0.6480446927374303 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case36/case36_06.json b/sleep_assessment/cases/case36/case36_06.json new file mode 100644 index 0000000000000000000000000000000000000000..1879f0282a78c866341e1cfbf1917759338149ec --- /dev/null +++ b/sleep_assessment/cases/case36/case36_06.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "label_path": "data/sleep/SHHS1_01.npy", + "instruction": "Use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate stage-specific Apnea-Hypopnea Indices for the whole recording. Assign each apnea or hypopnea event to the sleep stage at the event start. REM-AHI is the number of apnea plus hypopnea events starting during R/REM divided by REM sleep time in hours. NREM-AHI is the number starting during N1, N2, or N3 divided by NREM sleep time in hours. Events starting during Wake are excluded from both indices. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry. Report the final REM-AHI and NREM-AHI as events per hour." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final REM-specific AHI and NREM-specific AHI.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"rem_ahi\" and \"nrem_ahi\".\n4. Each value must be a FLOAT or null.\n5. Normalize REM AHI to rem_ahi and NREM AHI to nrem_ahi.\n6. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"rem_ahi\": , \"nrem_ahi\": }", + "metrics": [ + { + "metric_id": "rem_apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "rem_ahi", + "weight": 50, + "params": { + "gt_value": 47.05882352941177, + "tolerance": 4.705882352941177 + } + }, + { + "metric_id": "nrem_apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "nrem_ahi", + "weight": 50, + "params": { + "gt_value": 8.320493066255779, + "tolerance": 0.8320493066255779 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case36/case36_07.json b/sleep_assessment/cases/case36/case36_07.json new file mode 100644 index 0000000000000000000000000000000000000000..885142c8ee5de98eb316db666e7648c3ca3c22ba --- /dev/null +++ b/sleep_assessment/cases/case36/case36_07.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "label_path": "data/sleep/SHHS1_02.npy", + "instruction": "Use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate stage-specific Apnea-Hypopnea Indices for the whole recording. Assign each apnea or hypopnea event to the sleep stage at the event start. REM-AHI is the number of apnea plus hypopnea events starting during R/REM divided by REM sleep time in hours. NREM-AHI is the number starting during N1, N2, or N3 divided by NREM sleep time in hours. Events starting during Wake are excluded from both indices. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry. Report the final REM-AHI and NREM-AHI as events per hour." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final REM-specific AHI and NREM-specific AHI.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"rem_ahi\" and \"nrem_ahi\".\n4. Each value must be a FLOAT or null.\n5. Normalize REM AHI to rem_ahi and NREM AHI to nrem_ahi.\n6. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"rem_ahi\": , \"nrem_ahi\": }", + "metrics": [ + { + "metric_id": "rem_apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "rem_ahi", + "weight": 50, + "params": { + "gt_value": 55.08196721311476, + "tolerance": 5.508196721311476 + } + }, + { + "metric_id": "nrem_apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "nrem_ahi", + "weight": 50, + "params": { + "gt_value": 55.84158415841584, + "tolerance": 5.584158415841585 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case36/case36_08.json b/sleep_assessment/cases/case36/case36_08.json new file mode 100644 index 0000000000000000000000000000000000000000..f2733a5d8492d87b7cf041e20bd0aaca0697d8c4 --- /dev/null +++ b/sleep_assessment/cases/case36/case36_08.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "label_path": "data/sleep/SHHS1_03.npy", + "instruction": "Use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate stage-specific Apnea-Hypopnea Indices for the whole recording. Assign each apnea or hypopnea event to the sleep stage at the event start. REM-AHI is the number of apnea plus hypopnea events starting during R/REM divided by REM sleep time in hours. NREM-AHI is the number starting during N1, N2, or N3 divided by NREM sleep time in hours. Events starting during Wake are excluded from both indices. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry. Report the final REM-AHI and NREM-AHI as events per hour." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final REM-specific AHI and NREM-specific AHI.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"rem_ahi\" and \"nrem_ahi\".\n4. Each value must be a FLOAT or null.\n5. Normalize REM AHI to rem_ahi and NREM AHI to nrem_ahi.\n6. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"rem_ahi\": , \"nrem_ahi\": }", + "metrics": [ + { + "metric_id": "rem_apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "rem_ahi", + "weight": 50, + "params": { + "gt_value": 60.0, + "tolerance": 6.0 + } + }, + { + "metric_id": "nrem_apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "nrem_ahi", + "weight": 50, + "params": { + "gt_value": 12.551505546751187, + "tolerance": 1.2551505546751187 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case36/case36_09.json b/sleep_assessment/cases/case36/case36_09.json new file mode 100644 index 0000000000000000000000000000000000000000..8549f0c7687ac756bde751f94160e85d3b5962f8 --- /dev/null +++ b/sleep_assessment/cases/case36/case36_09.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "label_path": "data/sleep/SHHS1_04.npy", + "instruction": "Use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate stage-specific Apnea-Hypopnea Indices for the whole recording. Assign each apnea or hypopnea event to the sleep stage at the event start. REM-AHI is the number of apnea plus hypopnea events starting during R/REM divided by REM sleep time in hours. NREM-AHI is the number starting during N1, N2, or N3 divided by NREM sleep time in hours. Events starting during Wake are excluded from both indices. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry. Report the final REM-AHI and NREM-AHI as events per hour." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final REM-specific AHI and NREM-specific AHI.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"rem_ahi\" and \"nrem_ahi\".\n4. Each value must be a FLOAT or null.\n5. Normalize REM AHI to rem_ahi and NREM AHI to nrem_ahi.\n6. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"rem_ahi\": , \"nrem_ahi\": }", + "metrics": [ + { + "metric_id": "rem_apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "rem_ahi", + "weight": 50, + "params": { + "gt_value": 69.13043478260869, + "tolerance": 6.913043478260869 + } + }, + { + "metric_id": "nrem_apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "nrem_ahi", + "weight": 50, + "params": { + "gt_value": 12.941176470588236, + "tolerance": 1.2941176470588236 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case36/case36_10.json b/sleep_assessment/cases/case36/case36_10.json new file mode 100644 index 0000000000000000000000000000000000000000..749240ceff6e5b219de9f381aa8d2afa4bebbca8 --- /dev/null +++ b/sleep_assessment/cases/case36/case36_10.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "label_path": "data/sleep/SHHS1_05.npy", + "instruction": "Use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate stage-specific Apnea-Hypopnea Indices for the whole recording. Assign each apnea or hypopnea event to the sleep stage at the event start. REM-AHI is the number of apnea plus hypopnea events starting during R/REM divided by REM sleep time in hours. NREM-AHI is the number starting during N1, N2, or N3 divided by NREM sleep time in hours. Events starting during Wake are excluded from both indices. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry. Report the final REM-AHI and NREM-AHI as events per hour." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final REM-specific AHI and NREM-specific AHI.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"rem_ahi\" and \"nrem_ahi\".\n4. Each value must be a FLOAT or null.\n5. Normalize REM AHI to rem_ahi and NREM AHI to nrem_ahi.\n6. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"rem_ahi\": , \"nrem_ahi\": }", + "metrics": [ + { + "metric_id": "rem_apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "rem_ahi", + "weight": 50, + "params": { + "gt_value": 60.659340659340664, + "tolerance": 6.065934065934067 + } + }, + { + "metric_id": "nrem_apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "nrem_ahi", + "weight": 50, + "params": { + "gt_value": 5.731895223420647, + "tolerance": 0.5731895223420648 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case36/case36_11.json b/sleep_assessment/cases/case36/case36_11.json new file mode 100644 index 0000000000000000000000000000000000000000..ca7a165866e6ebf68a8cbb30643f0936c179038b --- /dev/null +++ b/sleep_assessment/cases/case36/case36_11.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "label_path": "data/sleep/Physionet2018_01.npy", + "instruction": "Use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate stage-specific Apnea-Hypopnea Indices for the whole recording. Assign each apnea or hypopnea event to the sleep stage at the event start. REM-AHI is the number of apnea plus hypopnea events starting during R/REM divided by REM sleep time in hours. NREM-AHI is the number starting during N1, N2, or N3 divided by NREM sleep time in hours. Events starting during Wake are excluded from both indices. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry. Report the final REM-AHI and NREM-AHI as events per hour." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final REM-specific AHI and NREM-specific AHI.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"rem_ahi\" and \"nrem_ahi\".\n4. Each value must be a FLOAT or null.\n5. Normalize REM AHI to rem_ahi and NREM AHI to nrem_ahi.\n6. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"rem_ahi\": , \"nrem_ahi\": }", + "metrics": [ + { + "metric_id": "rem_apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "rem_ahi", + "weight": 50, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + }, + { + "metric_id": "nrem_apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "nrem_ahi", + "weight": 50, + "params": { + "gt_value": 10.881612090680102, + "tolerance": 1.0881612090680102 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case36/case36_12.json b/sleep_assessment/cases/case36/case36_12.json new file mode 100644 index 0000000000000000000000000000000000000000..200955e193a4f9fa79d584e59330055ff1a6c7e8 --- /dev/null +++ b/sleep_assessment/cases/case36/case36_12.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "label_path": "data/sleep/Physionet2018_04.npy", + "instruction": "Use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate stage-specific Apnea-Hypopnea Indices for the whole recording. Assign each apnea or hypopnea event to the sleep stage at the event start. REM-AHI is the number of apnea plus hypopnea events starting during R/REM divided by REM sleep time in hours. NREM-AHI is the number starting during N1, N2, or N3 divided by NREM sleep time in hours. Events starting during Wake are excluded from both indices. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry. Report the final REM-AHI and NREM-AHI as events per hour." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final REM-specific AHI and NREM-specific AHI.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"rem_ahi\" and \"nrem_ahi\".\n4. Each value must be a FLOAT or null.\n5. Normalize REM AHI to rem_ahi and NREM AHI to nrem_ahi.\n6. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"rem_ahi\": , \"nrem_ahi\": }", + "metrics": [ + { + "metric_id": "rem_apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "rem_ahi", + "weight": 50, + "params": { + "gt_value": 13.333333333333334, + "tolerance": 1.3333333333333335 + } + }, + { + "metric_id": "nrem_apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "nrem_ahi", + "weight": 50, + "params": { + "gt_value": 31.203566121842496, + "tolerance": 3.1203566121842496 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case36/case36_13.json b/sleep_assessment/cases/case36/case36_13.json new file mode 100644 index 0000000000000000000000000000000000000000..c7fc0fad2145d918d56d4a1cf6c7eec42abb987e --- /dev/null +++ b/sleep_assessment/cases/case36/case36_13.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "label_path": "data/sleep/Physionet2018_05.npy", + "instruction": "Use the provided sleep EEG/PSG file and paired sleep-stage label file to calculate stage-specific Apnea-Hypopnea Indices for the whole recording. Assign each apnea or hypopnea event to the sleep stage at the event start. REM-AHI is the number of apnea plus hypopnea events starting during R/REM divided by REM sleep time in hours. NREM-AHI is the number starting during N1, N2, or N3 divided by NREM sleep time in hours. Events starting during Wake are excluded from both indices. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry. Report the final REM-AHI and NREM-AHI as events per hour." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final REM-specific AHI and NREM-specific AHI.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"rem_ahi\" and \"nrem_ahi\".\n4. Each value must be a FLOAT or null.\n5. Normalize REM AHI to rem_ahi and NREM AHI to nrem_ahi.\n6. If a value is missing or unclear, return null for that value.\n\n### OUTPUT TEMPLATE\n{\"rem_ahi\": , \"nrem_ahi\": }", + "metrics": [ + { + "metric_id": "rem_apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "rem_ahi", + "weight": 50, + "params": { + "gt_value": 18.823529411764707, + "tolerance": 1.8823529411764708 + } + }, + { + "metric_id": "nrem_apnea_hypopnea_index_accuracy", + "type": "numeric_check", + "target_key": "nrem_ahi", + "weight": 50, + "params": { + "gt_value": 34.43609022556391, + "tolerance": 3.443609022556391 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_01.json b/sleep_assessment/cases/case37/case37_01.json new file mode 100644 index 0000000000000000000000000000000000000000..3784c95ebef6e5fbbfe6e563532a77e5114960af --- /dev/null +++ b/sleep_assessment/cases/case37/case37_01.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 173 to minute 174. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_01.edf. The analyzed 60-second window is 173-174 minutes. Ground truth: the first target event is a apnea annotation (OA) from 10380 to 10410 seconds, followed by an arousal from 10410 to 10440 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as apnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_02.json b/sleep_assessment/cases/case37/case37_02.json new file mode 100644 index 0000000000000000000000000000000000000000..47884aafa75a484f8932fa5605d6a091e46c1e8b --- /dev/null +++ b/sleep_assessment/cases/case37/case37_02.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 174.5 to minute 175.5. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_01.edf. The analyzed 60-second window is 174.5-175.5 minutes. Ground truth: the first target event is a hypopnea annotation (OH) from 10470 to 10500 seconds, followed by an arousal from 10500 to 10530 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_03.json b/sleep_assessment/cases/case37/case37_03.json new file mode 100644 index 0000000000000000000000000000000000000000..d5913ef3d57e2a8eb1fdc3de440121f6a4bff549 --- /dev/null +++ b/sleep_assessment/cases/case37/case37_03.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 183 to minute 184. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_01.edf. The analyzed 60-second window is 183-184 minutes. Ground truth: the first target event is a apnea annotation (OA) from 10980 to 11010 seconds, followed by an arousal from 11010 to 11040 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as apnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_04.json b/sleep_assessment/cases/case37/case37_04.json new file mode 100644 index 0000000000000000000000000000000000000000..6a5fdcc3900509cd4d950ca686aa44a53f5ca937 --- /dev/null +++ b/sleep_assessment/cases/case37/case37_04.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 222.5 to minute 223.5. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_01.edf. The analyzed 60-second window is 222.5-223.5 minutes. Ground truth: the first target event is a apnea annotation (CA) from 13350 to 13380 seconds, followed by an arousal from 13380 to 13410 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as apnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_05.json b/sleep_assessment/cases/case37/case37_05.json new file mode 100644 index 0000000000000000000000000000000000000000..3c138cb43ccfb5f2cac9408b2314f8b0afe1d5f7 --- /dev/null +++ b/sleep_assessment/cases/case37/case37_05.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "label_path": "data/sleep/ISRUC_02.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 91.5 to minute 92.5. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_02.edf. The analyzed 60-second window is 91.5-92.5 minutes. Ground truth: the first target event is a hypopnea annotation (MH) from 5490 to 5520 seconds, followed by an arousal from 5520 to 5550 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_06.json b/sleep_assessment/cases/case37/case37_06.json new file mode 100644 index 0000000000000000000000000000000000000000..8ca5fa4bcaa07dc956261e1b71c88800fbc48f83 --- /dev/null +++ b/sleep_assessment/cases/case37/case37_06.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "label_path": "data/sleep/ISRUC_02.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 93 to minute 94. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_02.edf. The analyzed 60-second window is 93-94 minutes. Ground truth: the first target event is a hypopnea annotation (MH) from 5580 to 5610 seconds, followed by an arousal from 5610 to 5640 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_07.json b/sleep_assessment/cases/case37/case37_07.json new file mode 100644 index 0000000000000000000000000000000000000000..b25ed038445a124a258211ebdf7703212e098139 --- /dev/null +++ b/sleep_assessment/cases/case37/case37_07.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "label_path": "data/sleep/ISRUC_02.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 132 to minute 133. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_02.edf. The analyzed 60-second window is 132-133 minutes. Ground truth: the first target event is a hypopnea annotation (MH) from 7920 to 7950 seconds, followed by an arousal from 7950 to 7980 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_08.json b/sleep_assessment/cases/case37/case37_08.json new file mode 100644 index 0000000000000000000000000000000000000000..3012d08531475dff471d7d13c6b4fc5b7d6b70d1 --- /dev/null +++ b/sleep_assessment/cases/case37/case37_08.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "label_path": "data/sleep/ISRUC_02.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 147 to minute 148. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_02.edf. The analyzed 60-second window is 147-148 minutes. Ground truth: the first target event is a hypopnea annotation (MH) from 8820 to 8850 seconds, followed by an arousal from 8850 to 8880 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_09.json b/sleep_assessment/cases/case37/case37_09.json new file mode 100644 index 0000000000000000000000000000000000000000..fcad02c8935cfd2fe60908f9b40702632e722ced --- /dev/null +++ b/sleep_assessment/cases/case37/case37_09.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "label_path": "data/sleep/ISRUC_03.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 168.5 to minute 169.5. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_03.edf. The analyzed 60-second window is 168.5-169.5 minutes. Ground truth: the first target event is a apnea annotation (CA) from 10110 to 10140 seconds, followed by an arousal from 10140 to 10170 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as apnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_10.json b/sleep_assessment/cases/case37/case37_10.json new file mode 100644 index 0000000000000000000000000000000000000000..e067f021ddab5f9186526a51222587ce7d0393f7 --- /dev/null +++ b/sleep_assessment/cases/case37/case37_10.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "label_path": "data/sleep/ISRUC_03.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 336 to minute 337. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_03.edf. The analyzed 60-second window is 336-337 minutes. Ground truth: the first target event is a hypopnea annotation (OH) from 20160 to 20190 seconds, followed by an arousal from 20190 to 20220 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_11.json b/sleep_assessment/cases/case37/case37_11.json new file mode 100644 index 0000000000000000000000000000000000000000..5e976d46f5dbcc34acb045d1041cb49c6c8556b4 --- /dev/null +++ b/sleep_assessment/cases/case37/case37_11.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "label_path": "data/sleep/ISRUC_03.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 339.5 to minute 340.5. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_03.edf. The analyzed 60-second window is 339.5-340.5 minutes. Ground truth: the first target event is a hypopnea annotation (OH) from 20370 to 20400 seconds, followed by an arousal from 20400 to 20430 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_12.json b/sleep_assessment/cases/case37/case37_12.json new file mode 100644 index 0000000000000000000000000000000000000000..ca0d913951fe562a5a3bb48bda642ff7e4a712cf --- /dev/null +++ b/sleep_assessment/cases/case37/case37_12.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "label_path": "data/sleep/ISRUC_03.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 363 to minute 364. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_03.edf. The analyzed 60-second window is 363-364 minutes. Ground truth: the first target event is a hypopnea annotation (OH) from 21780 to 21810 seconds, followed by an arousal from 21810 to 21840 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_13.json b/sleep_assessment/cases/case37/case37_13.json new file mode 100644 index 0000000000000000000000000000000000000000..2bbc5e771e0b36b5208ab8577e9d1256c29c8b3d --- /dev/null +++ b/sleep_assessment/cases/case37/case37_13.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "ISRUC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_04.edf", + "label_path": "data/sleep/ISRUC_04.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 342.5 to minute 343.5. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_04.edf. The analyzed 60-second window is 342.5-343.5 minutes. Ground truth: the first target event is a hypopnea annotation (MH) from 20550 to 20580 seconds, followed by an arousal from 20580 to 20610 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_14.json b/sleep_assessment/cases/case37/case37_14.json new file mode 100644 index 0000000000000000000000000000000000000000..7ef1ac14d8ffa34cbb15d5812faf9160eb9b7563 --- /dev/null +++ b/sleep_assessment/cases/case37/case37_14.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "ISRUC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_04.edf", + "label_path": "data/sleep/ISRUC_04.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 345.5 to minute 346.5. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_04.edf. The analyzed 60-second window is 345.5-346.5 minutes. Ground truth: the first target event is a hypopnea annotation (MH) from 20730 to 20760 seconds, followed by an arousal from 20760 to 20790 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_15.json b/sleep_assessment/cases/case37/case37_15.json new file mode 100644 index 0000000000000000000000000000000000000000..c74f0b761d40ed60e0eeefca9ab613e908c25107 --- /dev/null +++ b/sleep_assessment/cases/case37/case37_15.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 179 to minute 180. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_01.edf. The analyzed 60-second window is 179-180 minutes. Ground truth: the first target event is a hypopnea annotation (OH) from 10740 to 10770 seconds, followed by an arousal from 10770 to 10800 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_16.json b/sleep_assessment/cases/case37/case37_16.json new file mode 100644 index 0000000000000000000000000000000000000000..ed1a9066a9b0278145a1da25a59288b09f7c9096 --- /dev/null +++ b/sleep_assessment/cases/case37/case37_16.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 237 to minute 238. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_01.edf. The analyzed 60-second window is 237-238 minutes. Ground truth: the first target event is a apnea annotation (CA) from 14220 to 14250 seconds, followed by an arousal from 14250 to 14280 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as apnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_17.json b/sleep_assessment/cases/case37/case37_17.json new file mode 100644 index 0000000000000000000000000000000000000000..147c6f987e2a48049deadf5a19a09e8dba968c32 --- /dev/null +++ b/sleep_assessment/cases/case37/case37_17.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "label_path": "data/sleep/ISRUC_05.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 80 to minute 81. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_05.edf. The analyzed 60-second window is 80-81 minutes. Ground truth: the first target event is a hypopnea annotation (MH) from 4800 to 4830 seconds, followed by an arousal from 4830 to 4860 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_18.json b/sleep_assessment/cases/case37/case37_18.json new file mode 100644 index 0000000000000000000000000000000000000000..7e721e963a96a03499837004b1ef9f6db75c07b9 --- /dev/null +++ b/sleep_assessment/cases/case37/case37_18.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "label_path": "data/sleep/ISRUC_05.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 130 to minute 131. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_05.edf. The analyzed 60-second window is 130-131 minutes. Ground truth: the first target event is a hypopnea annotation (MH) from 7800 to 7830 seconds, followed by an arousal from 7830 to 7860 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_19.json b/sleep_assessment/cases/case37/case37_19.json new file mode 100644 index 0000000000000000000000000000000000000000..b7e8a9e20a21a15422ae1530903c81650ccc040e --- /dev/null +++ b/sleep_assessment/cases/case37/case37_19.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "label_path": "data/sleep/ISRUC_05.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 267.5 to minute 268.5. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_05.edf. The analyzed 60-second window is 267.5-268.5 minutes. Ground truth: the first target event is a hypopnea annotation (OH) from 16050 to 16080 seconds, followed by an arousal from 16080 to 16110 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_20.json b/sleep_assessment/cases/case37/case37_20.json new file mode 100644 index 0000000000000000000000000000000000000000..565b4a274ed72b1daa7c43ab1291fd0a37c47881 --- /dev/null +++ b/sleep_assessment/cases/case37/case37_20.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "label_path": "data/sleep/ISRUC_05.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 315.5 to minute 316.5. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_05.edf. The analyzed 60-second window is 315.5-316.5 minutes. Ground truth: the first target event is a hypopnea annotation (OH) from 18930 to 18960 seconds, followed by an arousal from 18960 to 18990 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_21.json b/sleep_assessment/cases/case37/case37_21.json new file mode 100644 index 0000000000000000000000000000000000000000..90836508791abf02a71dba3c17a3167a43047fab --- /dev/null +++ b/sleep_assessment/cases/case37/case37_21.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "label_path": "data/sleep/SHHS1_01.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 304.753333 to minute 305.753333. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_01.edf. The analyzed 60-second window is 304.753333-305.753333 minutes. Ground truth: the first target event is a hypopnea annotation (Hypopnea) from 18296.8 to 18324.9 seconds, followed by an arousal from 18325.1 to 18345.2 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_22.json b/sleep_assessment/cases/case37/case37_22.json new file mode 100644 index 0000000000000000000000000000000000000000..a17c64cc6063d6f99988dd018154d08bd0e3125c --- /dev/null +++ b/sleep_assessment/cases/case37/case37_22.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "label_path": "data/sleep/SHHS1_01.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 419.753333 to minute 420.753333. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_01.edf. The analyzed 60-second window is 419.753333-420.753333 minutes. Ground truth: the first target event is a hypopnea annotation (Hypopnea) from 25208.2 to 25221.4 seconds, followed by an arousal from 25222.8 to 25245.2 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_23.json b/sleep_assessment/cases/case37/case37_23.json new file mode 100644 index 0000000000000000000000000000000000000000..c9072b319edbcdf5d19690e8ffd95484fbf91bea --- /dev/null +++ b/sleep_assessment/cases/case37/case37_23.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "label_path": "data/sleep/SHHS1_01.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 150.731667 to minute 151.731667. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_01.edf. The analyzed 60-second window is 150.731667-151.731667 minutes. Ground truth: the first target event is a hypopnea annotation (Hypopnea) from 9066 to 9095 seconds, followed by an arousal from 9096.5 to 9103.9 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_24.json b/sleep_assessment/cases/case37/case37_24.json new file mode 100644 index 0000000000000000000000000000000000000000..69301c6e21fd9dbad297409308ec3627a136e0bb --- /dev/null +++ b/sleep_assessment/cases/case37/case37_24.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "label_path": "data/sleep/SHHS1_01.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 439.416667 to minute 440.416667. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_01.edf. The analyzed 60-second window is 439.416667-440.416667 minutes. Ground truth: the first target event is a hypopnea annotation (Hypopnea) from 26382 to 26397 seconds, followed by an arousal from 26399.7 to 26410.4 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_25.json b/sleep_assessment/cases/case37/case37_25.json new file mode 100644 index 0000000000000000000000000000000000000000..ee3d3acd4c115e96cc8fb172eed937878c916d0e --- /dev/null +++ b/sleep_assessment/cases/case37/case37_25.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "label_path": "data/sleep/SHHS1_02.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 17.951667 to minute 18.951667. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_02.edf. The analyzed 60-second window is 17.951667-18.951667 minutes. Ground truth: the first target event is a hypopnea annotation (Hypopnea) from 1080 to 1111 seconds, followed by an arousal from 1111 to 1128.8 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_26.json b/sleep_assessment/cases/case37/case37_26.json new file mode 100644 index 0000000000000000000000000000000000000000..e47e14682da7b6ccaba3da14252fa8a18eaff5ea --- /dev/null +++ b/sleep_assessment/cases/case37/case37_26.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "label_path": "data/sleep/SHHS1_02.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 49.585 to minute 50.585. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_02.edf. The analyzed 60-second window is 49.585-50.585 minutes. Ground truth: the first target event is a apnea annotation (Obstructive Apnea) from 2977 to 3024 seconds, followed by an arousal from 3025.5 to 3035.1 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as apnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_27.json b/sleep_assessment/cases/case37/case37_27.json new file mode 100644 index 0000000000000000000000000000000000000000..f14deb265582b3b9f774c043d0f81b94b1cc3f52 --- /dev/null +++ b/sleep_assessment/cases/case37/case37_27.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "label_path": "data/sleep/SHHS1_02.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 76.83 to minute 77.83. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_02.edf. The analyzed 60-second window is 76.83-77.83 minutes. Ground truth: the first target event is a apnea annotation (Obstructive Apnea) from 4626 to 4650 seconds, followed by an arousal from 4652.5 to 4669.8 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as apnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_28.json b/sleep_assessment/cases/case37/case37_28.json new file mode 100644 index 0000000000000000000000000000000000000000..9ad258906c6e3b43316c9949d35f6b9c2202b17b --- /dev/null +++ b/sleep_assessment/cases/case37/case37_28.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "label_path": "data/sleep/SHHS1_02.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 197.353333 to minute 198.353333. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_02.edf. The analyzed 60-second window is 197.353333-198.353333 minutes. Ground truth: the first target event is a hypopnea annotation (Hypopnea) from 11845 to 11866 seconds, followed by an arousal from 11866.5 to 11877.4 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_29.json b/sleep_assessment/cases/case37/case37_29.json new file mode 100644 index 0000000000000000000000000000000000000000..ba22dc78c959eec61f2c60df5e851276b806d014 --- /dev/null +++ b/sleep_assessment/cases/case37/case37_29.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "label_path": "data/sleep/SHHS1_03.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 231.753333 to minute 232.753333. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_03.edf. The analyzed 60-second window is 231.753333-232.753333 minutes. Ground truth: the first target event is a apnea annotation (Obstructive Apnea) from 13912.7 to 13944 seconds, followed by an arousal from 13944.1 to 13965.2 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as apnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_30.json b/sleep_assessment/cases/case37/case37_30.json new file mode 100644 index 0000000000000000000000000000000000000000..f254f506db1dd1d448d87a77f576273699ebfe46 --- /dev/null +++ b/sleep_assessment/cases/case37/case37_30.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "label_path": "data/sleep/SHHS1_03.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 146.128333 to minute 147.128333. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_03.edf. The analyzed 60-second window is 146.128333-147.128333 minutes. Ground truth: the first target event is a hypopnea annotation (Hypopnea) from 8804.5 to 8817.7 seconds, followed by an arousal from 8821.4 to 8827.7 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_31.json b/sleep_assessment/cases/case37/case37_31.json new file mode 100644 index 0000000000000000000000000000000000000000..7838448a73c430eb1d576cade7d1dbadefea23e1 --- /dev/null +++ b/sleep_assessment/cases/case37/case37_31.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "label_path": "data/sleep/SHHS1_03.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 152.875 to minute 153.875. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_03.edf. The analyzed 60-second window is 152.875-153.875 minutes. Ground truth: the first target event is a hypopnea annotation (Hypopnea) from 9214.1 to 9224.5 seconds, followed by an arousal from 9228.3 to 9232.5 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_32.json b/sleep_assessment/cases/case37/case37_32.json new file mode 100644 index 0000000000000000000000000000000000000000..17fcd356360fffcdc8eb8f8c5e1b7d00631fe810 --- /dev/null +++ b/sleep_assessment/cases/case37/case37_32.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "label_path": "data/sleep/SHHS1_03.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 213.486667 to minute 214.486667. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_03.edf. The analyzed 60-second window is 213.486667-214.486667 minutes. Ground truth: the first target event is a hypopnea annotation (Hypopnea) from 12842.2 to 12857 seconds, followed by an arousal from 12863.9 to 12869.2 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_33.json b/sleep_assessment/cases/case37/case37_33.json new file mode 100644 index 0000000000000000000000000000000000000000..e15e1f4cedea35251c89c7daf4fad36d33ad1f36 --- /dev/null +++ b/sleep_assessment/cases/case37/case37_33.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "label_path": "data/sleep/SHHS1_04.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 145.213333 to minute 146.213333. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_04.edf. The analyzed 60-second window is 145.213333-146.213333 minutes. Ground truth: the first target event is a hypopnea annotation (Hypopnea) from 8744.8 to 8766.6 seconds, followed by an arousal from 8766.6 to 8772.8 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_34.json b/sleep_assessment/cases/case37/case37_34.json new file mode 100644 index 0000000000000000000000000000000000000000..562e141172a41334896e3846bdc434a721941ebe --- /dev/null +++ b/sleep_assessment/cases/case37/case37_34.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "label_path": "data/sleep/SHHS1_04.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 61.52 to minute 62.52. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_04.edf. The analyzed 60-second window is 61.52-62.52 minutes. Ground truth: the first target event is a hypopnea annotation (Hypopnea) from 3714.3 to 3730.1 seconds, followed by an arousal from 3730.2 to 3741.6 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_35.json b/sleep_assessment/cases/case37/case37_35.json new file mode 100644 index 0000000000000000000000000000000000000000..a0f733ee77c9df6a14660c677ab4599bc087e393 --- /dev/null +++ b/sleep_assessment/cases/case37/case37_35.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "label_path": "data/sleep/SHHS1_04.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 297.263333 to minute 298.263333. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_04.edf. The analyzed 60-second window is 297.263333-298.263333 minutes. Ground truth: the first target event is a hypopnea annotation (Hypopnea) from 17857 to 17881 seconds, followed by an arousal from 17881.9 to 17895.8 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_36.json b/sleep_assessment/cases/case37/case37_36.json new file mode 100644 index 0000000000000000000000000000000000000000..44165403a01e5096cbb30231723d1786e6ece8e5 --- /dev/null +++ b/sleep_assessment/cases/case37/case37_36.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "label_path": "data/sleep/SHHS1_04.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 224.8 to minute 225.8. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_04.edf. The analyzed 60-second window is 224.8-225.8 minutes. Ground truth: the first target event is a hypopnea annotation (Hypopnea) from 13508 to 13530 seconds, followed by an arousal from 13531 to 13548 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_37.json b/sleep_assessment/cases/case37/case37_37.json new file mode 100644 index 0000000000000000000000000000000000000000..e5cc8fb33be406a881dd86800cf63a4c8d38c37d --- /dev/null +++ b/sleep_assessment/cases/case37/case37_37.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "label_path": "data/sleep/SHHS1_05.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 288.38 to minute 289.38. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_05.edf. The analyzed 60-second window is 288.38-289.38 minutes. Ground truth: the first target event is a hypopnea annotation (Hypopnea) from 17332.1 to 17352.6 seconds, followed by an arousal from 17352.8 to 17362.8 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_38.json b/sleep_assessment/cases/case37/case37_38.json new file mode 100644 index 0000000000000000000000000000000000000000..4530f0c41ad61b22c1f00c01ab07b2c392694a33 --- /dev/null +++ b/sleep_assessment/cases/case37/case37_38.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "label_path": "data/sleep/SHHS1_05.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 389.641667 to minute 390.641667. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_05.edf. The analyzed 60-second window is 389.641667-390.641667 minutes. Ground truth: the first target event is a hypopnea annotation (Hypopnea) from 23382.6 to 23397.1 seconds, followed by an arousal from 23398.2 to 23415.2 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_39.json b/sleep_assessment/cases/case37/case37_39.json new file mode 100644 index 0000000000000000000000000000000000000000..cc4efc09e387ac6b0a4d842960f8cb26bacfd6bf --- /dev/null +++ b/sleep_assessment/cases/case37/case37_39.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "label_path": "data/sleep/SHHS1_05.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 397.076667 to minute 398.076667. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_05.edf. The analyzed 60-second window is 397.076667-398.076667 minutes. Ground truth: the first target event is a apnea annotation (Central Apnea) from 23825.3 to 23845.5 seconds, followed by an arousal from 23850.7 to 23857 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as apnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case37/case37_40.json b/sleep_assessment/cases/case37/case37_40.json new file mode 100644 index 0000000000000000000000000000000000000000..0ef54a8423005c589c1549ff4eb19349541a0771 --- /dev/null +++ b/sleep_assessment/cases/case37/case37_40.json @@ -0,0 +1,27 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "label_path": "data/sleep/SHHS1_05.npy", + "instruction": "Given the sleep file and paired sleep-stage label file, analyze the 60-second segment from minute 81.556667 to minute 82.556667. Identify the sleep events in chronological order, distinguishing apnea, hypopnea, arousal, and no event. Determine whether an apnea or hypopnea is followed by an arousal, then briefly analyze the likely impact of this event sequence on sleep continuity, sleep fragmentation, and restorative sleep. Describe an association supported by the temporal sequence rather than claiming that timing alone proves causality. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-event analysis reports.\n\n### TASK\nPreserve the complete substantive response for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or text outside the JSON object.\n3. The key must be EXACTLY \"full_text\".\n4. The value must be the complete report text as a string or null.\n5. If no substantive report is present, return null.\n\n### OUTPUT TEMPLATE\n{\"full_text\": }", + "metrics": [ + { + "metric_id": "respiratory_event_arousal_impact_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 100, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_05.edf. The analyzed 60-second window is 81.556667-82.556667 minutes. Ground truth: the first target event is a hypopnea annotation (Hypopnea) from 4922.2 to 4940.5 seconds, followed by an arousal from 4942.4 to 4953.4 seconds. Score the response additively using exactly these three criteria. Award 0.3 if it correctly identifies the first event as hypopnea; a correct specific subtype or source label is acceptable when it maps to that class. Award 0.3 if it correctly identifies a subsequent arousal and preserves the respiratory-event-then-arousal order; do not award this criterion if the order is reversed. Award 0.4 if it reasonably explains that this respiratory-event/arousal sequence is associated with sleep interruption, reduced sleep continuity, sleep fragmentation, or impaired restorative sleep. Do not require a definitive causal claim; wording such as associated with, may contribute to, or likely contributes to is correct. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case38/case38_01.json b/sleep_assessment/cases/case38/case38_01.json new file mode 100644 index 0000000000000000000000000000000000000000..1412b54954b7f74d065d5158dc7dd1b45d05af81 --- /dev/null +++ b/sleep_assessment/cases/case38/case38_01.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Given the sleep EEG/PSG file and paired sleep-stage label file, analyze deep sleep disruption across the whole night. Use the sleep labels with 0=W, 1=N1, 2=N2, 3=N3, and 4=R/REM. Count arousal events whose start time falls in an N3 epoch. Compute N3-ARI as the number of N3 arousal events divided by N3 sleep time in hours. Also count how many N3 arousal events are followed by a lighter or wake stage in the next 30-second epoch, where W, N1, N2, or R/REM all count as non-N3 transitions. Report the N3-ARI, the N3 arousal transition count, and judge whether nighttime deep sleep shows arousal-like disruption. Arousal-like disruption is present when the transition count is greater than 0." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final N3 arousal index, N3 arousal transition count, and preserve the report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"n3_ari\", \"n3_arousal_transition_count\", and \"full_text\".\n4. \"n3_ari\" must be a FLOAT or null.\n5. \"n3_arousal_transition_count\" must be a FLOAT or null.\n6. \"full_text\" must be the complete substantive report text as a string or null.\n7. If a numeric value is missing or unclear, return null for that numeric key. If no substantive report is present, return null for \"full_text\".\n\n### OUTPUT TEMPLATE\n{\"n3_ari\": , \"n3_arousal_transition_count\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "n3_arousal_index_accuracy", + "type": "numeric_check", + "target_key": "n3_ari", + "weight": 35, + "params": { + "gt_value": 1.5584415584415585, + "tolerance": 0.15584415584415587 + } + }, + { + "metric_id": "n3_arousal_transition_count_accuracy", + "type": "numeric_check", + "target_key": "n3_arousal_transition_count", + "weight": 35, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + }, + { + "metric_id": "n3_arousal_like_disruption_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 30, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_01.edf. Ground truth: N3 arousal transition count is 0, and arousal-like disruption is no. The decision threshold is strict: disruption is present if and only if the N3 arousal transition count is greater than 0. Score additively using exactly these criteria. Award 0.4 if the response correctly states whether arousal-like disruption is present. Award 0.3 if it correctly uses or is consistent with the transition-count threshold; for this sample that means recognizing 0 transition event(s) relative to the >0 rule. Award 0.3 if the clinical interpretation is directionally correct: for a positive case, N3 arousal followed by W/N1/N2/R indicates disrupted deep sleep continuity, sleep fragmentation, or impaired restorative sleep; for a negative case, there is no observed N3 arousal followed by a non-N3 stage under this rule. Do not require a definitive causal claim; associated with, suggests, or may contribute to is acceptable. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case38/case38_02.json b/sleep_assessment/cases/case38/case38_02.json new file mode 100644 index 0000000000000000000000000000000000000000..8285941cbf86d0c654ccfbb1202ae6a6ef26a9d6 --- /dev/null +++ b/sleep_assessment/cases/case38/case38_02.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "label_path": "data/sleep/ISRUC_02.npy", + "instruction": "Given the sleep EEG/PSG file and paired sleep-stage label file, analyze deep sleep disruption across the whole night. Use the sleep labels with 0=W, 1=N1, 2=N2, 3=N3, and 4=R/REM. Count arousal events whose start time falls in an N3 epoch. Compute N3-ARI as the number of N3 arousal events divided by N3 sleep time in hours. Also count how many N3 arousal events are followed by a lighter or wake stage in the next 30-second epoch, where W, N1, N2, or R/REM all count as non-N3 transitions. Report the N3-ARI, the N3 arousal transition count, and judge whether nighttime deep sleep shows arousal-like disruption. Arousal-like disruption is present when the transition count is greater than 0." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final N3 arousal index, N3 arousal transition count, and preserve the report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"n3_ari\", \"n3_arousal_transition_count\", and \"full_text\".\n4. \"n3_ari\" must be a FLOAT or null.\n5. \"n3_arousal_transition_count\" must be a FLOAT or null.\n6. \"full_text\" must be the complete substantive report text as a string or null.\n7. If a numeric value is missing or unclear, return null for that numeric key. If no substantive report is present, return null for \"full_text\".\n\n### OUTPUT TEMPLATE\n{\"n3_ari\": , \"n3_arousal_transition_count\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "n3_arousal_index_accuracy", + "type": "numeric_check", + "target_key": "n3_ari", + "weight": 35, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + }, + { + "metric_id": "n3_arousal_transition_count_accuracy", + "type": "numeric_check", + "target_key": "n3_arousal_transition_count", + "weight": 35, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + }, + { + "metric_id": "n3_arousal_like_disruption_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 30, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_02.edf. Ground truth: N3 arousal transition count is 0, and arousal-like disruption is no. The decision threshold is strict: disruption is present if and only if the N3 arousal transition count is greater than 0. Score additively using exactly these criteria. Award 0.4 if the response correctly states whether arousal-like disruption is present. Award 0.3 if it correctly uses or is consistent with the transition-count threshold; for this sample that means recognizing 0 transition event(s) relative to the >0 rule. Award 0.3 if the clinical interpretation is directionally correct: for a positive case, N3 arousal followed by W/N1/N2/R indicates disrupted deep sleep continuity, sleep fragmentation, or impaired restorative sleep; for a negative case, there is no observed N3 arousal followed by a non-N3 stage under this rule. Do not require a definitive causal claim; associated with, suggests, or may contribute to is acceptable. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case38/case38_03.json b/sleep_assessment/cases/case38/case38_03.json new file mode 100644 index 0000000000000000000000000000000000000000..76e478cacdd07d8431ab2dc1c606731188467ae2 --- /dev/null +++ b/sleep_assessment/cases/case38/case38_03.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "label_path": "data/sleep/ISRUC_03.npy", + "instruction": "Given the sleep EEG/PSG file and paired sleep-stage label file, analyze deep sleep disruption across the whole night. Use the sleep labels with 0=W, 1=N1, 2=N2, 3=N3, and 4=R/REM. Count arousal events whose start time falls in an N3 epoch. Compute N3-ARI as the number of N3 arousal events divided by N3 sleep time in hours. Also count how many N3 arousal events are followed by a lighter or wake stage in the next 30-second epoch, where W, N1, N2, or R/REM all count as non-N3 transitions. Report the N3-ARI, the N3 arousal transition count, and judge whether nighttime deep sleep shows arousal-like disruption. Arousal-like disruption is present when the transition count is greater than 0." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final N3 arousal index, N3 arousal transition count, and preserve the report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"n3_ari\", \"n3_arousal_transition_count\", and \"full_text\".\n4. \"n3_ari\" must be a FLOAT or null.\n5. \"n3_arousal_transition_count\" must be a FLOAT or null.\n6. \"full_text\" must be the complete substantive report text as a string or null.\n7. If a numeric value is missing or unclear, return null for that numeric key. If no substantive report is present, return null for \"full_text\".\n\n### OUTPUT TEMPLATE\n{\"n3_ari\": , \"n3_arousal_transition_count\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "n3_arousal_index_accuracy", + "type": "numeric_check", + "target_key": "n3_ari", + "weight": 35, + "params": { + "gt_value": 0.6936416184971098, + "tolerance": 0.06936416184971099 + } + }, + { + "metric_id": "n3_arousal_transition_count_accuracy", + "type": "numeric_check", + "target_key": "n3_arousal_transition_count", + "weight": 35, + "params": { + "gt_value": 1.0, + "tolerance": 0.0 + } + }, + { + "metric_id": "n3_arousal_like_disruption_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 30, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_03.edf. Ground truth: N3 arousal transition count is 1, and arousal-like disruption is yes. The decision threshold is strict: disruption is present if and only if the N3 arousal transition count is greater than 0. Score additively using exactly these criteria. Award 0.4 if the response correctly states whether arousal-like disruption is present. Award 0.3 if it correctly uses or is consistent with the transition-count threshold; for this sample that means recognizing 1 transition event(s) relative to the >0 rule. Award 0.3 if the clinical interpretation is directionally correct: for a positive case, N3 arousal followed by W/N1/N2/R indicates disrupted deep sleep continuity, sleep fragmentation, or impaired restorative sleep; for a negative case, there is no observed N3 arousal followed by a non-N3 stage under this rule. Do not require a definitive causal claim; associated with, suggests, or may contribute to is acceptable. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case38/case38_04.json b/sleep_assessment/cases/case38/case38_04.json new file mode 100644 index 0000000000000000000000000000000000000000..db94cb93743bc806ce9c4ed0d284213e3ebef126 --- /dev/null +++ b/sleep_assessment/cases/case38/case38_04.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "ISRUC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_04.edf", + "label_path": "data/sleep/ISRUC_04.npy", + "instruction": "Given the sleep EEG/PSG file and paired sleep-stage label file, analyze deep sleep disruption across the whole night. Use the sleep labels with 0=W, 1=N1, 2=N2, 3=N3, and 4=R/REM. Count arousal events whose start time falls in an N3 epoch. Compute N3-ARI as the number of N3 arousal events divided by N3 sleep time in hours. Also count how many N3 arousal events are followed by a lighter or wake stage in the next 30-second epoch, where W, N1, N2, or R/REM all count as non-N3 transitions. Report the N3-ARI, the N3 arousal transition count, and judge whether nighttime deep sleep shows arousal-like disruption. Arousal-like disruption is present when the transition count is greater than 0." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final N3 arousal index, N3 arousal transition count, and preserve the report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"n3_ari\", \"n3_arousal_transition_count\", and \"full_text\".\n4. \"n3_ari\" must be a FLOAT or null.\n5. \"n3_arousal_transition_count\" must be a FLOAT or null.\n6. \"full_text\" must be the complete substantive report text as a string or null.\n7. If a numeric value is missing or unclear, return null for that numeric key. If no substantive report is present, return null for \"full_text\".\n\n### OUTPUT TEMPLATE\n{\"n3_ari\": , \"n3_arousal_transition_count\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "n3_arousal_index_accuracy", + "type": "numeric_check", + "target_key": "n3_ari", + "weight": 35, + "params": { + "gt_value": 1.1214953271028036, + "tolerance": 0.11214953271028037 + } + }, + { + "metric_id": "n3_arousal_transition_count_accuracy", + "type": "numeric_check", + "target_key": "n3_arousal_transition_count", + "weight": 35, + "params": { + "gt_value": 1.0, + "tolerance": 0.0 + } + }, + { + "metric_id": "n3_arousal_like_disruption_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 30, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_04.edf. Ground truth: N3 arousal transition count is 1, and arousal-like disruption is yes. The decision threshold is strict: disruption is present if and only if the N3 arousal transition count is greater than 0. Score additively using exactly these criteria. Award 0.4 if the response correctly states whether arousal-like disruption is present. Award 0.3 if it correctly uses or is consistent with the transition-count threshold; for this sample that means recognizing 1 transition event(s) relative to the >0 rule. Award 0.3 if the clinical interpretation is directionally correct: for a positive case, N3 arousal followed by W/N1/N2/R indicates disrupted deep sleep continuity, sleep fragmentation, or impaired restorative sleep; for a negative case, there is no observed N3 arousal followed by a non-N3 stage under this rule. Do not require a definitive causal claim; associated with, suggests, or may contribute to is acceptable. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case38/case38_05.json b/sleep_assessment/cases/case38/case38_05.json new file mode 100644 index 0000000000000000000000000000000000000000..0ab7adafbfc33a4d57c695330afa3854f54c3ab9 --- /dev/null +++ b/sleep_assessment/cases/case38/case38_05.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "label_path": "data/sleep/ISRUC_05.npy", + "instruction": "Given the sleep EEG/PSG file and paired sleep-stage label file, analyze deep sleep disruption across the whole night. Use the sleep labels with 0=W, 1=N1, 2=N2, 3=N3, and 4=R/REM. Count arousal events whose start time falls in an N3 epoch. Compute N3-ARI as the number of N3 arousal events divided by N3 sleep time in hours. Also count how many N3 arousal events are followed by a lighter or wake stage in the next 30-second epoch, where W, N1, N2, or R/REM all count as non-N3 transitions. Report the N3-ARI, the N3 arousal transition count, and judge whether nighttime deep sleep shows arousal-like disruption. Arousal-like disruption is present when the transition count is greater than 0." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final N3 arousal index, N3 arousal transition count, and preserve the report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"n3_ari\", \"n3_arousal_transition_count\", and \"full_text\".\n4. \"n3_ari\" must be a FLOAT or null.\n5. \"n3_arousal_transition_count\" must be a FLOAT or null.\n6. \"full_text\" must be the complete substantive report text as a string or null.\n7. If a numeric value is missing or unclear, return null for that numeric key. If no substantive report is present, return null for \"full_text\".\n\n### OUTPUT TEMPLATE\n{\"n3_ari\": , \"n3_arousal_transition_count\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "n3_arousal_index_accuracy", + "type": "numeric_check", + "target_key": "n3_ari", + "weight": 35, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + }, + { + "metric_id": "n3_arousal_transition_count_accuracy", + "type": "numeric_check", + "target_key": "n3_arousal_transition_count", + "weight": 35, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + }, + { + "metric_id": "n3_arousal_like_disruption_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 30, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_05.edf. Ground truth: N3 arousal transition count is 0, and arousal-like disruption is no. The decision threshold is strict: disruption is present if and only if the N3 arousal transition count is greater than 0. Score additively using exactly these criteria. Award 0.4 if the response correctly states whether arousal-like disruption is present. Award 0.3 if it correctly uses or is consistent with the transition-count threshold; for this sample that means recognizing 0 transition event(s) relative to the >0 rule. Award 0.3 if the clinical interpretation is directionally correct: for a positive case, N3 arousal followed by W/N1/N2/R indicates disrupted deep sleep continuity, sleep fragmentation, or impaired restorative sleep; for a negative case, there is no observed N3 arousal followed by a non-N3 stage under this rule. Do not require a definitive causal claim; associated with, suggests, or may contribute to is acceptable. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case38/case38_06.json b/sleep_assessment/cases/case38/case38_06.json new file mode 100644 index 0000000000000000000000000000000000000000..c0a7785b4fca21d603b761c4721c5b0731d63bdb --- /dev/null +++ b/sleep_assessment/cases/case38/case38_06.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "label_path": "data/sleep/SHHS1_01.npy", + "instruction": "Given the sleep EEG/PSG file and paired sleep-stage label file, analyze deep sleep disruption across the whole night. Use the sleep labels with 0=W, 1=N1, 2=N2, 3=N3, and 4=R/REM. Count arousal events whose start time falls in an N3 epoch. Compute N3-ARI as the number of N3 arousal events divided by N3 sleep time in hours. Also count how many N3 arousal events are followed by a lighter or wake stage in the next 30-second epoch, where W, N1, N2, or R/REM all count as non-N3 transitions. Report the N3-ARI, the N3 arousal transition count, and judge whether nighttime deep sleep shows arousal-like disruption. Arousal-like disruption is present when the transition count is greater than 0." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final N3 arousal index, N3 arousal transition count, and preserve the report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"n3_ari\", \"n3_arousal_transition_count\", and \"full_text\".\n4. \"n3_ari\" must be a FLOAT or null.\n5. \"n3_arousal_transition_count\" must be a FLOAT or null.\n6. \"full_text\" must be the complete substantive report text as a string or null.\n7. If a numeric value is missing or unclear, return null for that numeric key. If no substantive report is present, return null for \"full_text\".\n\n### OUTPUT TEMPLATE\n{\"n3_ari\": , \"n3_arousal_transition_count\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "n3_arousal_index_accuracy", + "type": "numeric_check", + "target_key": "n3_ari", + "weight": 35, + "params": { + "gt_value": 5.793103448275862, + "tolerance": 0.5793103448275863 + } + }, + { + "metric_id": "n3_arousal_transition_count_accuracy", + "type": "numeric_check", + "target_key": "n3_arousal_transition_count", + "weight": 35, + "params": { + "gt_value": 5.0, + "tolerance": 1.0 + } + }, + { + "metric_id": "n3_arousal_like_disruption_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 30, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_01.edf. Ground truth: N3 arousal transition count is 5, and arousal-like disruption is yes. The decision threshold is strict: disruption is present if and only if the N3 arousal transition count is greater than 0. Score additively using exactly these criteria. Award 0.4 if the response correctly states whether arousal-like disruption is present. Award 0.3 if it correctly uses or is consistent with the transition-count threshold; for this sample that means recognizing 5 transition event(s) relative to the >0 rule. Award 0.3 if the clinical interpretation is directionally correct: for a positive case, N3 arousal followed by W/N1/N2/R indicates disrupted deep sleep continuity, sleep fragmentation, or impaired restorative sleep; for a negative case, there is no observed N3 arousal followed by a non-N3 stage under this rule. Do not require a definitive causal claim; associated with, suggests, or may contribute to is acceptable. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case38/case38_07.json b/sleep_assessment/cases/case38/case38_07.json new file mode 100644 index 0000000000000000000000000000000000000000..d5378ff44b1ef047d7cb32de734426bc242e97b9 --- /dev/null +++ b/sleep_assessment/cases/case38/case38_07.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "label_path": "data/sleep/SHHS1_02.npy", + "instruction": "Given the sleep EEG/PSG file and paired sleep-stage label file, analyze deep sleep disruption across the whole night. Use the sleep labels with 0=W, 1=N1, 2=N2, 3=N3, and 4=R/REM. Count arousal events whose start time falls in an N3 epoch. Compute N3-ARI as the number of N3 arousal events divided by N3 sleep time in hours. Also count how many N3 arousal events are followed by a lighter or wake stage in the next 30-second epoch, where W, N1, N2, or R/REM all count as non-N3 transitions. Report the N3-ARI, the N3 arousal transition count, and judge whether nighttime deep sleep shows arousal-like disruption. Arousal-like disruption is present when the transition count is greater than 0." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final N3 arousal index, N3 arousal transition count, and preserve the report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"n3_ari\", \"n3_arousal_transition_count\", and \"full_text\".\n4. \"n3_ari\" must be a FLOAT or null.\n5. \"n3_arousal_transition_count\" must be a FLOAT or null.\n6. \"full_text\" must be the complete substantive report text as a string or null.\n7. If a numeric value is missing or unclear, return null for that numeric key. If no substantive report is present, return null for \"full_text\".\n\n### OUTPUT TEMPLATE\n{\"n3_ari\": , \"n3_arousal_transition_count\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "n3_arousal_index_accuracy", + "type": "numeric_check", + "target_key": "n3_ari", + "weight": 35, + "params": { + "gt_value": 17.704918032786885, + "tolerance": 1.7704918032786887 + } + }, + { + "metric_id": "n3_arousal_transition_count_accuracy", + "type": "numeric_check", + "target_key": "n3_arousal_transition_count", + "weight": 35, + "params": { + "gt_value": 8.0, + "tolerance": 1.0 + } + }, + { + "metric_id": "n3_arousal_like_disruption_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 30, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_02.edf. Ground truth: N3 arousal transition count is 8, and arousal-like disruption is yes. The decision threshold is strict: disruption is present if and only if the N3 arousal transition count is greater than 0. Score additively using exactly these criteria. Award 0.4 if the response correctly states whether arousal-like disruption is present. Award 0.3 if it correctly uses or is consistent with the transition-count threshold; for this sample that means recognizing 8 transition event(s) relative to the >0 rule. Award 0.3 if the clinical interpretation is directionally correct: for a positive case, N3 arousal followed by W/N1/N2/R indicates disrupted deep sleep continuity, sleep fragmentation, or impaired restorative sleep; for a negative case, there is no observed N3 arousal followed by a non-N3 stage under this rule. Do not require a definitive causal claim; associated with, suggests, or may contribute to is acceptable. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case38/case38_08.json b/sleep_assessment/cases/case38/case38_08.json new file mode 100644 index 0000000000000000000000000000000000000000..239fc2392bbec9d2fdfeb7e2aac83980510fed14 --- /dev/null +++ b/sleep_assessment/cases/case38/case38_08.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "label_path": "data/sleep/SHHS1_03.npy", + "instruction": "Given the sleep EEG/PSG file and paired sleep-stage label file, analyze deep sleep disruption across the whole night. Use the sleep labels with 0=W, 1=N1, 2=N2, 3=N3, and 4=R/REM. Count arousal events whose start time falls in an N3 epoch. Compute N3-ARI as the number of N3 arousal events divided by N3 sleep time in hours. Also count how many N3 arousal events are followed by a lighter or wake stage in the next 30-second epoch, where W, N1, N2, or R/REM all count as non-N3 transitions. Report the N3-ARI, the N3 arousal transition count, and judge whether nighttime deep sleep shows arousal-like disruption. Arousal-like disruption is present when the transition count is greater than 0." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final N3 arousal index, N3 arousal transition count, and preserve the report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"n3_ari\", \"n3_arousal_transition_count\", and \"full_text\".\n4. \"n3_ari\" must be a FLOAT or null.\n5. \"n3_arousal_transition_count\" must be a FLOAT or null.\n6. \"full_text\" must be the complete substantive report text as a string or null.\n7. If a numeric value is missing or unclear, return null for that numeric key. If no substantive report is present, return null for \"full_text\".\n\n### OUTPUT TEMPLATE\n{\"n3_ari\": , \"n3_arousal_transition_count\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "n3_arousal_index_accuracy", + "type": "numeric_check", + "target_key": "n3_ari", + "weight": 35, + "params": { + "gt_value": 11.335504885993487, + "tolerance": 1.1335504885993488 + } + }, + { + "metric_id": "n3_arousal_transition_count_accuracy", + "type": "numeric_check", + "target_key": "n3_arousal_transition_count", + "weight": 35, + "params": { + "gt_value": 8.0, + "tolerance": 1.0 + } + }, + { + "metric_id": "n3_arousal_like_disruption_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 30, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_03.edf. Ground truth: N3 arousal transition count is 8, and arousal-like disruption is yes. The decision threshold is strict: disruption is present if and only if the N3 arousal transition count is greater than 0. Score additively using exactly these criteria. Award 0.4 if the response correctly states whether arousal-like disruption is present. Award 0.3 if it correctly uses or is consistent with the transition-count threshold; for this sample that means recognizing 8 transition event(s) relative to the >0 rule. Award 0.3 if the clinical interpretation is directionally correct: for a positive case, N3 arousal followed by W/N1/N2/R indicates disrupted deep sleep continuity, sleep fragmentation, or impaired restorative sleep; for a negative case, there is no observed N3 arousal followed by a non-N3 stage under this rule. Do not require a definitive causal claim; associated with, suggests, or may contribute to is acceptable. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case38/case38_09.json b/sleep_assessment/cases/case38/case38_09.json new file mode 100644 index 0000000000000000000000000000000000000000..7aa2e34138ecd1b3c5cc1c98edf4ab93462e17d5 --- /dev/null +++ b/sleep_assessment/cases/case38/case38_09.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "label_path": "data/sleep/SHHS1_04.npy", + "instruction": "Given the sleep EEG/PSG file and paired sleep-stage label file, analyze deep sleep disruption across the whole night. Use the sleep labels with 0=W, 1=N1, 2=N2, 3=N3, and 4=R/REM. Count arousal events whose start time falls in an N3 epoch. Compute N3-ARI as the number of N3 arousal events divided by N3 sleep time in hours. Also count how many N3 arousal events are followed by a lighter or wake stage in the next 30-second epoch, where W, N1, N2, or R/REM all count as non-N3 transitions. Report the N3-ARI, the N3 arousal transition count, and judge whether nighttime deep sleep shows arousal-like disruption. Arousal-like disruption is present when the transition count is greater than 0." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final N3 arousal index, N3 arousal transition count, and preserve the report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"n3_ari\", \"n3_arousal_transition_count\", and \"full_text\".\n4. \"n3_ari\" must be a FLOAT or null.\n5. \"n3_arousal_transition_count\" must be a FLOAT or null.\n6. \"full_text\" must be the complete substantive report text as a string or null.\n7. If a numeric value is missing or unclear, return null for that numeric key. If no substantive report is present, return null for \"full_text\".\n\n### OUTPUT TEMPLATE\n{\"n3_ari\": , \"n3_arousal_transition_count\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "n3_arousal_index_accuracy", + "type": "numeric_check", + "target_key": "n3_ari", + "weight": 35, + "params": { + "gt_value": 2.2857142857142856, + "tolerance": 0.22857142857142856 + } + }, + { + "metric_id": "n3_arousal_transition_count_accuracy", + "type": "numeric_check", + "target_key": "n3_arousal_transition_count", + "weight": 35, + "params": { + "gt_value": 4.0, + "tolerance": 1.0 + } + }, + { + "metric_id": "n3_arousal_like_disruption_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 30, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_04.edf. Ground truth: N3 arousal transition count is 4, and arousal-like disruption is yes. The decision threshold is strict: disruption is present if and only if the N3 arousal transition count is greater than 0. Score additively using exactly these criteria. Award 0.4 if the response correctly states whether arousal-like disruption is present. Award 0.3 if it correctly uses or is consistent with the transition-count threshold; for this sample that means recognizing 4 transition event(s) relative to the >0 rule. Award 0.3 if the clinical interpretation is directionally correct: for a positive case, N3 arousal followed by W/N1/N2/R indicates disrupted deep sleep continuity, sleep fragmentation, or impaired restorative sleep; for a negative case, there is no observed N3 arousal followed by a non-N3 stage under this rule. Do not require a definitive causal claim; associated with, suggests, or may contribute to is acceptable. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case38/case38_10.json b/sleep_assessment/cases/case38/case38_10.json new file mode 100644 index 0000000000000000000000000000000000000000..4b1c339e63f7c8a32adb335a509420ab49d356c1 --- /dev/null +++ b/sleep_assessment/cases/case38/case38_10.json @@ -0,0 +1,47 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "label_path": "data/sleep/SHHS1_05.npy", + "instruction": "Given the sleep EEG/PSG file and paired sleep-stage label file, analyze deep sleep disruption across the whole night. Use the sleep labels with 0=W, 1=N1, 2=N2, 3=N3, and 4=R/REM. Count arousal events whose start time falls in an N3 epoch. Compute N3-ARI as the number of N3 arousal events divided by N3 sleep time in hours. Also count how many N3 arousal events are followed by a lighter or wake stage in the next 30-second epoch, where W, N1, N2, or R/REM all count as non-N3 transitions. Report the N3-ARI, the N3 arousal transition count, and judge whether nighttime deep sleep shows arousal-like disruption. Arousal-like disruption is present when the transition count is greater than 0." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the final N3 arousal index, N3 arousal transition count, and preserve the report text for semantic judging.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"n3_ari\", \"n3_arousal_transition_count\", and \"full_text\".\n4. \"n3_ari\" must be a FLOAT or null.\n5. \"n3_arousal_transition_count\" must be a FLOAT or null.\n6. \"full_text\" must be the complete substantive report text as a string or null.\n7. If a numeric value is missing or unclear, return null for that numeric key. If no substantive report is present, return null for \"full_text\".\n\n### OUTPUT TEMPLATE\n{\"n3_ari\": , \"n3_arousal_transition_count\": , \"full_text\": }", + "metrics": [ + { + "metric_id": "n3_arousal_index_accuracy", + "type": "numeric_check", + "target_key": "n3_ari", + "weight": 35, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + }, + { + "metric_id": "n3_arousal_transition_count_accuracy", + "type": "numeric_check", + "target_key": "n3_arousal_transition_count", + "weight": 35, + "params": { + "gt_value": 0.0, + "tolerance": 1.0 + } + }, + { + "metric_id": "n3_arousal_like_disruption_semantic_consistency", + "type": "semantic_consistency", + "target_key": "full_text", + "weight": 30, + "params": { + "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is SHHS1_05.edf. Ground truth: N3 arousal transition count is 0, and arousal-like disruption is no. The decision threshold is strict: disruption is present if and only if the N3 arousal transition count is greater than 0. Score additively using exactly these criteria. Award 0.4 if the response correctly states whether arousal-like disruption is present. Award 0.3 if it correctly uses or is consistent with the transition-count threshold; for this sample that means recognizing 0 transition event(s) relative to the >0 rule. Award 0.3 if the clinical interpretation is directionally correct: for a positive case, N3 arousal followed by W/N1/N2/R indicates disrupted deep sleep continuity, sleep fragmentation, or impaired restorative sleep; for a negative case, there is no observed N3 arousal followed by a non-N3 stage under this rule. Do not require a definitive causal claim; associated with, suggests, or may contribute to is acceptable. The final score must be the sum of the earned criteria and therefore one of 0.0, 0.3, 0.4, 0.6, 0.7, or 1.0. Output only one valid JSON object with exactly one key: {\"score\": }. Do not output Markdown, explanations, or additional fields." + } + } + ] + } +} diff --git a/sleep_assessment/cases/case39/case39_01.json b/sleep_assessment/cases/case39/case39_01.json new file mode 100644 index 0000000000000000000000000000000000000000..e4e9e0415a368e0a225d69109afb351315c75660 --- /dev/null +++ b/sleep_assessment/cases/case39/case39_01.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "ISRUC_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Given the sleep EEG/PSG file and paired sleep-stage label file, assess whole-night respiratory event clustering. Divide the recording into fixed, non-overlapping 5-minute windows starting at 0 minutes: 0-5, 5-10, 10-15, and so on. Use only complete 5-minute windows. Count apnea and hypopnea candidate events by their event onset time; an event belongs to exactly one half-open window [start, end). A cluster window is a 5-minute window with at least 3 apnea/hypopnea candidate events. For each cluster window, determine the dominant sleep stage from the ten 30-second sleep-stage labels in that window. If the dominant stage is tied, use the sleep stages at the candidate event onset epochs within the window to break the tie; if still tied, use W, N1, N2, N3, R order. Report the total number of cluster windows and the dominant stage across all cluster windows. If no cluster windows occur, report dominant stage as NULL. Allowed stage outputs are W, N1, N2, N3, R, and NULL. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the total respiratory cluster-window count and dominant cluster sleep stage.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"cluster_window_count\" and \"dominant_cluster_stage\".\n4. \"cluster_window_count\" must be a FLOAT or null.\n5. \"dominant_cluster_stage\" must be one of \"W\", \"N1\", \"N2\", \"N3\", \"R\", \"NULL\", or null.\n6. If there are no respiratory cluster windows, normalize the stage to \"NULL\".\n7. If a value is missing or unclear, return null for that key.\n\n### OUTPUT TEMPLATE\n{\"cluster_window_count\": , \"dominant_cluster_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|\"NULL\"|null>}", + "metrics": [ + { + "metric_id": "respiratory_cluster_window_count_accuracy", + "type": "numeric_check", + "target_key": "cluster_window_count", + "weight": 60, + "params": { + "gt_value": 31.0, + "tolerance": 6.0 + } + }, + { + "metric_id": "respiratory_cluster_dominant_stage_accuracy", + "type": "categorical_check", + "target_key": "dominant_cluster_stage", + "weight": 40, + "params": { + "gt_value": "N2" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case39/case39_02.json b/sleep_assessment/cases/case39/case39_02.json new file mode 100644 index 0000000000000000000000000000000000000000..db877eec0f98808e989d516a53303b8ef4371e7c --- /dev/null +++ b/sleep_assessment/cases/case39/case39_02.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "ISRUC_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "label_path": "data/sleep/ISRUC_02.npy", + "instruction": "Given the sleep EEG/PSG file and paired sleep-stage label file, assess whole-night respiratory event clustering. Divide the recording into fixed, non-overlapping 5-minute windows starting at 0 minutes: 0-5, 5-10, 10-15, and so on. Use only complete 5-minute windows. Count apnea and hypopnea candidate events by their event onset time; an event belongs to exactly one half-open window [start, end). A cluster window is a 5-minute window with at least 3 apnea/hypopnea candidate events. For each cluster window, determine the dominant sleep stage from the ten 30-second sleep-stage labels in that window. If the dominant stage is tied, use the sleep stages at the candidate event onset epochs within the window to break the tie; if still tied, use W, N1, N2, N3, R order. Report the total number of cluster windows and the dominant stage across all cluster windows. If no cluster windows occur, report dominant stage as NULL. Allowed stage outputs are W, N1, N2, N3, R, and NULL. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the total respiratory cluster-window count and dominant cluster sleep stage.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"cluster_window_count\" and \"dominant_cluster_stage\".\n4. \"cluster_window_count\" must be a FLOAT or null.\n5. \"dominant_cluster_stage\" must be one of \"W\", \"N1\", \"N2\", \"N3\", \"R\", \"NULL\", or null.\n6. If there are no respiratory cluster windows, normalize the stage to \"NULL\".\n7. If a value is missing or unclear, return null for that key.\n\n### OUTPUT TEMPLATE\n{\"cluster_window_count\": , \"dominant_cluster_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|\"NULL\"|null>}", + "metrics": [ + { + "metric_id": "respiratory_cluster_window_count_accuracy", + "type": "numeric_check", + "target_key": "cluster_window_count", + "weight": 60, + "params": { + "gt_value": 3.0, + "tolerance": 0.0 + } + }, + { + "metric_id": "respiratory_cluster_dominant_stage_accuracy", + "type": "categorical_check", + "target_key": "dominant_cluster_stage", + "weight": 40, + "params": { + "gt_value": "N1" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case39/case39_03.json b/sleep_assessment/cases/case39/case39_03.json new file mode 100644 index 0000000000000000000000000000000000000000..233caec930173bade017e5c84a049743ff220a7e --- /dev/null +++ b/sleep_assessment/cases/case39/case39_03.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "ISRUC_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "label_path": "data/sleep/ISRUC_03.npy", + "instruction": "Given the sleep EEG/PSG file and paired sleep-stage label file, assess whole-night respiratory event clustering. Divide the recording into fixed, non-overlapping 5-minute windows starting at 0 minutes: 0-5, 5-10, 10-15, and so on. Use only complete 5-minute windows. Count apnea and hypopnea candidate events by their event onset time; an event belongs to exactly one half-open window [start, end). A cluster window is a 5-minute window with at least 3 apnea/hypopnea candidate events. For each cluster window, determine the dominant sleep stage from the ten 30-second sleep-stage labels in that window. If the dominant stage is tied, use the sleep stages at the candidate event onset epochs within the window to break the tie; if still tied, use W, N1, N2, N3, R order. Report the total number of cluster windows and the dominant stage across all cluster windows. If no cluster windows occur, report dominant stage as NULL. Allowed stage outputs are W, N1, N2, N3, R, and NULL. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the total respiratory cluster-window count and dominant cluster sleep stage.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"cluster_window_count\" and \"dominant_cluster_stage\".\n4. \"cluster_window_count\" must be a FLOAT or null.\n5. \"dominant_cluster_stage\" must be one of \"W\", \"N1\", \"N2\", \"N3\", \"R\", \"NULL\", or null.\n6. If there are no respiratory cluster windows, normalize the stage to \"NULL\".\n7. If a value is missing or unclear, return null for that key.\n\n### OUTPUT TEMPLATE\n{\"cluster_window_count\": , \"dominant_cluster_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|\"NULL\"|null>}", + "metrics": [ + { + "metric_id": "respiratory_cluster_window_count_accuracy", + "type": "numeric_check", + "target_key": "cluster_window_count", + "weight": 60, + "params": { + "gt_value": 1.0, + "tolerance": 0.0 + } + }, + { + "metric_id": "respiratory_cluster_dominant_stage_accuracy", + "type": "categorical_check", + "target_key": "dominant_cluster_stage", + "weight": 40, + "params": { + "gt_value": "N1" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case39/case39_04.json b/sleep_assessment/cases/case39/case39_04.json new file mode 100644 index 0000000000000000000000000000000000000000..1c94a7216e59a924e145e6c18860ca43be51e909 --- /dev/null +++ b/sleep_assessment/cases/case39/case39_04.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "ISRUC_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_04.edf", + "label_path": "data/sleep/ISRUC_04.npy", + "instruction": "Given the sleep EEG/PSG file and paired sleep-stage label file, assess whole-night respiratory event clustering. Divide the recording into fixed, non-overlapping 5-minute windows starting at 0 minutes: 0-5, 5-10, 10-15, and so on. Use only complete 5-minute windows. Count apnea and hypopnea candidate events by their event onset time; an event belongs to exactly one half-open window [start, end). A cluster window is a 5-minute window with at least 3 apnea/hypopnea candidate events. For each cluster window, determine the dominant sleep stage from the ten 30-second sleep-stage labels in that window. If the dominant stage is tied, use the sleep stages at the candidate event onset epochs within the window to break the tie; if still tied, use W, N1, N2, N3, R order. Report the total number of cluster windows and the dominant stage across all cluster windows. If no cluster windows occur, report dominant stage as NULL. Allowed stage outputs are W, N1, N2, N3, R, and NULL. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the total respiratory cluster-window count and dominant cluster sleep stage.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"cluster_window_count\" and \"dominant_cluster_stage\".\n4. \"cluster_window_count\" must be a FLOAT or null.\n5. \"dominant_cluster_stage\" must be one of \"W\", \"N1\", \"N2\", \"N3\", \"R\", \"NULL\", or null.\n6. If there are no respiratory cluster windows, normalize the stage to \"NULL\".\n7. If a value is missing or unclear, return null for that key.\n\n### OUTPUT TEMPLATE\n{\"cluster_window_count\": , \"dominant_cluster_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|\"NULL\"|null>}", + "metrics": [ + { + "metric_id": "respiratory_cluster_window_count_accuracy", + "type": "numeric_check", + "target_key": "cluster_window_count", + "weight": 60, + "params": { + "gt_value": 1.0, + "tolerance": 0.0 + } + }, + { + "metric_id": "respiratory_cluster_dominant_stage_accuracy", + "type": "categorical_check", + "target_key": "dominant_cluster_stage", + "weight": 40, + "params": { + "gt_value": "R" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case39/case39_05.json b/sleep_assessment/cases/case39/case39_05.json new file mode 100644 index 0000000000000000000000000000000000000000..a2bd51c2390e653fe9c9eeafedf73225066ec70c --- /dev/null +++ b/sleep_assessment/cases/case39/case39_05.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "ISRUC_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "label_path": "data/sleep/ISRUC_05.npy", + "instruction": "Given the sleep EEG/PSG file and paired sleep-stage label file, assess whole-night respiratory event clustering. Divide the recording into fixed, non-overlapping 5-minute windows starting at 0 minutes: 0-5, 5-10, 10-15, and so on. Use only complete 5-minute windows. Count apnea and hypopnea candidate events by their event onset time; an event belongs to exactly one half-open window [start, end). A cluster window is a 5-minute window with at least 3 apnea/hypopnea candidate events. For each cluster window, determine the dominant sleep stage from the ten 30-second sleep-stage labels in that window. If the dominant stage is tied, use the sleep stages at the candidate event onset epochs within the window to break the tie; if still tied, use W, N1, N2, N3, R order. Report the total number of cluster windows and the dominant stage across all cluster windows. If no cluster windows occur, report dominant stage as NULL. Allowed stage outputs are W, N1, N2, N3, R, and NULL. Respiratory channels for ISRUC: X5 = snore; X6 = Flow-1 (pressure-based airflow); DC3 = Flow-2 airflow; X7 and X8 = abdominal respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the total respiratory cluster-window count and dominant cluster sleep stage.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"cluster_window_count\" and \"dominant_cluster_stage\".\n4. \"cluster_window_count\" must be a FLOAT or null.\n5. \"dominant_cluster_stage\" must be one of \"W\", \"N1\", \"N2\", \"N3\", \"R\", \"NULL\", or null.\n6. If there are no respiratory cluster windows, normalize the stage to \"NULL\".\n7. If a value is missing or unclear, return null for that key.\n\n### OUTPUT TEMPLATE\n{\"cluster_window_count\": , \"dominant_cluster_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|\"NULL\"|null>}", + "metrics": [ + { + "metric_id": "respiratory_cluster_window_count_accuracy", + "type": "numeric_check", + "target_key": "cluster_window_count", + "weight": 60, + "params": { + "gt_value": 5.0, + "tolerance": 2.0 + } + }, + { + "metric_id": "respiratory_cluster_dominant_stage_accuracy", + "type": "categorical_check", + "target_key": "dominant_cluster_stage", + "weight": 40, + "params": { + "gt_value": "N2" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case39/case39_06.json b/sleep_assessment/cases/case39/case39_06.json new file mode 100644 index 0000000000000000000000000000000000000000..541af8b96777f68339db0e41036b9997cf670dd7 --- /dev/null +++ b/sleep_assessment/cases/case39/case39_06.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "SHHS1_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "label_path": "data/sleep/SHHS1_01.npy", + "instruction": "Given the sleep EEG/PSG file and paired sleep-stage label file, assess whole-night respiratory event clustering. Divide the recording into fixed, non-overlapping 5-minute windows starting at 0 minutes: 0-5, 5-10, 10-15, and so on. Use only complete 5-minute windows. Count apnea and hypopnea candidate events by their event onset time; an event belongs to exactly one half-open window [start, end). A cluster window is a 5-minute window with at least 3 apnea/hypopnea candidate events. For each cluster window, determine the dominant sleep stage from the ten 30-second sleep-stage labels in that window. If the dominant stage is tied, use the sleep stages at the candidate event onset epochs within the window to break the tie; if still tied, use W, N1, N2, N3, R order. Report the total number of cluster windows and the dominant stage across all cluster windows. If no cluster windows occur, report dominant stage as NULL. Allowed stage outputs are W, N1, N2, N3, R, and NULL. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the total respiratory cluster-window count and dominant cluster sleep stage.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"cluster_window_count\" and \"dominant_cluster_stage\".\n4. \"cluster_window_count\" must be a FLOAT or null.\n5. \"dominant_cluster_stage\" must be one of \"W\", \"N1\", \"N2\", \"N3\", \"R\", \"NULL\", or null.\n6. If there are no respiratory cluster windows, normalize the stage to \"NULL\".\n7. If a value is missing or unclear, return null for that key.\n\n### OUTPUT TEMPLATE\n{\"cluster_window_count\": , \"dominant_cluster_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|\"NULL\"|null>}", + "metrics": [ + { + "metric_id": "respiratory_cluster_window_count_accuracy", + "type": "numeric_check", + "target_key": "cluster_window_count", + "weight": 60, + "params": { + "gt_value": 12.0, + "tolerance": 4.0 + } + }, + { + "metric_id": "respiratory_cluster_dominant_stage_accuracy", + "type": "categorical_check", + "target_key": "dominant_cluster_stage", + "weight": 40, + "params": { + "gt_value": "R" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case39/case39_07.json b/sleep_assessment/cases/case39/case39_07.json new file mode 100644 index 0000000000000000000000000000000000000000..5737857a7a5615df89d60d03f4c0f0fc383e6f41 --- /dev/null +++ b/sleep_assessment/cases/case39/case39_07.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "SHHS1_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "label_path": "data/sleep/SHHS1_02.npy", + "instruction": "Given the sleep EEG/PSG file and paired sleep-stage label file, assess whole-night respiratory event clustering. Divide the recording into fixed, non-overlapping 5-minute windows starting at 0 minutes: 0-5, 5-10, 10-15, and so on. Use only complete 5-minute windows. Count apnea and hypopnea candidate events by their event onset time; an event belongs to exactly one half-open window [start, end). A cluster window is a 5-minute window with at least 3 apnea/hypopnea candidate events. For each cluster window, determine the dominant sleep stage from the ten 30-second sleep-stage labels in that window. If the dominant stage is tied, use the sleep stages at the candidate event onset epochs within the window to break the tie; if still tied, use W, N1, N2, N3, R order. Report the total number of cluster windows and the dominant stage across all cluster windows. If no cluster windows occur, report dominant stage as NULL. Allowed stage outputs are W, N1, N2, N3, R, and NULL. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the total respiratory cluster-window count and dominant cluster sleep stage.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"cluster_window_count\" and \"dominant_cluster_stage\".\n4. \"cluster_window_count\" must be a FLOAT or null.\n5. \"dominant_cluster_stage\" must be one of \"W\", \"N1\", \"N2\", \"N3\", \"R\", \"NULL\", or null.\n6. If there are no respiratory cluster windows, normalize the stage to \"NULL\".\n7. If a value is missing or unclear, return null for that key.\n\n### OUTPUT TEMPLATE\n{\"cluster_window_count\": , \"dominant_cluster_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|\"NULL\"|null>}", + "metrics": [ + { + "metric_id": "respiratory_cluster_window_count_accuracy", + "type": "numeric_check", + "target_key": "cluster_window_count", + "weight": 60, + "params": { + "gt_value": 32.0, + "tolerance": 6.0 + } + }, + { + "metric_id": "respiratory_cluster_dominant_stage_accuracy", + "type": "categorical_check", + "target_key": "dominant_cluster_stage", + "weight": 40, + "params": { + "gt_value": "N2" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case39/case39_08.json b/sleep_assessment/cases/case39/case39_08.json new file mode 100644 index 0000000000000000000000000000000000000000..5284a3d1f85c05d9b86127d2e8a5974c050ad2bd --- /dev/null +++ b/sleep_assessment/cases/case39/case39_08.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "SHHS1_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "label_path": "data/sleep/SHHS1_03.npy", + "instruction": "Given the sleep EEG/PSG file and paired sleep-stage label file, assess whole-night respiratory event clustering. Divide the recording into fixed, non-overlapping 5-minute windows starting at 0 minutes: 0-5, 5-10, 10-15, and so on. Use only complete 5-minute windows. Count apnea and hypopnea candidate events by their event onset time; an event belongs to exactly one half-open window [start, end). A cluster window is a 5-minute window with at least 3 apnea/hypopnea candidate events. For each cluster window, determine the dominant sleep stage from the ten 30-second sleep-stage labels in that window. If the dominant stage is tied, use the sleep stages at the candidate event onset epochs within the window to break the tie; if still tied, use W, N1, N2, N3, R order. Report the total number of cluster windows and the dominant stage across all cluster windows. If no cluster windows occur, report dominant stage as NULL. Allowed stage outputs are W, N1, N2, N3, R, and NULL. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the total respiratory cluster-window count and dominant cluster sleep stage.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"cluster_window_count\" and \"dominant_cluster_stage\".\n4. \"cluster_window_count\" must be a FLOAT or null.\n5. \"dominant_cluster_stage\" must be one of \"W\", \"N1\", \"N2\", \"N3\", \"R\", \"NULL\", or null.\n6. If there are no respiratory cluster windows, normalize the stage to \"NULL\".\n7. If a value is missing or unclear, return null for that key.\n\n### OUTPUT TEMPLATE\n{\"cluster_window_count\": , \"dominant_cluster_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|\"NULL\"|null>}", + "metrics": [ + { + "metric_id": "respiratory_cluster_window_count_accuracy", + "type": "numeric_check", + "target_key": "cluster_window_count", + "weight": 60, + "params": { + "gt_value": 20.0, + "tolerance": 4.0 + } + }, + { + "metric_id": "respiratory_cluster_dominant_stage_accuracy", + "type": "categorical_check", + "target_key": "dominant_cluster_stage", + "weight": 40, + "params": { + "gt_value": "R" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case39/case39_09.json b/sleep_assessment/cases/case39/case39_09.json new file mode 100644 index 0000000000000000000000000000000000000000..bfddd1a267f3755f1ba65de112a13488e8b8813d --- /dev/null +++ b/sleep_assessment/cases/case39/case39_09.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "SHHS1_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "label_path": "data/sleep/SHHS1_04.npy", + "instruction": "Given the sleep EEG/PSG file and paired sleep-stage label file, assess whole-night respiratory event clustering. Divide the recording into fixed, non-overlapping 5-minute windows starting at 0 minutes: 0-5, 5-10, 10-15, and so on. Use only complete 5-minute windows. Count apnea and hypopnea candidate events by their event onset time; an event belongs to exactly one half-open window [start, end). A cluster window is a 5-minute window with at least 3 apnea/hypopnea candidate events. For each cluster window, determine the dominant sleep stage from the ten 30-second sleep-stage labels in that window. If the dominant stage is tied, use the sleep stages at the candidate event onset epochs within the window to break the tie; if still tied, use W, N1, N2, N3, R order. Report the total number of cluster windows and the dominant stage across all cluster windows. If no cluster windows occur, report dominant stage as NULL. Allowed stage outputs are W, N1, N2, N3, R, and NULL. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the total respiratory cluster-window count and dominant cluster sleep stage.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"cluster_window_count\" and \"dominant_cluster_stage\".\n4. \"cluster_window_count\" must be a FLOAT or null.\n5. \"dominant_cluster_stage\" must be one of \"W\", \"N1\", \"N2\", \"N3\", \"R\", \"NULL\", or null.\n6. If there are no respiratory cluster windows, normalize the stage to \"NULL\".\n7. If a value is missing or unclear, return null for that key.\n\n### OUTPUT TEMPLATE\n{\"cluster_window_count\": , \"dominant_cluster_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|\"NULL\"|null>}", + "metrics": [ + { + "metric_id": "respiratory_cluster_window_count_accuracy", + "type": "numeric_check", + "target_key": "cluster_window_count", + "weight": 60, + "params": { + "gt_value": 21.0, + "tolerance": 5.0 + } + }, + { + "metric_id": "respiratory_cluster_dominant_stage_accuracy", + "type": "categorical_check", + "target_key": "dominant_cluster_stage", + "weight": 40, + "params": { + "gt_value": "R" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case39/case39_10.json b/sleep_assessment/cases/case39/case39_10.json new file mode 100644 index 0000000000000000000000000000000000000000..c917a1d02f022dbf5a1dff9d41d4d88b055583a3 --- /dev/null +++ b/sleep_assessment/cases/case39/case39_10.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "SHHS1_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "label_path": "data/sleep/SHHS1_05.npy", + "instruction": "Given the sleep EEG/PSG file and paired sleep-stage label file, assess whole-night respiratory event clustering. Divide the recording into fixed, non-overlapping 5-minute windows starting at 0 minutes: 0-5, 5-10, 10-15, and so on. Use only complete 5-minute windows. Count apnea and hypopnea candidate events by their event onset time; an event belongs to exactly one half-open window [start, end). A cluster window is a 5-minute window with at least 3 apnea/hypopnea candidate events. For each cluster window, determine the dominant sleep stage from the ten 30-second sleep-stage labels in that window. If the dominant stage is tied, use the sleep stages at the candidate event onset epochs within the window to break the tie; if still tied, use W, N1, N2, N3, R order. Report the total number of cluster windows and the dominant stage across all cluster windows. If no cluster windows occur, report dominant stage as NULL. Allowed stage outputs are W, N1, N2, N3, R, and NULL. Respiratory channels for SHHS1: THOR RES = thoracic respiratory effort; ABDO RES = abdominal respiratory effort; NEW AIR = nasal/oral airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the total respiratory cluster-window count and dominant cluster sleep stage.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"cluster_window_count\" and \"dominant_cluster_stage\".\n4. \"cluster_window_count\" must be a FLOAT or null.\n5. \"dominant_cluster_stage\" must be one of \"W\", \"N1\", \"N2\", \"N3\", \"R\", \"NULL\", or null.\n6. If there are no respiratory cluster windows, normalize the stage to \"NULL\".\n7. If a value is missing or unclear, return null for that key.\n\n### OUTPUT TEMPLATE\n{\"cluster_window_count\": , \"dominant_cluster_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|\"NULL\"|null>}", + "metrics": [ + { + "metric_id": "respiratory_cluster_window_count_accuracy", + "type": "numeric_check", + "target_key": "cluster_window_count", + "weight": 60, + "params": { + "gt_value": 15.0, + "tolerance": 4.0 + } + }, + { + "metric_id": "respiratory_cluster_dominant_stage_accuracy", + "type": "categorical_check", + "target_key": "dominant_cluster_stage", + "weight": 40, + "params": { + "gt_value": "R" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case39/case39_11.json b/sleep_assessment/cases/case39/case39_11.json new file mode 100644 index 0000000000000000000000000000000000000000..5a0913b096b8fad9729661f71d3ded4ac96e2d96 --- /dev/null +++ b/sleep_assessment/cases/case39/case39_11.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "Physionet2018_01.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "label_path": "data/sleep/Physionet2018_01.npy", + "instruction": "Given the sleep EEG/PSG file and paired sleep-stage label file, assess whole-night respiratory event clustering. Divide the recording into fixed, non-overlapping 5-minute windows starting at 0 minutes: 0-5, 5-10, 10-15, and so on. Use only complete 5-minute windows. Count apnea and hypopnea candidate events by their event onset time; an event belongs to exactly one half-open window [start, end). A cluster window is a 5-minute window with at least 3 apnea/hypopnea candidate events. For each cluster window, determine the dominant sleep stage from the ten 30-second sleep-stage labels in that window. If the dominant stage is tied, use the sleep stages at the candidate event onset epochs within the window to break the tie; if still tied, use W, N1, N2, N3, R order. Report the total number of cluster windows and the dominant stage across all cluster windows. If no cluster windows occur, report dominant stage as NULL. Allowed stage outputs are W, N1, N2, N3, R, and NULL. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the total respiratory cluster-window count and dominant cluster sleep stage.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"cluster_window_count\" and \"dominant_cluster_stage\".\n4. \"cluster_window_count\" must be a FLOAT or null.\n5. \"dominant_cluster_stage\" must be one of \"W\", \"N1\", \"N2\", \"N3\", \"R\", \"NULL\", or null.\n6. If there are no respiratory cluster windows, normalize the stage to \"NULL\".\n7. If a value is missing or unclear, return null for that key.\n\n### OUTPUT TEMPLATE\n{\"cluster_window_count\": , \"dominant_cluster_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|\"NULL\"|null>}", + "metrics": [ + { + "metric_id": "respiratory_cluster_window_count_accuracy", + "type": "numeric_check", + "target_key": "cluster_window_count", + "weight": 60, + "params": { + "gt_value": 4.0, + "tolerance": 2.0 + } + }, + { + "metric_id": "respiratory_cluster_dominant_stage_accuracy", + "type": "categorical_check", + "target_key": "dominant_cluster_stage", + "weight": 40, + "params": { + "gt_value": "N2" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case39/case39_12.json b/sleep_assessment/cases/case39/case39_12.json new file mode 100644 index 0000000000000000000000000000000000000000..8930dc43be51e6f340a1a2e8665780221833bb01 --- /dev/null +++ b/sleep_assessment/cases/case39/case39_12.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "Physionet2018_02.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "label_path": "data/sleep/Physionet2018_02.npy", + "instruction": "Given the sleep EEG/PSG file and paired sleep-stage label file, assess whole-night respiratory event clustering. Divide the recording into fixed, non-overlapping 5-minute windows starting at 0 minutes: 0-5, 5-10, 10-15, and so on. Use only complete 5-minute windows. Count apnea and hypopnea candidate events by their event onset time; an event belongs to exactly one half-open window [start, end). A cluster window is a 5-minute window with at least 3 apnea/hypopnea candidate events. For each cluster window, determine the dominant sleep stage from the ten 30-second sleep-stage labels in that window. If the dominant stage is tied, use the sleep stages at the candidate event onset epochs within the window to break the tie; if still tied, use W, N1, N2, N3, R order. Report the total number of cluster windows and the dominant stage across all cluster windows. If no cluster windows occur, report dominant stage as NULL. Allowed stage outputs are W, N1, N2, N3, R, and NULL. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the total respiratory cluster-window count and dominant cluster sleep stage.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"cluster_window_count\" and \"dominant_cluster_stage\".\n4. \"cluster_window_count\" must be a FLOAT or null.\n5. \"dominant_cluster_stage\" must be one of \"W\", \"N1\", \"N2\", \"N3\", \"R\", \"NULL\", or null.\n6. If there are no respiratory cluster windows, normalize the stage to \"NULL\".\n7. If a value is missing or unclear, return null for that key.\n\n### OUTPUT TEMPLATE\n{\"cluster_window_count\": , \"dominant_cluster_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|\"NULL\"|null>}", + "metrics": [ + { + "metric_id": "respiratory_cluster_window_count_accuracy", + "type": "numeric_check", + "target_key": "cluster_window_count", + "weight": 60, + "params": { + "gt_value": 23.0, + "tolerance": 5.0 + } + }, + { + "metric_id": "respiratory_cluster_dominant_stage_accuracy", + "type": "categorical_check", + "target_key": "dominant_cluster_stage", + "weight": 40, + "params": { + "gt_value": "N2" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case39/case39_13.json b/sleep_assessment/cases/case39/case39_13.json new file mode 100644 index 0000000000000000000000000000000000000000..efd2d08815dd3861e25687e158698d93fcfe105c --- /dev/null +++ b/sleep_assessment/cases/case39/case39_13.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "Physionet2018_03.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_03.edf", + "label_path": "data/sleep/Physionet2018_03.npy", + "instruction": "Given the sleep EEG/PSG file and paired sleep-stage label file, assess whole-night respiratory event clustering. Divide the recording into fixed, non-overlapping 5-minute windows starting at 0 minutes: 0-5, 5-10, 10-15, and so on. Use only complete 5-minute windows. Count apnea and hypopnea candidate events by their event onset time; an event belongs to exactly one half-open window [start, end). A cluster window is a 5-minute window with at least 3 apnea/hypopnea candidate events. For each cluster window, determine the dominant sleep stage from the ten 30-second sleep-stage labels in that window. If the dominant stage is tied, use the sleep stages at the candidate event onset epochs within the window to break the tie; if still tied, use W, N1, N2, N3, R order. Report the total number of cluster windows and the dominant stage across all cluster windows. If no cluster windows occur, report dominant stage as NULL. Allowed stage outputs are W, N1, N2, N3, R, and NULL. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the total respiratory cluster-window count and dominant cluster sleep stage.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"cluster_window_count\" and \"dominant_cluster_stage\".\n4. \"cluster_window_count\" must be a FLOAT or null.\n5. \"dominant_cluster_stage\" must be one of \"W\", \"N1\", \"N2\", \"N3\", \"R\", \"NULL\", or null.\n6. If there are no respiratory cluster windows, normalize the stage to \"NULL\".\n7. If a value is missing or unclear, return null for that key.\n\n### OUTPUT TEMPLATE\n{\"cluster_window_count\": , \"dominant_cluster_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|\"NULL\"|null>}", + "metrics": [ + { + "metric_id": "respiratory_cluster_window_count_accuracy", + "type": "numeric_check", + "target_key": "cluster_window_count", + "weight": 60, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + }, + { + "metric_id": "respiratory_cluster_dominant_stage_accuracy", + "type": "categorical_check", + "target_key": "dominant_cluster_stage", + "weight": 40, + "params": { + "gt_value": "NULL" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case39/case39_14.json b/sleep_assessment/cases/case39/case39_14.json new file mode 100644 index 0000000000000000000000000000000000000000..10b0a6e37ff89401dcd0a27b8600e113c87b7ec3 --- /dev/null +++ b/sleep_assessment/cases/case39/case39_14.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "Physionet2018_04.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "label_path": "data/sleep/Physionet2018_04.npy", + "instruction": "Given the sleep EEG/PSG file and paired sleep-stage label file, assess whole-night respiratory event clustering. Divide the recording into fixed, non-overlapping 5-minute windows starting at 0 minutes: 0-5, 5-10, 10-15, and so on. Use only complete 5-minute windows. Count apnea and hypopnea candidate events by their event onset time; an event belongs to exactly one half-open window [start, end). A cluster window is a 5-minute window with at least 3 apnea/hypopnea candidate events. For each cluster window, determine the dominant sleep stage from the ten 30-second sleep-stage labels in that window. If the dominant stage is tied, use the sleep stages at the candidate event onset epochs within the window to break the tie; if still tied, use W, N1, N2, N3, R order. Report the total number of cluster windows and the dominant stage across all cluster windows. If no cluster windows occur, report dominant stage as NULL. Allowed stage outputs are W, N1, N2, N3, R, and NULL. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the total respiratory cluster-window count and dominant cluster sleep stage.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"cluster_window_count\" and \"dominant_cluster_stage\".\n4. \"cluster_window_count\" must be a FLOAT or null.\n5. \"dominant_cluster_stage\" must be one of \"W\", \"N1\", \"N2\", \"N3\", \"R\", \"NULL\", or null.\n6. If there are no respiratory cluster windows, normalize the stage to \"NULL\".\n7. If a value is missing or unclear, return null for that key.\n\n### OUTPUT TEMPLATE\n{\"cluster_window_count\": , \"dominant_cluster_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|\"NULL\"|null>}", + "metrics": [ + { + "metric_id": "respiratory_cluster_window_count_accuracy", + "type": "numeric_check", + "target_key": "cluster_window_count", + "weight": 60, + "params": { + "gt_value": 25.0, + "tolerance": 5.0 + } + }, + { + "metric_id": "respiratory_cluster_dominant_stage_accuracy", + "type": "categorical_check", + "target_key": "dominant_cluster_stage", + "weight": 40, + "params": { + "gt_value": "N2" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case39/case39_15.json b/sleep_assessment/cases/case39/case39_15.json new file mode 100644 index 0000000000000000000000000000000000000000..0ae1b1bb3961e624cedf8e08ba16f889d3aeb1b5 --- /dev/null +++ b/sleep_assessment/cases/case39/case39_15.json @@ -0,0 +1,37 @@ +{ + "meta_info": { + "case_id": "Physionet2018_05.edf", + "bench_subset": "NeuroBench-Sleep", + "difficult": 2, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "label_path": "data/sleep/Physionet2018_05.npy", + "instruction": "Given the sleep EEG/PSG file and paired sleep-stage label file, assess whole-night respiratory event clustering. Divide the recording into fixed, non-overlapping 5-minute windows starting at 0 minutes: 0-5, 5-10, 10-15, and so on. Use only complete 5-minute windows. Count apnea and hypopnea candidate events by their event onset time; an event belongs to exactly one half-open window [start, end). A cluster window is a 5-minute window with at least 3 apnea/hypopnea candidate events. For each cluster window, determine the dominant sleep stage from the ten 30-second sleep-stage labels in that window. If the dominant stage is tied, use the sleep stages at the candidate event onset epochs within the window to break the tie; if still tied, use W, N1, N2, N3, R order. Report the total number of cluster windows and the dominant stage across all cluster windows. If no cluster windows occur, report dominant stage as NULL. Allowed stage outputs are W, N1, N2, N3, R, and NULL. Respiratory channels for Physionet2018: ABD = abdominal respiratory effort; CHEST = thoracic respiratory effort; AIRFLOW = airflow; SaO2 = oximetry." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the total respiratory cluster-window count and dominant cluster sleep stage.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"cluster_window_count\" and \"dominant_cluster_stage\".\n4. \"cluster_window_count\" must be a FLOAT or null.\n5. \"dominant_cluster_stage\" must be one of \"W\", \"N1\", \"N2\", \"N3\", \"R\", \"NULL\", or null.\n6. If there are no respiratory cluster windows, normalize the stage to \"NULL\".\n7. If a value is missing or unclear, return null for that key.\n\n### OUTPUT TEMPLATE\n{\"cluster_window_count\": , \"dominant_cluster_stage\": <\"W\"|\"N1\"|\"N2\"|\"N3\"|\"R\"|\"NULL\"|null>}", + "metrics": [ + { + "metric_id": "respiratory_cluster_window_count_accuracy", + "type": "numeric_check", + "target_key": "cluster_window_count", + "weight": 60, + "params": { + "gt_value": 42.0, + "tolerance": 6.0 + } + }, + { + "metric_id": "respiratory_cluster_dominant_stage_accuracy", + "type": "categorical_check", + "target_key": "dominant_cluster_stage", + "weight": 40, + "params": { + "gt_value": "N2" + } + } + ] + } +} \ No newline at end of file diff --git a/sleep_assessment/cases/case40/case40_01.json b/sleep_assessment/cases/case40/case40_01.json new file mode 100644 index 0000000000000000000000000000000000000000..2174a53758492d1d41b6564d3253579b7f9b8d35 --- /dev/null +++ b/sleep_assessment/cases/case40/case40_01.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "case40_01", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Use the sleep file and paired sleep-stage labels to calculate sleep-period ODI3 and ODI4. ISRUC channels: SaO2=oxygen saturation; X5=snore; X6/DC3=airflow; X7/X8=abdominal effort. Report final ODI3 and ODI4 in events/hour." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the whole-night sleep oxygen desaturation indices ODI3 and ODI4.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"odi3\" and \"odi4\".\n4. Both values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"odi3\": , \"odi4\": }", + "metrics": [ + { + "metric_id": "oxygen_desaturation_index_3_accuracy", + "type": "numeric_check", + "target_key": "odi3", + "weight": 50, + "params": { + "gt_value": 2.727272727272727, + "tolerance": 0.13636363636363635 + } + }, + { + "metric_id": "oxygen_desaturation_index_4_accuracy", + "type": "numeric_check", + "target_key": "odi4", + "weight": 50, + "params": { + "gt_value": 2.727272727272727, + "tolerance": 0.13636363636363635 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case40/case40_02.json b/sleep_assessment/cases/case40/case40_02.json new file mode 100644 index 0000000000000000000000000000000000000000..45a0d0384c7499c61c610010c93075a6778c37c9 --- /dev/null +++ b/sleep_assessment/cases/case40/case40_02.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "case40_02", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "label_path": "data/sleep/ISRUC_02.npy", + "instruction": "Use the sleep file and paired sleep-stage labels to calculate sleep-period ODI3 and ODI4. ISRUC channels: SaO2=oxygen saturation; X5=snore; X6/DC3=airflow; X7/X8=abdominal effort. Report final ODI3 and ODI4 in events/hour." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the whole-night sleep oxygen desaturation indices ODI3 and ODI4.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"odi3\" and \"odi4\".\n4. Both values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"odi3\": , \"odi4\": }", + "metrics": [ + { + "metric_id": "oxygen_desaturation_index_3_accuracy", + "type": "numeric_check", + "target_key": "odi3", + "weight": 50, + "params": { + "gt_value": 0.3337969401947149, + "tolerance": 0.016689847009735744 + } + }, + { + "metric_id": "oxygen_desaturation_index_4_accuracy", + "type": "numeric_check", + "target_key": "odi4", + "weight": 50, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case40/case40_03.json b/sleep_assessment/cases/case40/case40_03.json new file mode 100644 index 0000000000000000000000000000000000000000..cac0577e7942573d01871d40583b692bbd33b4b2 --- /dev/null +++ b/sleep_assessment/cases/case40/case40_03.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "case40_03", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "label_path": "data/sleep/ISRUC_03.npy", + "instruction": "Use the sleep file and paired sleep-stage labels to calculate sleep-period ODI3 and ODI4. ISRUC channels: SaO2=oxygen saturation; X5=snore; X6/DC3=airflow; X7/X8=abdominal effort. Report final ODI3 and ODI4 in events/hour." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the whole-night sleep oxygen desaturation indices ODI3 and ODI4.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"odi3\" and \"odi4\".\n4. Both values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"odi3\": , \"odi4\": }", + "metrics": [ + { + "metric_id": "oxygen_desaturation_index_3_accuracy", + "type": "numeric_check", + "target_key": "odi3", + "weight": 50, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + }, + { + "metric_id": "oxygen_desaturation_index_4_accuracy", + "type": "numeric_check", + "target_key": "odi4", + "weight": 50, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case40/case40_04.json b/sleep_assessment/cases/case40/case40_04.json new file mode 100644 index 0000000000000000000000000000000000000000..ce0c03cf9f89fc6e2ec763f691bf65542e70f9db --- /dev/null +++ b/sleep_assessment/cases/case40/case40_04.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "case40_04", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "label_path": "data/sleep/ISRUC_05.npy", + "instruction": "Use the sleep file and paired sleep-stage labels to calculate sleep-period ODI3 and ODI4. ISRUC channels: SaO2=oxygen saturation; X5=snore; X6/DC3=airflow; X7/X8=abdominal effort. Report final ODI3 and ODI4 in events/hour." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the whole-night sleep oxygen desaturation indices ODI3 and ODI4.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"odi3\" and \"odi4\".\n4. Both values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"odi3\": , \"odi4\": }", + "metrics": [ + { + "metric_id": "oxygen_desaturation_index_3_accuracy", + "type": "numeric_check", + "target_key": "odi3", + "weight": 50, + "params": { + "gt_value": 1.8652849740932642, + "tolerance": 0.09326424870466321 + } + }, + { + "metric_id": "oxygen_desaturation_index_4_accuracy", + "type": "numeric_check", + "target_key": "odi4", + "weight": 50, + "params": { + "gt_value": 0.20725388601036268, + "tolerance": 0.010362694300518135 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case40/case40_05.json b/sleep_assessment/cases/case40/case40_05.json new file mode 100644 index 0000000000000000000000000000000000000000..4c19c6f8249a03eaecad430b0e8c4dc5e6e9ad7e --- /dev/null +++ b/sleep_assessment/cases/case40/case40_05.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "case40_05", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "label_path": "data/sleep/SHHS1_01.npy", + "instruction": "Use the sleep file and paired sleep-stage labels to calculate sleep-period ODI3 and ODI4. SHHS1 channels: SaO2=oxygen saturation; THOR RES/ABDO RES=effort; NEW AIR/AIRFLOW=airflow. Report final ODI3 and ODI4 in events/hour." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the whole-night sleep oxygen desaturation indices ODI3 and ODI4.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"odi3\" and \"odi4\".\n4. Both values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"odi3\": , \"odi4\": }", + "metrics": [ + { + "metric_id": "oxygen_desaturation_index_3_accuracy", + "type": "numeric_check", + "target_key": "odi3", + "weight": 50, + "params": { + "gt_value": 1.4380825565912116, + "tolerance": 0.07190412782956059 + } + }, + { + "metric_id": "oxygen_desaturation_index_4_accuracy", + "type": "numeric_check", + "target_key": "odi4", + "weight": 50, + "params": { + "gt_value": 0.6391478029294274, + "tolerance": 0.03195739014647137 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case40/case40_06.json b/sleep_assessment/cases/case40/case40_06.json new file mode 100644 index 0000000000000000000000000000000000000000..e30cd06bf39e39e9919feec18f1c21e5aaae7cd8 --- /dev/null +++ b/sleep_assessment/cases/case40/case40_06.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "case40_06", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "label_path": "data/sleep/SHHS1_02.npy", + "instruction": "Use the sleep file and paired sleep-stage labels to calculate sleep-period ODI3 and ODI4. SHHS1 channels: SaO2=oxygen saturation; THOR RES/ABDO RES=effort; NEW AIR/AIRFLOW=airflow. Report final ODI3 and ODI4 in events/hour." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the whole-night sleep oxygen desaturation indices ODI3 and ODI4.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"odi3\" and \"odi4\".\n4. Both values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"odi3\": , \"odi4\": }", + "metrics": [ + { + "metric_id": "oxygen_desaturation_index_3_accuracy", + "type": "numeric_check", + "target_key": "odi3", + "weight": 50, + "params": { + "gt_value": 7.582417582417583, + "tolerance": 0.37912087912087916 + } + }, + { + "metric_id": "oxygen_desaturation_index_4_accuracy", + "type": "numeric_check", + "target_key": "odi4", + "weight": 50, + "params": { + "gt_value": 6.593406593406594, + "tolerance": 0.3296703296703297 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case40/case40_07.json b/sleep_assessment/cases/case40/case40_07.json new file mode 100644 index 0000000000000000000000000000000000000000..772393231bcc60634952e486fc843abe4161ba1a --- /dev/null +++ b/sleep_assessment/cases/case40/case40_07.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "case40_07", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "label_path": "data/sleep/SHHS1_03.npy", + "instruction": "Use the sleep file and paired sleep-stage labels to calculate sleep-period ODI3 and ODI4. SHHS1 channels: SaO2=oxygen saturation; THOR RES/ABDO RES=effort; NEW AIR/AIRFLOW=airflow. Report final ODI3 and ODI4 in events/hour." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the whole-night sleep oxygen desaturation indices ODI3 and ODI4.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"odi3\" and \"odi4\".\n4. Both values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"odi3\": , \"odi4\": }", + "metrics": [ + { + "metric_id": "oxygen_desaturation_index_3_accuracy", + "type": "numeric_check", + "target_key": "odi3", + "weight": 50, + "params": { + "gt_value": 5.020920502092051, + "tolerance": 0.25104602510460255 + } + }, + { + "metric_id": "oxygen_desaturation_index_4_accuracy", + "type": "numeric_check", + "target_key": "odi4", + "weight": 50, + "params": { + "gt_value": 2.5104602510460254, + "tolerance": 0.12552301255230128 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case40/case40_08.json b/sleep_assessment/cases/case40/case40_08.json new file mode 100644 index 0000000000000000000000000000000000000000..8911a43bcde35152616c614ad5f6e5202c7683a9 --- /dev/null +++ b/sleep_assessment/cases/case40/case40_08.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "case40_08", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "label_path": "data/sleep/SHHS1_04.npy", + "instruction": "Use the sleep file and paired sleep-stage labels to calculate sleep-period ODI3 and ODI4. SHHS1 channels: SaO2=oxygen saturation; THOR RES/ABDO RES=effort; NEW AIR/AIRFLOW=airflow. Report final ODI3 and ODI4 in events/hour." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the whole-night sleep oxygen desaturation indices ODI3 and ODI4.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"odi3\" and \"odi4\".\n4. Both values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"odi3\": , \"odi4\": }", + "metrics": [ + { + "metric_id": "oxygen_desaturation_index_3_accuracy", + "type": "numeric_check", + "target_key": "odi3", + "weight": 50, + "params": { + "gt_value": 0.9966777408637874, + "tolerance": 0.04983388704318937 + } + }, + { + "metric_id": "oxygen_desaturation_index_4_accuracy", + "type": "numeric_check", + "target_key": "odi4", + "weight": 50, + "params": { + "gt_value": 0.5980066445182725, + "tolerance": 0.029900332225913623 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case40/case40_09.json b/sleep_assessment/cases/case40/case40_09.json new file mode 100644 index 0000000000000000000000000000000000000000..ef5ce910f06d24095057264105c8f7e2db9e607b --- /dev/null +++ b/sleep_assessment/cases/case40/case40_09.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "case40_09", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "label_path": "data/sleep/SHHS1_05.npy", + "instruction": "Use the sleep file and paired sleep-stage labels to calculate sleep-period ODI3 and ODI4. SHHS1 channels: SaO2=oxygen saturation; THOR RES/ABDO RES=effort; NEW AIR/AIRFLOW=airflow. Report final ODI3 and ODI4 in events/hour." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the whole-night sleep oxygen desaturation indices ODI3 and ODI4.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"odi3\" and \"odi4\".\n4. Both values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"odi3\": , \"odi4\": }", + "metrics": [ + { + "metric_id": "oxygen_desaturation_index_3_accuracy", + "type": "numeric_check", + "target_key": "odi3", + "weight": 50, + "params": { + "gt_value": 1.9459459459459458, + "tolerance": 0.0972972972972973 + } + }, + { + "metric_id": "oxygen_desaturation_index_4_accuracy", + "type": "numeric_check", + "target_key": "odi4", + "weight": 50, + "params": { + "gt_value": 1.6216216216216215, + "tolerance": 0.08108108108108109 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case40/case40_10.json b/sleep_assessment/cases/case40/case40_10.json new file mode 100644 index 0000000000000000000000000000000000000000..2664e5226fa40f44b882ebf3f3413283061f91cd --- /dev/null +++ b/sleep_assessment/cases/case40/case40_10.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "case40_10", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "label_path": "data/sleep/Physionet2018_01.npy", + "instruction": "Use the sleep file and paired sleep-stage labels to calculate sleep-period ODI3 and ODI4. Physionet2018 channels: SaO2=oxygen saturation; ABD/CHEST=effort; AIRFLOW=airflow. Report final ODI3 and ODI4 in events/hour." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the whole-night sleep oxygen desaturation indices ODI3 and ODI4.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"odi3\" and \"odi4\".\n4. Both values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"odi3\": , \"odi4\": }", + "metrics": [ + { + "metric_id": "oxygen_desaturation_index_3_accuracy", + "type": "numeric_check", + "target_key": "odi3", + "weight": 50, + "params": { + "gt_value": 8.535564853556485, + "tolerance": 0.42677824267782427 + } + }, + { + "metric_id": "oxygen_desaturation_index_4_accuracy", + "type": "numeric_check", + "target_key": "odi4", + "weight": 50, + "params": { + "gt_value": 2.510460251046025, + "tolerance": 0.12552301255230125 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case40/case40_11.json b/sleep_assessment/cases/case40/case40_11.json new file mode 100644 index 0000000000000000000000000000000000000000..0a738ef2fb26b0601066a496b6fe76c6aa4b294e --- /dev/null +++ b/sleep_assessment/cases/case40/case40_11.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "case40_11", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "label_path": "data/sleep/Physionet2018_02.npy", + "instruction": "Use the sleep file and paired sleep-stage labels to calculate sleep-period ODI3 and ODI4. Physionet2018 channels: SaO2=oxygen saturation; ABD/CHEST=effort; AIRFLOW=airflow. Report final ODI3 and ODI4 in events/hour." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the whole-night sleep oxygen desaturation indices ODI3 and ODI4.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"odi3\" and \"odi4\".\n4. Both values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"odi3\": , \"odi4\": }", + "metrics": [ + { + "metric_id": "oxygen_desaturation_index_3_accuracy", + "type": "numeric_check", + "target_key": "odi3", + "weight": 50, + "params": { + "gt_value": 9.391304347826086, + "tolerance": 0.4695652173913043 + } + }, + { + "metric_id": "oxygen_desaturation_index_4_accuracy", + "type": "numeric_check", + "target_key": "odi4", + "weight": 50, + "params": { + "gt_value": 8.869565217391305, + "tolerance": 0.44347826086956527 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case40/case40_12.json b/sleep_assessment/cases/case40/case40_12.json new file mode 100644 index 0000000000000000000000000000000000000000..e044db24f0b130ce98174f5a9f05ec4c0a3a1d43 --- /dev/null +++ b/sleep_assessment/cases/case40/case40_12.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "case40_12", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_03.edf", + "label_path": "data/sleep/Physionet2018_03.npy", + "instruction": "Use the sleep file and paired sleep-stage labels to calculate sleep-period ODI3 and ODI4. Physionet2018 channels: SaO2=oxygen saturation; ABD/CHEST=effort; AIRFLOW=airflow. Report final ODI3 and ODI4 in events/hour." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the whole-night sleep oxygen desaturation indices ODI3 and ODI4.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"odi3\" and \"odi4\".\n4. Both values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"odi3\": , \"odi4\": }", + "metrics": [ + { + "metric_id": "oxygen_desaturation_index_3_accuracy", + "type": "numeric_check", + "target_key": "odi3", + "weight": 50, + "params": { + "gt_value": 1.0619469026548674, + "tolerance": 0.05309734513274337 + } + }, + { + "metric_id": "oxygen_desaturation_index_4_accuracy", + "type": "numeric_check", + "target_key": "odi4", + "weight": 50, + "params": { + "gt_value": 0.5309734513274337, + "tolerance": 0.026548672566371685 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case40/case40_13.json b/sleep_assessment/cases/case40/case40_13.json new file mode 100644 index 0000000000000000000000000000000000000000..f89ccd650188bb0baf18590c9a6f43e713bd946f --- /dev/null +++ b/sleep_assessment/cases/case40/case40_13.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "case40_13", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "label_path": "data/sleep/Physionet2018_04.npy", + "instruction": "Use the sleep file and paired sleep-stage labels to calculate sleep-period ODI3 and ODI4. Physionet2018 channels: SaO2=oxygen saturation; ABD/CHEST=effort; AIRFLOW=airflow. Report final ODI3 and ODI4 in events/hour." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the whole-night sleep oxygen desaturation indices ODI3 and ODI4.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"odi3\" and \"odi4\".\n4. Both values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"odi3\": , \"odi4\": }", + "metrics": [ + { + "metric_id": "oxygen_desaturation_index_3_accuracy", + "type": "numeric_check", + "target_key": "odi3", + "weight": 50, + "params": { + "gt_value": 13.211009174311927, + "tolerance": 0.6605504587155964 + } + }, + { + "metric_id": "oxygen_desaturation_index_4_accuracy", + "type": "numeric_check", + "target_key": "odi4", + "weight": 50, + "params": { + "gt_value": 10.851900393184797, + "tolerance": 0.5425950196592398 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case40/case40_14.json b/sleep_assessment/cases/case40/case40_14.json new file mode 100644 index 0000000000000000000000000000000000000000..17a94b6323d0cb895321da491263132db69927cd --- /dev/null +++ b/sleep_assessment/cases/case40/case40_14.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "case40_14", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "label_path": "data/sleep/Physionet2018_05.npy", + "instruction": "Use the sleep file and paired sleep-stage labels to calculate sleep-period ODI3 and ODI4. Physionet2018 channels: SaO2=oxygen saturation; ABD/CHEST=effort; AIRFLOW=airflow. Report final ODI3 and ODI4 in events/hour." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract the whole-night sleep oxygen desaturation indices ODI3 and ODI4.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. The keys must be EXACTLY \"odi3\" and \"odi4\".\n4. Both values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"odi3\": , \"odi4\": }", + "metrics": [ + { + "metric_id": "oxygen_desaturation_index_3_accuracy", + "type": "numeric_check", + "target_key": "odi3", + "weight": 50, + "params": { + "gt_value": 10.266666666666667, + "tolerance": 0.5133333333333334 + } + }, + { + "metric_id": "oxygen_desaturation_index_4_accuracy", + "type": "numeric_check", + "target_key": "odi4", + "weight": 50, + "params": { + "gt_value": 8.933333333333334, + "tolerance": 0.4466666666666667 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case41/case41_01.json b/sleep_assessment/cases/case41/case41_01.json new file mode 100644 index 0000000000000000000000000000000000000000..2f8201e1257b1f14bc28933627a1d884070114c5 --- /dev/null +++ b/sleep_assessment/cases/case41/case41_01.json @@ -0,0 +1,48 @@ +{ + "meta_info": { + "case_id": "case41_01", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Use the sleep file and paired labels to calculate sleep mean SpO2, wake mean SpO2, and sleep minimum SpO2. Convert SaO2 to 1 Hz by per-second median; accept 50-100%, remove isolated <=2-second outliers more than 4 points from a centered 9-second median, interpolate only bounded internal gaps <=30 seconds, apply a centered 9-second median, and exclude all remaining missing seconds. Treat N1/N2/N3/R as sleep and W as wake. Report percentages. ISRUC: SaO2=oxygen saturation; X5=snore; X6/DC3=airflow; X7/X8=abdominal effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract sleep mean SpO2, wake mean SpO2, and sleep minimum SpO2.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. Keys must be EXACTLY \"sleep_mean_spo2\", \"wake_mean_spo2\", and \"sleep_min_spo2\".\n4. Values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"sleep_mean_spo2\": , \"wake_mean_spo2\": , \"sleep_min_spo2\": }", + "metrics": [ + { + "metric_id": "sleep_mean_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "sleep_mean_spo2", + "weight": 33, + "params": { + "gt_value": 95.31661652458806, + "tolerance": 4.765830826229403 + } + }, + { + "metric_id": "wake_mean_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "wake_mean_spo2", + "weight": 33, + "params": { + "gt_value": 96.99387057428399, + "tolerance": 4.8496935287142 + } + }, + { + "metric_id": "sleep_min_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "sleep_min_spo2", + "weight": 34, + "params": { + "gt_value": 80.18999154305723, + "tolerance": 4.009499577152861 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case41/case41_02.json b/sleep_assessment/cases/case41/case41_02.json new file mode 100644 index 0000000000000000000000000000000000000000..ce859a90851b0bc2c729200c230d3c86dc96c4cc --- /dev/null +++ b/sleep_assessment/cases/case41/case41_02.json @@ -0,0 +1,48 @@ +{ + "meta_info": { + "case_id": "case41_02", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "label_path": "data/sleep/ISRUC_02.npy", + "instruction": "Use the sleep file and paired labels to calculate sleep mean SpO2, wake mean SpO2, and sleep minimum SpO2. Convert SaO2 to 1 Hz by per-second median; accept 50-100%, remove isolated <=2-second outliers more than 4 points from a centered 9-second median, interpolate only bounded internal gaps <=30 seconds, apply a centered 9-second median, and exclude all remaining missing seconds. Treat N1/N2/N3/R as sleep and W as wake. Report percentages. ISRUC: SaO2=oxygen saturation; X5=snore; X6/DC3=airflow; X7/X8=abdominal effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract sleep mean SpO2, wake mean SpO2, and sleep minimum SpO2.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. Keys must be EXACTLY \"sleep_mean_spo2\", \"wake_mean_spo2\", and \"sleep_min_spo2\".\n4. Values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"sleep_mean_spo2\": , \"wake_mean_spo2\": , \"sleep_min_spo2\": }", + "metrics": [ + { + "metric_id": "sleep_mean_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "sleep_mean_spo2", + "weight": 33, + "params": { + "gt_value": 96.35141244472393, + "tolerance": 4.817570622236197 + } + }, + { + "metric_id": "wake_mean_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "wake_mean_spo2", + "weight": 33, + "params": { + "gt_value": 96.3768491908515, + "tolerance": 4.818842459542576 + } + }, + { + "metric_id": "sleep_min_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "sleep_min_spo2", + "weight": 34, + "params": { + "gt_value": 90.3370649259306, + "tolerance": 4.51685324629653 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case41/case41_03.json b/sleep_assessment/cases/case41/case41_03.json new file mode 100644 index 0000000000000000000000000000000000000000..54f01b11932c956112476c01515edfaa181bdd01 --- /dev/null +++ b/sleep_assessment/cases/case41/case41_03.json @@ -0,0 +1,48 @@ +{ + "meta_info": { + "case_id": "case41_03", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "label_path": "data/sleep/ISRUC_03.npy", + "instruction": "Use the sleep file and paired labels to calculate sleep mean SpO2, wake mean SpO2, and sleep minimum SpO2. Convert SaO2 to 1 Hz by per-second median; accept 50-100%, remove isolated <=2-second outliers more than 4 points from a centered 9-second median, interpolate only bounded internal gaps <=30 seconds, apply a centered 9-second median, and exclude all remaining missing seconds. Treat N1/N2/N3/R as sleep and W as wake. Report percentages. ISRUC: SaO2=oxygen saturation; X5=snore; X6/DC3=airflow; X7/X8=abdominal effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract sleep mean SpO2, wake mean SpO2, and sleep minimum SpO2.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. Keys must be EXACTLY \"sleep_mean_spo2\", \"wake_mean_spo2\", and \"sleep_min_spo2\".\n4. Values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"sleep_mean_spo2\": , \"wake_mean_spo2\": , \"sleep_min_spo2\": }", + "metrics": [ + { + "metric_id": "sleep_mean_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "sleep_mean_spo2", + "weight": 33, + "params": { + "gt_value": 99.68644538175248, + "tolerance": 4.984322269087624 + } + }, + { + "metric_id": "wake_mean_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "wake_mean_spo2", + "weight": 33, + "params": { + "gt_value": 99.7551652699975, + "tolerance": 4.987758263499876 + } + }, + { + "metric_id": "sleep_min_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "sleep_min_spo2", + "weight": 34, + "params": { + "gt_value": 97.4486915388724, + "tolerance": 4.87243457694362 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case41/case41_04.json b/sleep_assessment/cases/case41/case41_04.json new file mode 100644 index 0000000000000000000000000000000000000000..b5036c3783751566f56c1d4d408bdcd17c0e8ecf --- /dev/null +++ b/sleep_assessment/cases/case41/case41_04.json @@ -0,0 +1,48 @@ +{ + "meta_info": { + "case_id": "case41_04", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "label_path": "data/sleep/ISRUC_05.npy", + "instruction": "Use the sleep file and paired labels to calculate sleep mean SpO2, wake mean SpO2, and sleep minimum SpO2. Convert SaO2 to 1 Hz by per-second median; accept 50-100%, remove isolated <=2-second outliers more than 4 points from a centered 9-second median, interpolate only bounded internal gaps <=30 seconds, apply a centered 9-second median, and exclude all remaining missing seconds. Treat N1/N2/N3/R as sleep and W as wake. Report percentages. ISRUC: SaO2=oxygen saturation; X5=snore; X6/DC3=airflow; X7/X8=abdominal effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract sleep mean SpO2, wake mean SpO2, and sleep minimum SpO2.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. Keys must be EXACTLY \"sleep_mean_spo2\", \"wake_mean_spo2\", and \"sleep_min_spo2\".\n4. Values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"sleep_mean_spo2\": , \"wake_mean_spo2\": , \"sleep_min_spo2\": }", + "metrics": [ + { + "metric_id": "sleep_mean_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "sleep_mean_spo2", + "weight": 33, + "params": { + "gt_value": 98.01404911485601, + "tolerance": 4.900702455742801 + } + }, + { + "metric_id": "wake_mean_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "wake_mean_spo2", + "weight": 33, + "params": { + "gt_value": 99.07532700594251, + "tolerance": 4.953766350297126 + } + }, + { + "metric_id": "sleep_min_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "sleep_min_spo2", + "weight": 34, + "params": { + "gt_value": 92.3925677386931, + "tolerance": 4.6196283869346555 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case41/case41_05.json b/sleep_assessment/cases/case41/case41_05.json new file mode 100644 index 0000000000000000000000000000000000000000..315e8efeedd3de8dd3118bee21505b695d62ba93 --- /dev/null +++ b/sleep_assessment/cases/case41/case41_05.json @@ -0,0 +1,48 @@ +{ + "meta_info": { + "case_id": "case41_05", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "label_path": "data/sleep/SHHS1_01.npy", + "instruction": "Use the sleep file and paired labels to calculate sleep mean SpO2, wake mean SpO2, and sleep minimum SpO2. Convert SaO2 to 1 Hz by per-second median; accept 50-100%, remove isolated <=2-second outliers more than 4 points from a centered 9-second median, interpolate only bounded internal gaps <=30 seconds, apply a centered 9-second median, and exclude all remaining missing seconds. Treat N1/N2/N3/R as sleep and W as wake. Report percentages. SHHS1: SaO2=oxygen saturation; THOR RES/ABDO RES=effort; NEW AIR/AIRFLOW=airflow." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract sleep mean SpO2, wake mean SpO2, and sleep minimum SpO2.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. Keys must be EXACTLY \"sleep_mean_spo2\", \"wake_mean_spo2\", and \"sleep_min_spo2\".\n4. Values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"sleep_mean_spo2\": , \"wake_mean_spo2\": , \"sleep_min_spo2\": }", + "metrics": [ + { + "metric_id": "sleep_mean_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "sleep_mean_spo2", + "weight": 33, + "params": { + "gt_value": 93.06258265255414, + "tolerance": 4.653129132627707 + } + }, + { + "metric_id": "wake_mean_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "wake_mean_spo2", + "weight": 33, + "params": { + "gt_value": 93.52465920576844, + "tolerance": 4.676232960288423 + } + }, + { + "metric_id": "sleep_min_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "sleep_min_spo2", + "weight": 34, + "params": { + "gt_value": 85.30551749480685, + "tolerance": 4.265275874740342 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case41/case41_06.json b/sleep_assessment/cases/case41/case41_06.json new file mode 100644 index 0000000000000000000000000000000000000000..a7ce9ee14450f73c6146f67daadf314c73e480cd --- /dev/null +++ b/sleep_assessment/cases/case41/case41_06.json @@ -0,0 +1,48 @@ +{ + "meta_info": { + "case_id": "case41_06", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "label_path": "data/sleep/SHHS1_02.npy", + "instruction": "Use the sleep file and paired labels to calculate sleep mean SpO2, wake mean SpO2, and sleep minimum SpO2. Convert SaO2 to 1 Hz by per-second median; accept 50-100%, remove isolated <=2-second outliers more than 4 points from a centered 9-second median, interpolate only bounded internal gaps <=30 seconds, apply a centered 9-second median, and exclude all remaining missing seconds. Treat N1/N2/N3/R as sleep and W as wake. Report percentages. SHHS1: SaO2=oxygen saturation; THOR RES/ABDO RES=effort; NEW AIR/AIRFLOW=airflow." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract sleep mean SpO2, wake mean SpO2, and sleep minimum SpO2.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. Keys must be EXACTLY \"sleep_mean_spo2\", \"wake_mean_spo2\", and \"sleep_min_spo2\".\n4. Values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"sleep_mean_spo2\": , \"wake_mean_spo2\": , \"sleep_min_spo2\": }", + "metrics": [ + { + "metric_id": "sleep_mean_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "sleep_mean_spo2", + "weight": 33, + "params": { + "gt_value": 94.5460224126736, + "tolerance": 4.72730112063368 + } + }, + { + "metric_id": "wake_mean_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "wake_mean_spo2", + "weight": 33, + "params": { + "gt_value": 94.22105120816956, + "tolerance": 4.711052560408478 + } + }, + { + "metric_id": "sleep_min_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "sleep_min_spo2", + "weight": 34, + "params": { + "gt_value": 83.20339248179185, + "tolerance": 4.160169624089593 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case41/case41_07.json b/sleep_assessment/cases/case41/case41_07.json new file mode 100644 index 0000000000000000000000000000000000000000..e2ee691e485388c1a5736b712abbfaeb66f5de0c --- /dev/null +++ b/sleep_assessment/cases/case41/case41_07.json @@ -0,0 +1,48 @@ +{ + "meta_info": { + "case_id": "case41_07", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "label_path": "data/sleep/SHHS1_03.npy", + "instruction": "Use the sleep file and paired labels to calculate sleep mean SpO2, wake mean SpO2, and sleep minimum SpO2. Convert SaO2 to 1 Hz by per-second median; accept 50-100%, remove isolated <=2-second outliers more than 4 points from a centered 9-second median, interpolate only bounded internal gaps <=30 seconds, apply a centered 9-second median, and exclude all remaining missing seconds. Treat N1/N2/N3/R as sleep and W as wake. Report percentages. SHHS1: SaO2=oxygen saturation; THOR RES/ABDO RES=effort; NEW AIR/AIRFLOW=airflow." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract sleep mean SpO2, wake mean SpO2, and sleep minimum SpO2.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. Keys must be EXACTLY \"sleep_mean_spo2\", \"wake_mean_spo2\", and \"sleep_min_spo2\".\n4. Values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"sleep_mean_spo2\": , \"wake_mean_spo2\": , \"sleep_min_spo2\": }", + "metrics": [ + { + "metric_id": "sleep_mean_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "sleep_mean_spo2", + "weight": 33, + "params": { + "gt_value": 95.62549723032568, + "tolerance": 4.781274861516284 + } + }, + { + "metric_id": "wake_mean_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "wake_mean_spo2", + "weight": 33, + "params": { + "gt_value": 96.08619241225712, + "tolerance": 4.804309620612856 + } + }, + { + "metric_id": "sleep_min_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "sleep_min_spo2", + "weight": 34, + "params": { + "gt_value": 83.95213313180965, + "tolerance": 4.197606656590483 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case41/case41_08.json b/sleep_assessment/cases/case41/case41_08.json new file mode 100644 index 0000000000000000000000000000000000000000..96d7198f24f4f6c0cd61ef1c520242bb9a44f16b --- /dev/null +++ b/sleep_assessment/cases/case41/case41_08.json @@ -0,0 +1,48 @@ +{ + "meta_info": { + "case_id": "case41_08", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "label_path": "data/sleep/SHHS1_04.npy", + "instruction": "Use the sleep file and paired labels to calculate sleep mean SpO2, wake mean SpO2, and sleep minimum SpO2. Convert SaO2 to 1 Hz by per-second median; accept 50-100%, remove isolated <=2-second outliers more than 4 points from a centered 9-second median, interpolate only bounded internal gaps <=30 seconds, apply a centered 9-second median, and exclude all remaining missing seconds. Treat N1/N2/N3/R as sleep and W as wake. Report percentages. SHHS1: SaO2=oxygen saturation; THOR RES/ABDO RES=effort; NEW AIR/AIRFLOW=airflow." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract sleep mean SpO2, wake mean SpO2, and sleep minimum SpO2.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. Keys must be EXACTLY \"sleep_mean_spo2\", \"wake_mean_spo2\", and \"sleep_min_spo2\".\n4. Values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"sleep_mean_spo2\": , \"wake_mean_spo2\": , \"sleep_min_spo2\": }", + "metrics": [ + { + "metric_id": "sleep_mean_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "sleep_mean_spo2", + "weight": 33, + "params": { + "gt_value": 96.11838338081978, + "tolerance": 4.8059191690409895 + } + }, + { + "metric_id": "wake_mean_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "wake_mean_spo2", + "weight": 33, + "params": { + "gt_value": 96.28846366637599, + "tolerance": 4.8144231833188 + } + }, + { + "metric_id": "sleep_min_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "sleep_min_spo2", + "weight": 34, + "params": { + "gt_value": 92.28865598094711, + "tolerance": 4.614432799047356 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case41/case41_09.json b/sleep_assessment/cases/case41/case41_09.json new file mode 100644 index 0000000000000000000000000000000000000000..ac5fce2c7485c99797f9499b7b6e73e14521528e --- /dev/null +++ b/sleep_assessment/cases/case41/case41_09.json @@ -0,0 +1,48 @@ +{ + "meta_info": { + "case_id": "case41_09", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "label_path": "data/sleep/SHHS1_05.npy", + "instruction": "Use the sleep file and paired labels to calculate sleep mean SpO2, wake mean SpO2, and sleep minimum SpO2. Convert SaO2 to 1 Hz by per-second median; accept 50-100%, remove isolated <=2-second outliers more than 4 points from a centered 9-second median, interpolate only bounded internal gaps <=30 seconds, apply a centered 9-second median, and exclude all remaining missing seconds. Treat N1/N2/N3/R as sleep and W as wake. Report percentages. SHHS1: SaO2=oxygen saturation; THOR RES/ABDO RES=effort; NEW AIR/AIRFLOW=airflow." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract sleep mean SpO2, wake mean SpO2, and sleep minimum SpO2.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. Keys must be EXACTLY \"sleep_mean_spo2\", \"wake_mean_spo2\", and \"sleep_min_spo2\".\n4. Values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"sleep_mean_spo2\": , \"wake_mean_spo2\": , \"sleep_min_spo2\": }", + "metrics": [ + { + "metric_id": "sleep_mean_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "sleep_mean_spo2", + "weight": 33, + "params": { + "gt_value": 94.84243082865771, + "tolerance": 4.742121541432886 + } + }, + { + "metric_id": "wake_mean_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "wake_mean_spo2", + "weight": 33, + "params": { + "gt_value": 95.96471019336792, + "tolerance": 4.798235509668396 + } + }, + { + "metric_id": "sleep_min_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "sleep_min_spo2", + "weight": 34, + "params": { + "gt_value": 84.70099572079334, + "tolerance": 4.235049786039667 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case41/case41_10.json b/sleep_assessment/cases/case41/case41_10.json new file mode 100644 index 0000000000000000000000000000000000000000..d8e53c9cb64edb5f6f050d75c7b10974d8b79a2d --- /dev/null +++ b/sleep_assessment/cases/case41/case41_10.json @@ -0,0 +1,48 @@ +{ + "meta_info": { + "case_id": "case41_10", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "label_path": "data/sleep/Physionet2018_01.npy", + "instruction": "Use the sleep file and paired labels to calculate sleep mean SpO2, wake mean SpO2, and sleep minimum SpO2. Convert SaO2 to 1 Hz by per-second median; accept 50-100%, remove isolated <=2-second outliers more than 4 points from a centered 9-second median, interpolate only bounded internal gaps <=30 seconds, apply a centered 9-second median, and exclude all remaining missing seconds. Treat N1/N2/N3/R as sleep and W as wake. Report percentages. Physionet2018: SaO2=oxygen saturation; ABD/CHEST=effort; AIRFLOW=airflow." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract sleep mean SpO2, wake mean SpO2, and sleep minimum SpO2.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. Keys must be EXACTLY \"sleep_mean_spo2\", \"wake_mean_spo2\", and \"sleep_min_spo2\".\n4. Values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"sleep_mean_spo2\": , \"wake_mean_spo2\": , \"sleep_min_spo2\": }", + "metrics": [ + { + "metric_id": "sleep_mean_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "sleep_mean_spo2", + "weight": 33, + "params": { + "gt_value": 95.22697915185633, + "tolerance": 4.7613489575928165 + } + }, + { + "metric_id": "wake_mean_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "wake_mean_spo2", + "weight": 33, + "params": { + "gt_value": 96.37645633826935, + "tolerance": 4.8188228169134675 + } + }, + { + "metric_id": "sleep_min_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "sleep_min_spo2", + "weight": 34, + "params": { + "gt_value": 89.6039031990112, + "tolerance": 4.4801951599505605 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case41/case41_11.json b/sleep_assessment/cases/case41/case41_11.json new file mode 100644 index 0000000000000000000000000000000000000000..5e45617f60372ab77b92ce118d4cda69b57dd7e5 --- /dev/null +++ b/sleep_assessment/cases/case41/case41_11.json @@ -0,0 +1,48 @@ +{ + "meta_info": { + "case_id": "case41_11", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "label_path": "data/sleep/Physionet2018_02.npy", + "instruction": "Use the sleep file and paired labels to calculate sleep mean SpO2, wake mean SpO2, and sleep minimum SpO2. Convert SaO2 to 1 Hz by per-second median; accept 50-100%, remove isolated <=2-second outliers more than 4 points from a centered 9-second median, interpolate only bounded internal gaps <=30 seconds, apply a centered 9-second median, and exclude all remaining missing seconds. Treat N1/N2/N3/R as sleep and W as wake. Report percentages. Physionet2018: SaO2=oxygen saturation; ABD/CHEST=effort; AIRFLOW=airflow." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract sleep mean SpO2, wake mean SpO2, and sleep minimum SpO2.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. Keys must be EXACTLY \"sleep_mean_spo2\", \"wake_mean_spo2\", and \"sleep_min_spo2\".\n4. Values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"sleep_mean_spo2\": , \"wake_mean_spo2\": , \"sleep_min_spo2\": }", + "metrics": [ + { + "metric_id": "sleep_mean_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "sleep_mean_spo2", + "weight": 33, + "params": { + "gt_value": 90.44955375402417, + "tolerance": 4.522477687701208 + } + }, + { + "metric_id": "wake_mean_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "wake_mean_spo2", + "weight": 33, + "params": { + "gt_value": 92.04055215891857, + "tolerance": 4.602027607945929 + } + }, + { + "metric_id": "sleep_min_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "sleep_min_spo2", + "weight": 34, + "params": { + "gt_value": 71.7834291950438, + "tolerance": 3.58917145975219 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case41/case41_12.json b/sleep_assessment/cases/case41/case41_12.json new file mode 100644 index 0000000000000000000000000000000000000000..2401ba8babd4ca3e1a5817debc71e560e26093e6 --- /dev/null +++ b/sleep_assessment/cases/case41/case41_12.json @@ -0,0 +1,48 @@ +{ + "meta_info": { + "case_id": "case41_12", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_03.edf", + "label_path": "data/sleep/Physionet2018_03.npy", + "instruction": "Use the sleep file and paired labels to calculate sleep mean SpO2, wake mean SpO2, and sleep minimum SpO2. Convert SaO2 to 1 Hz by per-second median; accept 50-100%, remove isolated <=2-second outliers more than 4 points from a centered 9-second median, interpolate only bounded internal gaps <=30 seconds, apply a centered 9-second median, and exclude all remaining missing seconds. Treat N1/N2/N3/R as sleep and W as wake. Report percentages. Physionet2018: SaO2=oxygen saturation; ABD/CHEST=effort; AIRFLOW=airflow." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract sleep mean SpO2, wake mean SpO2, and sleep minimum SpO2.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. Keys must be EXACTLY \"sleep_mean_spo2\", \"wake_mean_spo2\", and \"sleep_min_spo2\".\n4. Values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"sleep_mean_spo2\": , \"wake_mean_spo2\": , \"sleep_min_spo2\": }", + "metrics": [ + { + "metric_id": "sleep_mean_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "sleep_mean_spo2", + "weight": 33, + "params": { + "gt_value": 94.27980709415505, + "tolerance": 4.713990354707753 + } + }, + { + "metric_id": "wake_mean_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "wake_mean_spo2", + "weight": 33, + "params": { + "gt_value": 95.4768836436887, + "tolerance": 4.773844182184435 + } + }, + { + "metric_id": "sleep_min_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "sleep_min_spo2", + "weight": 34, + "params": { + "gt_value": 91.58421359923094, + "tolerance": 4.579210679961547 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case41/case41_13.json b/sleep_assessment/cases/case41/case41_13.json new file mode 100644 index 0000000000000000000000000000000000000000..2a294a88ce6428fb71dc1b509b24342fec963f9e --- /dev/null +++ b/sleep_assessment/cases/case41/case41_13.json @@ -0,0 +1,48 @@ +{ + "meta_info": { + "case_id": "case41_13", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "label_path": "data/sleep/Physionet2018_04.npy", + "instruction": "Use the sleep file and paired labels to calculate sleep mean SpO2, wake mean SpO2, and sleep minimum SpO2. Convert SaO2 to 1 Hz by per-second median; accept 50-100%, remove isolated <=2-second outliers more than 4 points from a centered 9-second median, interpolate only bounded internal gaps <=30 seconds, apply a centered 9-second median, and exclude all remaining missing seconds. Treat N1/N2/N3/R as sleep and W as wake. Report percentages. Physionet2018: SaO2=oxygen saturation; ABD/CHEST=effort; AIRFLOW=airflow." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract sleep mean SpO2, wake mean SpO2, and sleep minimum SpO2.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. Keys must be EXACTLY \"sleep_mean_spo2\", \"wake_mean_spo2\", and \"sleep_min_spo2\".\n4. Values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"sleep_mean_spo2\": , \"wake_mean_spo2\": , \"sleep_min_spo2\": }", + "metrics": [ + { + "metric_id": "sleep_mean_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "sleep_mean_spo2", + "weight": 33, + "params": { + "gt_value": 94.13502105325266, + "tolerance": 4.706751052662633 + } + }, + { + "metric_id": "wake_mean_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "wake_mean_spo2", + "weight": 33, + "params": { + "gt_value": 94.45512625125403, + "tolerance": 4.7227563125627015 + } + }, + { + "metric_id": "sleep_min_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "sleep_min_spo2", + "weight": 34, + "params": { + "gt_value": 77.723536500412, + "tolerance": 3.8861768250206 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case41/case41_14.json b/sleep_assessment/cases/case41/case41_14.json new file mode 100644 index 0000000000000000000000000000000000000000..60649c1f390bf21d1265387927e69db96c0d6320 --- /dev/null +++ b/sleep_assessment/cases/case41/case41_14.json @@ -0,0 +1,48 @@ +{ + "meta_info": { + "case_id": "case41_14", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "label_path": "data/sleep/Physionet2018_05.npy", + "instruction": "Use the sleep file and paired labels to calculate sleep mean SpO2, wake mean SpO2, and sleep minimum SpO2. Convert SaO2 to 1 Hz by per-second median; accept 50-100%, remove isolated <=2-second outliers more than 4 points from a centered 9-second median, interpolate only bounded internal gaps <=30 seconds, apply a centered 9-second median, and exclude all remaining missing seconds. Treat N1/N2/N3/R as sleep and W as wake. Report percentages. Physionet2018: SaO2=oxygen saturation; ABD/CHEST=effort; AIRFLOW=airflow." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract sleep mean SpO2, wake mean SpO2, and sleep minimum SpO2.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. Keys must be EXACTLY \"sleep_mean_spo2\", \"wake_mean_spo2\", and \"sleep_min_spo2\".\n4. Values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"sleep_mean_spo2\": , \"wake_mean_spo2\": , \"sleep_min_spo2\": }", + "metrics": [ + { + "metric_id": "sleep_mean_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "sleep_mean_spo2", + "weight": 33, + "params": { + "gt_value": 91.31518765601054, + "tolerance": 4.565759382800527 + } + }, + { + "metric_id": "wake_mean_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "wake_mean_spo2", + "weight": 33, + "params": { + "gt_value": 95.84114814688597, + "tolerance": 4.7920574073442985 + } + }, + { + "metric_id": "sleep_min_oxygen_saturation_accuracy", + "type": "numeric_check", + "target_key": "sleep_min_spo2", + "weight": 34, + "params": { + "gt_value": 65.84181620123906, + "tolerance": 3.292090810061953 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case42/case42_01.json b/sleep_assessment/cases/case42/case42_01.json new file mode 100644 index 0000000000000000000000000000000000000000..4cb9e18b161d2c4e17076a8e81d81578584aac8f --- /dev/null +++ b/sleep_assessment/cases/case42/case42_01.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "case42_01", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_01.edf", + "label_path": "data/sleep/ISRUC_01.npy", + "instruction": "Use the sleep file and paired labels to calculate sleep-period T90 and T80 in cumulative minutes, not percentages. Convert SaO2 to 1 Hz by per-second median; accept 50-100%, remove isolated <=2-second outliers more than 4 points from a centered 9-second median, interpolate only bounded internal gaps <=30 seconds, apply a centered 9-second median, and exclude all remaining missing seconds without extrapolation. T90 counts valid N1/N2/N3/R seconds with SaO2 strictly below 90%; T80 uses strictly below 80%. ISRUC: SaO2=oxygen saturation; X5=snore; X6/DC3=airflow; X7/X8=abdominal effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract T90 and T80 hypoxemia burden in minutes.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. Keys must be EXACTLY \"t90_min\" and \"t80_min\".\n4. Values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"t90_min\": , \"t80_min\": }", + "metrics": [ + { + "metric_id": "t90_duration_burden_accuracy", + "type": "numeric_check", + "target_key": "t90_min", + "weight": 50, + "params": { + "gt_value": 18.183333333333334, + "tolerance": 0.9091666666666667 + } + }, + { + "metric_id": "t80_duration_burden_accuracy", + "type": "numeric_check", + "target_key": "t80_min", + "weight": 50, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case42/case42_02.json b/sleep_assessment/cases/case42/case42_02.json new file mode 100644 index 0000000000000000000000000000000000000000..dfbe9e3b7b262593d879759f70031b3da994b000 --- /dev/null +++ b/sleep_assessment/cases/case42/case42_02.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "case42_02", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_02.edf", + "label_path": "data/sleep/ISRUC_02.npy", + "instruction": "Use the sleep file and paired labels to calculate sleep-period T90 and T80 in cumulative minutes, not percentages. Convert SaO2 to 1 Hz by per-second median; accept 50-100%, remove isolated <=2-second outliers more than 4 points from a centered 9-second median, interpolate only bounded internal gaps <=30 seconds, apply a centered 9-second median, and exclude all remaining missing seconds without extrapolation. T90 counts valid N1/N2/N3/R seconds with SaO2 strictly below 90%; T80 uses strictly below 80%. ISRUC: SaO2=oxygen saturation; X5=snore; X6/DC3=airflow; X7/X8=abdominal effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract T90 and T80 hypoxemia burden in minutes.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. Keys must be EXACTLY \"t90_min\" and \"t80_min\".\n4. Values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"t90_min\": , \"t80_min\": }", + "metrics": [ + { + "metric_id": "t90_duration_burden_accuracy", + "type": "numeric_check", + "target_key": "t90_min", + "weight": 50, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + }, + { + "metric_id": "t80_duration_burden_accuracy", + "type": "numeric_check", + "target_key": "t80_min", + "weight": 50, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case42/case42_03.json b/sleep_assessment/cases/case42/case42_03.json new file mode 100644 index 0000000000000000000000000000000000000000..d67ae1123e660c44945440d399ceb7f3b5a4c3f5 --- /dev/null +++ b/sleep_assessment/cases/case42/case42_03.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "case42_03", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_03.edf", + "label_path": "data/sleep/ISRUC_03.npy", + "instruction": "Use the sleep file and paired labels to calculate sleep-period T90 and T80 in cumulative minutes, not percentages. Convert SaO2 to 1 Hz by per-second median; accept 50-100%, remove isolated <=2-second outliers more than 4 points from a centered 9-second median, interpolate only bounded internal gaps <=30 seconds, apply a centered 9-second median, and exclude all remaining missing seconds without extrapolation. T90 counts valid N1/N2/N3/R seconds with SaO2 strictly below 90%; T80 uses strictly below 80%. ISRUC: SaO2=oxygen saturation; X5=snore; X6/DC3=airflow; X7/X8=abdominal effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract T90 and T80 hypoxemia burden in minutes.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. Keys must be EXACTLY \"t90_min\" and \"t80_min\".\n4. Values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"t90_min\": , \"t80_min\": }", + "metrics": [ + { + "metric_id": "t90_duration_burden_accuracy", + "type": "numeric_check", + "target_key": "t90_min", + "weight": 50, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + }, + { + "metric_id": "t80_duration_burden_accuracy", + "type": "numeric_check", + "target_key": "t80_min", + "weight": 50, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case42/case42_04.json b/sleep_assessment/cases/case42/case42_04.json new file mode 100644 index 0000000000000000000000000000000000000000..5b9cfb91a366d8aedf0d561802071871f235b38c --- /dev/null +++ b/sleep_assessment/cases/case42/case42_04.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "case42_04", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "ISRUC" + }, + "agent_input": { + "data_path": "data/sleep/ISRUC_05.edf", + "label_path": "data/sleep/ISRUC_05.npy", + "instruction": "Use the sleep file and paired labels to calculate sleep-period T90 and T80 in cumulative minutes, not percentages. Convert SaO2 to 1 Hz by per-second median; accept 50-100%, remove isolated <=2-second outliers more than 4 points from a centered 9-second median, interpolate only bounded internal gaps <=30 seconds, apply a centered 9-second median, and exclude all remaining missing seconds without extrapolation. T90 counts valid N1/N2/N3/R seconds with SaO2 strictly below 90%; T80 uses strictly below 80%. ISRUC: SaO2=oxygen saturation; X5=snore; X6/DC3=airflow; X7/X8=abdominal effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract T90 and T80 hypoxemia burden in minutes.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. Keys must be EXACTLY \"t90_min\" and \"t80_min\".\n4. Values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"t90_min\": , \"t80_min\": }", + "metrics": [ + { + "metric_id": "t90_duration_burden_accuracy", + "type": "numeric_check", + "target_key": "t90_min", + "weight": 50, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + }, + { + "metric_id": "t80_duration_burden_accuracy", + "type": "numeric_check", + "target_key": "t80_min", + "weight": 50, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case42/case42_05.json b/sleep_assessment/cases/case42/case42_05.json new file mode 100644 index 0000000000000000000000000000000000000000..1b4dfaf70f9a5abcf579bc3fd9d0788f4edd772c --- /dev/null +++ b/sleep_assessment/cases/case42/case42_05.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "case42_05", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_01.edf", + "label_path": "data/sleep/SHHS1_01.npy", + "instruction": "Use the sleep file and paired labels to calculate sleep-period T90 and T80 in cumulative minutes, not percentages. Convert SaO2 to 1 Hz by per-second median; accept 50-100%, remove isolated <=2-second outliers more than 4 points from a centered 9-second median, interpolate only bounded internal gaps <=30 seconds, apply a centered 9-second median, and exclude all remaining missing seconds without extrapolation. T90 counts valid N1/N2/N3/R seconds with SaO2 strictly below 90%; T80 uses strictly below 80%. SHHS1: SaO2=oxygen saturation; THOR RES/ABDO RES=effort; NEW AIR/AIRFLOW=airflow." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract T90 and T80 hypoxemia burden in minutes.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. Keys must be EXACTLY \"t90_min\" and \"t80_min\".\n4. Values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"t90_min\": , \"t80_min\": }", + "metrics": [ + { + "metric_id": "t90_duration_burden_accuracy", + "type": "numeric_check", + "target_key": "t90_min", + "weight": 50, + "params": { + "gt_value": 7.866666666666666, + "tolerance": 0.3933333333333333 + } + }, + { + "metric_id": "t80_duration_burden_accuracy", + "type": "numeric_check", + "target_key": "t80_min", + "weight": 50, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case42/case42_06.json b/sleep_assessment/cases/case42/case42_06.json new file mode 100644 index 0000000000000000000000000000000000000000..c43822a00eca2fe69cf8e9b6b7793bf49c190f28 --- /dev/null +++ b/sleep_assessment/cases/case42/case42_06.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "case42_06", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_02.edf", + "label_path": "data/sleep/SHHS1_02.npy", + "instruction": "Use the sleep file and paired labels to calculate sleep-period T90 and T80 in cumulative minutes, not percentages. Convert SaO2 to 1 Hz by per-second median; accept 50-100%, remove isolated <=2-second outliers more than 4 points from a centered 9-second median, interpolate only bounded internal gaps <=30 seconds, apply a centered 9-second median, and exclude all remaining missing seconds without extrapolation. T90 counts valid N1/N2/N3/R seconds with SaO2 strictly below 90%; T80 uses strictly below 80%. SHHS1: SaO2=oxygen saturation; THOR RES/ABDO RES=effort; NEW AIR/AIRFLOW=airflow." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract T90 and T80 hypoxemia burden in minutes.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. Keys must be EXACTLY \"t90_min\" and \"t80_min\".\n4. Values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"t90_min\": , \"t80_min\": }", + "metrics": [ + { + "metric_id": "t90_duration_burden_accuracy", + "type": "numeric_check", + "target_key": "t90_min", + "weight": 50, + "params": { + "gt_value": 2.9833333333333334, + "tolerance": 0.14916666666666667 + } + }, + { + "metric_id": "t80_duration_burden_accuracy", + "type": "numeric_check", + "target_key": "t80_min", + "weight": 50, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case42/case42_07.json b/sleep_assessment/cases/case42/case42_07.json new file mode 100644 index 0000000000000000000000000000000000000000..aedf75cf44b13421d929ac81447b6bf5259dd311 --- /dev/null +++ b/sleep_assessment/cases/case42/case42_07.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "case42_07", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_03.edf", + "label_path": "data/sleep/SHHS1_03.npy", + "instruction": "Use the sleep file and paired labels to calculate sleep-period T90 and T80 in cumulative minutes, not percentages. Convert SaO2 to 1 Hz by per-second median; accept 50-100%, remove isolated <=2-second outliers more than 4 points from a centered 9-second median, interpolate only bounded internal gaps <=30 seconds, apply a centered 9-second median, and exclude all remaining missing seconds without extrapolation. T90 counts valid N1/N2/N3/R seconds with SaO2 strictly below 90%; T80 uses strictly below 80%. SHHS1: SaO2=oxygen saturation; THOR RES/ABDO RES=effort; NEW AIR/AIRFLOW=airflow." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract T90 and T80 hypoxemia burden in minutes.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. Keys must be EXACTLY \"t90_min\" and \"t80_min\".\n4. Values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"t90_min\": , \"t80_min\": }", + "metrics": [ + { + "metric_id": "t90_duration_burden_accuracy", + "type": "numeric_check", + "target_key": "t90_min", + "weight": 50, + "params": { + "gt_value": 0.9666666666666667, + "tolerance": 0.04833333333333334 + } + }, + { + "metric_id": "t80_duration_burden_accuracy", + "type": "numeric_check", + "target_key": "t80_min", + "weight": 50, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case42/case42_08.json b/sleep_assessment/cases/case42/case42_08.json new file mode 100644 index 0000000000000000000000000000000000000000..6d1d11b0faa26988cfea6c66c015ef867f3e5ac9 --- /dev/null +++ b/sleep_assessment/cases/case42/case42_08.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "case42_08", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_04.edf", + "label_path": "data/sleep/SHHS1_04.npy", + "instruction": "Use the sleep file and paired labels to calculate sleep-period T90 and T80 in cumulative minutes, not percentages. Convert SaO2 to 1 Hz by per-second median; accept 50-100%, remove isolated <=2-second outliers more than 4 points from a centered 9-second median, interpolate only bounded internal gaps <=30 seconds, apply a centered 9-second median, and exclude all remaining missing seconds without extrapolation. T90 counts valid N1/N2/N3/R seconds with SaO2 strictly below 90%; T80 uses strictly below 80%. SHHS1: SaO2=oxygen saturation; THOR RES/ABDO RES=effort; NEW AIR/AIRFLOW=airflow." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract T90 and T80 hypoxemia burden in minutes.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. Keys must be EXACTLY \"t90_min\" and \"t80_min\".\n4. Values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"t90_min\": , \"t80_min\": }", + "metrics": [ + { + "metric_id": "t90_duration_burden_accuracy", + "type": "numeric_check", + "target_key": "t90_min", + "weight": 50, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + }, + { + "metric_id": "t80_duration_burden_accuracy", + "type": "numeric_check", + "target_key": "t80_min", + "weight": 50, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case42/case42_09.json b/sleep_assessment/cases/case42/case42_09.json new file mode 100644 index 0000000000000000000000000000000000000000..a11a02182bdbabe398b0a357f6b5eb5ff5f25bdc --- /dev/null +++ b/sleep_assessment/cases/case42/case42_09.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "case42_09", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "SHHS1" + }, + "agent_input": { + "data_path": "data/sleep/SHHS1_05.edf", + "label_path": "data/sleep/SHHS1_05.npy", + "instruction": "Use the sleep file and paired labels to calculate sleep-period T90 and T80 in cumulative minutes, not percentages. Convert SaO2 to 1 Hz by per-second median; accept 50-100%, remove isolated <=2-second outliers more than 4 points from a centered 9-second median, interpolate only bounded internal gaps <=30 seconds, apply a centered 9-second median, and exclude all remaining missing seconds without extrapolation. T90 counts valid N1/N2/N3/R seconds with SaO2 strictly below 90%; T80 uses strictly below 80%. SHHS1: SaO2=oxygen saturation; THOR RES/ABDO RES=effort; NEW AIR/AIRFLOW=airflow." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract T90 and T80 hypoxemia burden in minutes.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. Keys must be EXACTLY \"t90_min\" and \"t80_min\".\n4. Values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"t90_min\": , \"t80_min\": }", + "metrics": [ + { + "metric_id": "t90_duration_burden_accuracy", + "type": "numeric_check", + "target_key": "t90_min", + "weight": 50, + "params": { + "gt_value": 2.75, + "tolerance": 0.1375 + } + }, + { + "metric_id": "t80_duration_burden_accuracy", + "type": "numeric_check", + "target_key": "t80_min", + "weight": 50, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case42/case42_10.json b/sleep_assessment/cases/case42/case42_10.json new file mode 100644 index 0000000000000000000000000000000000000000..beba1e060ecd7d4d0c424feaa3eb7bf8288cdf59 --- /dev/null +++ b/sleep_assessment/cases/case42/case42_10.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "case42_10", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_01.edf", + "label_path": "data/sleep/Physionet2018_01.npy", + "instruction": "Use the sleep file and paired labels to calculate sleep-period T90 and T80 in cumulative minutes, not percentages. Convert SaO2 to 1 Hz by per-second median; accept 50-100%, remove isolated <=2-second outliers more than 4 points from a centered 9-second median, interpolate only bounded internal gaps <=30 seconds, apply a centered 9-second median, and exclude all remaining missing seconds without extrapolation. T90 counts valid N1/N2/N3/R seconds with SaO2 strictly below 90%; T80 uses strictly below 80%. Physionet2018: SaO2=oxygen saturation; ABD/CHEST=effort; AIRFLOW=airflow." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract T90 and T80 hypoxemia burden in minutes.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. Keys must be EXACTLY \"t90_min\" and \"t80_min\".\n4. Values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"t90_min\": , \"t80_min\": }", + "metrics": [ + { + "metric_id": "t90_duration_burden_accuracy", + "type": "numeric_check", + "target_key": "t90_min", + "weight": 50, + "params": { + "gt_value": 0.15, + "tolerance": 0.0075 + } + }, + { + "metric_id": "t80_duration_burden_accuracy", + "type": "numeric_check", + "target_key": "t80_min", + "weight": 50, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case42/case42_11.json b/sleep_assessment/cases/case42/case42_11.json new file mode 100644 index 0000000000000000000000000000000000000000..5ec4cd94a1e5475e15461a3edbe5ff22e4e803cf --- /dev/null +++ b/sleep_assessment/cases/case42/case42_11.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "case42_11", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_02.edf", + "label_path": "data/sleep/Physionet2018_02.npy", + "instruction": "Use the sleep file and paired labels to calculate sleep-period T90 and T80 in cumulative minutes, not percentages. Convert SaO2 to 1 Hz by per-second median; accept 50-100%, remove isolated <=2-second outliers more than 4 points from a centered 9-second median, interpolate only bounded internal gaps <=30 seconds, apply a centered 9-second median, and exclude all remaining missing seconds without extrapolation. T90 counts valid N1/N2/N3/R seconds with SaO2 strictly below 90%; T80 uses strictly below 80%. Physionet2018: SaO2=oxygen saturation; ABD/CHEST=effort; AIRFLOW=airflow." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract T90 and T80 hypoxemia burden in minutes.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. Keys must be EXACTLY \"t90_min\" and \"t80_min\".\n4. Values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"t90_min\": , \"t80_min\": }", + "metrics": [ + { + "metric_id": "t90_duration_burden_accuracy", + "type": "numeric_check", + "target_key": "t90_min", + "weight": 50, + "params": { + "gt_value": 97.08333333333333, + "tolerance": 4.854166666666667 + } + }, + { + "metric_id": "t80_duration_burden_accuracy", + "type": "numeric_check", + "target_key": "t80_min", + "weight": 50, + "params": { + "gt_value": 2.966666666666667, + "tolerance": 0.14833333333333334 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case42/case42_12.json b/sleep_assessment/cases/case42/case42_12.json new file mode 100644 index 0000000000000000000000000000000000000000..4f4b440423b63e66dc6c579a55fab6cc9706bec8 --- /dev/null +++ b/sleep_assessment/cases/case42/case42_12.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "case42_12", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_03.edf", + "label_path": "data/sleep/Physionet2018_03.npy", + "instruction": "Use the sleep file and paired labels to calculate sleep-period T90 and T80 in cumulative minutes, not percentages. Convert SaO2 to 1 Hz by per-second median; accept 50-100%, remove isolated <=2-second outliers more than 4 points from a centered 9-second median, interpolate only bounded internal gaps <=30 seconds, apply a centered 9-second median, and exclude all remaining missing seconds without extrapolation. T90 counts valid N1/N2/N3/R seconds with SaO2 strictly below 90%; T80 uses strictly below 80%. Physionet2018: SaO2=oxygen saturation; ABD/CHEST=effort; AIRFLOW=airflow." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract T90 and T80 hypoxemia burden in minutes.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. Keys must be EXACTLY \"t90_min\" and \"t80_min\".\n4. Values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"t90_min\": , \"t80_min\": }", + "metrics": [ + { + "metric_id": "t90_duration_burden_accuracy", + "type": "numeric_check", + "target_key": "t90_min", + "weight": 50, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + }, + { + "metric_id": "t80_duration_burden_accuracy", + "type": "numeric_check", + "target_key": "t80_min", + "weight": 50, + "params": { + "gt_value": 0.0, + "tolerance": 0.0 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case42/case42_13.json b/sleep_assessment/cases/case42/case42_13.json new file mode 100644 index 0000000000000000000000000000000000000000..5652cd00caecbddfa71eb244c46af10480965ab9 --- /dev/null +++ b/sleep_assessment/cases/case42/case42_13.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "case42_13", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_04.edf", + "label_path": "data/sleep/Physionet2018_04.npy", + "instruction": "Use the sleep file and paired labels to calculate sleep-period T90 and T80 in cumulative minutes, not percentages. Convert SaO2 to 1 Hz by per-second median; accept 50-100%, remove isolated <=2-second outliers more than 4 points from a centered 9-second median, interpolate only bounded internal gaps <=30 seconds, apply a centered 9-second median, and exclude all remaining missing seconds without extrapolation. T90 counts valid N1/N2/N3/R seconds with SaO2 strictly below 90%; T80 uses strictly below 80%. Physionet2018: SaO2=oxygen saturation; ABD/CHEST=effort; AIRFLOW=airflow." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract T90 and T80 hypoxemia burden in minutes.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. Keys must be EXACTLY \"t90_min\" and \"t80_min\".\n4. Values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"t90_min\": , \"t80_min\": }", + "metrics": [ + { + "metric_id": "t90_duration_burden_accuracy", + "type": "numeric_check", + "target_key": "t90_min", + "weight": 50, + "params": { + "gt_value": 28.233333333333334, + "tolerance": 1.4116666666666668 + } + }, + { + "metric_id": "t80_duration_burden_accuracy", + "type": "numeric_check", + "target_key": "t80_min", + "weight": 50, + "params": { + "gt_value": 0.3, + "tolerance": 0.015 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case42/case42_14.json b/sleep_assessment/cases/case42/case42_14.json new file mode 100644 index 0000000000000000000000000000000000000000..a910f7e9c672688fa5046c3d64a8afd6f82d8907 --- /dev/null +++ b/sleep_assessment/cases/case42/case42_14.json @@ -0,0 +1,38 @@ +{ + "meta_info": { + "case_id": "case42_14", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1.5, + "original_dataset": "Physionet2018" + }, + "agent_input": { + "data_path": "data/sleep/Physionet2018_05.edf", + "label_path": "data/sleep/Physionet2018_05.npy", + "instruction": "Use the sleep file and paired labels to calculate sleep-period T90 and T80 in cumulative minutes, not percentages. Convert SaO2 to 1 Hz by per-second median; accept 50-100%, remove isolated <=2-second outliers more than 4 points from a centered 9-second median, interpolate only bounded internal gaps <=30 seconds, apply a centered 9-second median, and exclude all remaining missing seconds without extrapolation. T90 counts valid N1/N2/N3/R seconds with SaO2 strictly below 90%; T80 uses strictly below 80%. Physionet2018: SaO2=oxygen saturation; ABD/CHEST=effort; AIRFLOW=airflow." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports.\n\n### TASK\nExtract T90 and T80 hypoxemia burden in minutes.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown, explanations, units, or extra fields.\n3. Keys must be EXACTLY \"t90_min\" and \"t80_min\".\n4. Values must be FLOAT values or null.\n5. Return null for a missing or unclear value.\n\n### OUTPUT TEMPLATE\n{\"t90_min\": , \"t80_min\": }", + "metrics": [ + { + "metric_id": "t90_duration_burden_accuracy", + "type": "numeric_check", + "target_key": "t90_min", + "weight": 50, + "params": { + "gt_value": 142.18333333333334, + "tolerance": 7.109166666666667 + } + }, + { + "metric_id": "t80_duration_burden_accuracy", + "type": "numeric_check", + "target_key": "t80_min", + "weight": 50, + "params": { + "gt_value": 13.033333333333333, + "tolerance": 0.6516666666666667 + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_01.json b/sleep_assessment/cases/case43/case43_01.json new file mode 100644 index 0000000000000000000000000000000000000000..b7d83fa69ef5150d34aa30e7b24f07cb0230d39c --- /dev/null +++ b/sleep_assessment/cases/case43/case43_01.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q01", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: In a standard adult PSG, which EEG rhythm is most typical over the occipital region during relaxed wakefulness with eyes closed?\nA. 0.5-2 Hz slow waves\nB. 8-13 Hz alpha rhythm\nC. 4-7 Hz theta rhythm\nD. 11-16 Hz sleep spindles" + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "B" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_02.json b/sleep_assessment/cases/case43/case43_02.json new file mode 100644 index 0000000000000000000000000000000000000000..5b48edb26fc55f164ab9a9e5b215f96d4a0dced5 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_02.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q02", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which statement about scoring stage N1 is correct?\nA. Sleep spindles must be present before N1 can be scored.\nB. Slow eye movements occur only in REM sleep.\nC. Chin EMG activity is usually higher than during wakefulness.\nD. Vertex sharp waves may occur, but they are not required for scoring N1." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "D" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_03.json b/sleep_assessment/cases/case43/case43_03.json new file mode 100644 index 0000000000000000000000000000000000000000..e4bcc9b605325e8ee96ab139e2a70220405ae5b4 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_03.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q03", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which description best characterizes vertex sharp waves?\nA. They are a type of triphasic wave typically seen in hepatic encephalopathy.\nB. They usually last at least 1 second and are maximal frontally.\nC. They last less than 0.5 seconds and are most prominent over the central region.\nD. They have a frequency of 11-16 Hz and are maximal over the central region." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "C" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_04.json b/sleep_assessment/cases/case43/case43_04.json new file mode 100644 index 0000000000000000000000000000000000000000..c07ef128af44507e2d624109a2e07603777f6923 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_04.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q04", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which feature is essential to the AASM definition of a K-complex?\nA. A total duration of at least 0.5 seconds, with a negative sharp wave immediately followed by a positive component.\nB. Its frequency must be 12-14 Hz.\nC. It can count toward N2 only when accompanied by an arousal.\nD. It occurs only during REM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "A" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_05.json b/sleep_assessment/cases/case43/case43_05.json new file mode 100644 index 0000000000000000000000000000000000000000..829bd689b9856d7530b039acaf874e7f900ae904 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_05.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q05", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: When the criteria for N3 are not met, which rule best describes the onset of N2?\nA. N2 begins whenever chin EMG activity decreases.\nB. N2 begins whenever any 4-7 Hz theta activity appears.\nC. N2 begins when any eye movement occurs in the preceding 10 seconds of the epoch.\nD. N2 begins when a K-complex not associated with an arousal or a sleep spindle occurs in the first half of the current epoch or the second half of the preceding epoch." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "D" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_06.json b/sleep_assessment/cases/case43/case43_06.json new file mode 100644 index 0000000000000000000000000000000000000000..5e8880d892e4ae83d98c77fca944096bc13ae8cb --- /dev/null +++ b/sleep_assessment/cases/case43/case43_06.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q06", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which statement about sleep spindles is correct?\nA. They are 11-16 Hz, most commonly 12-14 Hz, and last at least 0.5 seconds.\nB. They are 2-6 Hz and last at least 0.5 seconds.\nC. They must last at least 3 seconds to be scored.\nD. They occur only in REM sleep and are most prominent over the occipital region." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "A" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_07.json b/sleep_assessment/cases/case43/case43_07.json new file mode 100644 index 0000000000000000000000000000000000000000..d460c05bc456c47c63a8cde89418f78f2555d759 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_07.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q07", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which criterion most appropriately defines stage N3?\nA. Any occurrence of delta activity is sufficient for N3.\nB. In a 30-second epoch, frontal slow-wave activity occupies at least 20%, with a frequency of 0.5-2 Hz and a peak-to-peak amplitude greater than 75 microvolts.\nC. Any K-complex is sufficient for N3.\nD. A reduction in eye movements alone is sufficient for N3." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "B" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_08.json b/sleep_assessment/cases/case43/case43_08.json new file mode 100644 index 0000000000000000000000000000000000000000..c99f52bd70a8a335512b0d34ac0acfbf462b8b37 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_08.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q08", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which combination best describes definite REM sleep according to the AASM?\nA. High-amplitude delta activity, high chin muscle tone, and no eye movements.\nB. Sustained alpha rhythm, high chin muscle tone, and a flat EOG.\nC. Low-amplitude mixed-frequency EEG, low chin muscle tone, and rapid eye movements.\nD. Frequent spindles, increased muscle tone, and slow eye movements." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "C" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_09.json b/sleep_assessment/cases/case43/case43_09.json new file mode 100644 index 0000000000000000000000000000000000000000..6672803d3bf1ea3fab6d3205fb5b72d9780e11e4 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_09.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q09", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: After definite stage R has begun, how should a subsequent epoch be scored if it contains no new rapid eye movements but still shows low-amplitude mixed-frequency EEG and low chin muscle tone, contains no K-complexes or spindles, and has no arousal or other stage-R termination marker?\nA. It must be rescored as wakefulness.\nB. It must be rescored as N1.\nC. It must be rescored as N2.\nD. It may continue to be scored as stage R." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "D" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_10.json b/sleep_assessment/cases/case43/case43_10.json new file mode 100644 index 0000000000000000000000000000000000000000..e0dcc68b2c7cc28d0df8825d062813dceb8e242a --- /dev/null +++ b/sleep_assessment/cases/case43/case43_10.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q10", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which statement about sawtooth waves is correct?\nA. They are typically 11-16 Hz and are required for N2.\nB. They are typically 2-6 Hz, are most prominent centrally, and support REM scoring but are not required for REM.\nC. Their presence is sufficient to diagnose REM sleep behavior disorder.\nD. They occur only in N3 and are usually accompanied by high-amplitude slow waves." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "B" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_11.json b/sleep_assessment/cases/case43/case43_11.json new file mode 100644 index 0000000000000000000000000000000000000000..933b106e9b796fc98f1b6d742868a834d27e3231 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_11.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q11", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which option best matches the AASM arousal scoring rule?\nA. An increase in heart rate alone is sufficient to score an arousal.\nB. Any EEG frequency shift lasting at least 1 second is sufficient.\nC. An abrupt shift to alpha, theta, or activity above 16 Hz other than spindles lasts at least 3 seconds after at least 10 seconds of stable sleep; in stage R, at least 1 second of increased chin EMG is also required.\nD. Any sleep spindle is followed by an arousal." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "C" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_12.json b/sleep_assessment/cases/case43/case43_12.json new file mode 100644 index 0000000000000000000000000000000000000000..e32aec6ea7b520801c8ada43f2cc5a4542d8d905 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_12.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q12", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: What is the core adult PSG scoring criterion for apnea?\nA. Airflow decreases by at least 50% for at least 5 seconds with desaturation.\nB. Airflow decreases by at least 30% for at least 10 seconds with an arousal.\nC. The respiratory sensor signal decreases by at least 90% from baseline for at least 90% of the event duration, and the event lasts at least 10 seconds.\nD. Any oxygen desaturation of at least 3% defines an apnea." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "C" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_13.json b/sleep_assessment/cases/case43/case43_13.json new file mode 100644 index 0000000000000000000000000000000000000000..0477ea2c9e13dc617f9ea13b61f5310354058158 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_13.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q13", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which statement correctly distinguishes obstructive, central, and mixed apnea?\nA. Central apnea has persistent or increasing respiratory effort throughout the event.\nB. The three types are distinguished only by the magnitude of SpO2 reduction.\nC. Obstructive apnea has no respiratory effort throughout the event.\nD. Mixed apnea begins without respiratory effort and is followed by resumption of respiratory effort." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "D" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_14.json b/sleep_assessment/cases/case43/case43_14.json new file mode 100644 index 0000000000000000000000000000000000000000..6348b1c54b5691a2735026b2ad29fe7db5cd828a --- /dev/null +++ b/sleep_assessment/cases/case43/case43_14.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q14", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which option gives the recommended adult hypopnea scoring rule?\nA. Airflow decreases by at least 30% for at least 10 seconds and is accompanied by at least 3% desaturation or an arousal.\nB. Thoracoabdominal paradox and snoring must both be present.\nC. Hypopnea can be scored only when SpO2 decreases by at least 4%; an arousal does not qualify.\nD. Airflow decreases by at least 90% for at least 10 seconds." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "A" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_15.json b/sleep_assessment/cases/case43/case43_15.json new file mode 100644 index 0000000000000000000000000000000000000000..82c68cf04c99a09252547eb7b1bd615f358787d8 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_15.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q15", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which option best defines a respiratory effort-related arousal (RERA)?\nA. Every snoring event is a RERA.\nB. Every central apnea is a RERA.\nC. Every respiratory event with at least 3% SpO2 reduction is a RERA.\nD. A sequence lasting at least 10 seconds with increasing respiratory effort or flattening of the inspiratory nasal-pressure waveform that leads to an arousal but does not meet apnea or hypopnea criteria." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "D" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_16.json b/sleep_assessment/cases/case43/case43_16.json new file mode 100644 index 0000000000000000000000000000000000000000..5b9453c21b1d436a27a9f660f0bfa740b7f86096 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_16.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q16", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which statement about scoring sleep respiratory events during supplemental oxygen is correct?\nA. Supplemental oxygen may blunt desaturation; if a diagnostic night is performed with oxygen, this should be noted in the report summary.\nB. All hypopneas should be rescored as apneas whenever oxygen is used.\nC. Supplemental oxygen makes desaturation more pronounced and therefore makes hypopnea easier to score.\nD. The AASM requires subtracting 3% from every SpO2 value before scoring patients receiving oxygen." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "A" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_17.json b/sleep_assessment/cases/case43/case43_17.json new file mode 100644 index 0000000000000000000000000000000000000000..e7daedc418d422d7d207a4304fc39205ba398c15 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_17.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q17", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which adult OSA severity classification by AHI is correct?\nA. Mild: 5-14.9/h; moderate: 15-29.9/h; severe: at least 30/h.\nB. Mild: 0-5/h; moderate: 6-20/h; severe: greater than 20/h.\nC. Mild: 5-9.9/h; moderate: 10-19.9/h; severe: at least 20/h.\nD. Mild: 1-4.9/h; moderate: 5-14.9/h; severe: at least 15/h." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "A" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_18.json b/sleep_assessment/cases/case43/case43_18.json new file mode 100644 index 0000000000000000000000000000000000000000..e78d4fa2d7d02b634c2ca31e57774da33b41f4ae --- /dev/null +++ b/sleep_assessment/cases/case43/case43_18.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q18", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which statement about body position and OSA is best supported by the literature?\nA. Adult OSA is usually worse in the lateral position and milder when supine.\nB. Position affects OSA only in children, not in adults.\nC. Adult OSA is often worse when supine and milder in the lateral position.\nD. Position affects snoring but not OSA severity." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "C" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_19.json b/sleep_assessment/cases/case43/case43_19.json new file mode 100644 index 0000000000000000000000000000000000000000..5f62091e427fa079559a27d1a22c388e58dd3523 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_19.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q19", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: According to the 2024 Australasian Sleep Association guideline, which set best represents the minimum channels for an adult Type 1 laboratory diagnostic PSG?\nA. Video alone is sufficient, so EEG, EOG, and EMG are unnecessary.\nB. Nasal airflow, oximetry, and a thoracic belt alone are equivalent to PSG.\nC. At least 3 EEG channels, 2 EOG channels, 1 chin EMG channel, 2 leg EMG channels, ECG, oximetry, airflow, thoracic and abdominal movement, body position, and snoring.\nD. One EEG channel, one EOG channel, no leg EMG, and no ECG." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "C" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_20.json b/sleep_assessment/cases/case43/case43_20.json new file mode 100644 index 0000000000000000000000000000000000000000..53258416ef5be1a181d21a42f9092db3c036f8e6 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_20.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q20", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which statement about digital sampling of sleep signals is consistent with the Nyquist theorem?\nA. The sampling rate only needs to equal the highest frequency of interest.\nB. The sampling rate is determined mainly by screen resolution and is unrelated to signal frequency.\nC. A lower sampling rate always reduces distortion.\nD. The sampling rate should be greater than twice the highest frequency of interest." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "D" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_21.json b/sleep_assessment/cases/case43/case43_21.json new file mode 100644 index 0000000000000000000000000000000000000000..156df2baf962207f070ec5f183792e49aedc61f6 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_21.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q21", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: According to the AAST PSG technical guideline, which sampling-rate setting is most appropriate for EEG, EOG, EMG, ECG, and snoring channels?\nA. Minimum 200 Hz; ideal or recommended 500 Hz.\nB. Exactly 1000 Hz is mandatory.\nC. Minimum 25 Hz; recommended 50 Hz.\nD. Minimum 100 Hz; recommended 128 Hz." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "A" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_22.json b/sleep_assessment/cases/case43/case43_22.json new file mode 100644 index 0000000000000000000000000000000000000000..d32b91e7e777a1db00113f3eb3678680f42de896 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_22.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q22", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: What is the most likely consequence of incorrectly lowering the EEG high-frequency filter cutoff from 35 Hz to 15 Hz?\nA. Sleep-spindle detail is preserved more effectively.\nB. EEG components near 15 Hz, including sleep spindles, are markedly attenuated and distorted.\nC. Only SpO2 is affected; EEG is unchanged.\nD. Electrode impedance improves without any spectral effect." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "B" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_23.json b/sleep_assessment/cases/case43/case43_23.json new file mode 100644 index 0000000000000000000000000000000000000000..1a3353acf81268a4fe9874d5d4f4f5df32f22ac1 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_23.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q23", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which statement best matches technical guidance for electrode impedance in adult PSG?\nA. The upper limit is 5 kOhm for EEG/EOG and 10 kOhm for EMG, and paired electrode impedances should be matched as closely as possible.\nB. EEG may be as high as 50 kOhm, while EMG must be below 1 kOhm.\nC. All channels must have impedances of at least 20 kOhm to remain stable.\nD. Impedance affects only ECG and does not affect EEG, EOG, or EMG." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "A" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_24.json b/sleep_assessment/cases/case43/case43_24.json new file mode 100644 index 0000000000000000000000000000000000000000..7f47dbcb5cb4a9c3aebd7557bb96f2daa3931bad --- /dev/null +++ b/sleep_assessment/cases/case43/case43_24.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q24", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: If a PSG or sleep-EEG channel suddenly shows a nonphysiologic step-like deflection and a loose electrode is suspected, what should be considered first?\nA. A typical K-complex.\nB. A lead-pop or loose-electrode artifact.\nC. A sawtooth wave.\nD. New cortical slow-wave activity." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "B" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_25.json b/sleep_assessment/cases/case43/case43_25.json new file mode 100644 index 0000000000000000000000000000000000000000..d7ef52dee4804401383e0a8c32bc66e0f2a043f5 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_25.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q25", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which option correctly defines a significant leg movement (LM) event according to the AASM?\nA. It lasts 0.5-10 seconds and the EMG amplitude rises at least 8 microvolts above resting baseline.\nB. Any movement associated with an arousal is an LM regardless of duration.\nC. An amplitude increase greater than 2 microvolts is sufficient.\nD. It must last longer than 30 seconds." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "A" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_26.json b/sleep_assessment/cases/case43/case43_26.json new file mode 100644 index 0000000000000000000000000000000000000000..6b2890c0cd4c864023fb976cb152da7352112bc1 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_26.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q26", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which criterion defines a periodic leg movement series?\nA. More than 10 leg movements over the entire night are sufficient.\nB. At least 4 consecutive LMs with onset intervals of 5-90 seconds; movements in the two legs beginning less than 5 seconds apart count as one movement.\nC. The interval must be fixed at exactly 20 seconds.\nD. At least 2 leg movements are sufficient, with no restriction on interval." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "B" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_27.json b/sleep_assessment/cases/case43/case43_27.json new file mode 100644 index 0000000000000000000000000000000000000000..f52ca3a18127879edbf3c63dde8f8d6494033ff0 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_27.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q27", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which AASM rule applies to scoring leg movements adjacent to respiratory events?\nA. Exclusion is needed only during REM sleep.\nB. Any movement accompanied by desaturation must be included as a PLM.\nC. Every movement occurring during sleep counts as a PLM.\nD. A leg movement occurring from 0.5 seconds before respiratory-event onset through 0.5 seconds after event end should not be scored as an LM." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "D" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_28.json b/sleep_assessment/cases/case43/case43_28.json new file mode 100644 index 0000000000000000000000000000000000000000..2af9194518dfec5caa70c027d3bb090e97db6aac --- /dev/null +++ b/sleep_assessment/cases/case43/case43_28.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q28", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which statement correctly distinguishes restless legs syndrome (RLS) from periodic limb movement disorder (PLMD)?\nA. An elevated PLMI always proves RLS.\nB. RLS requires PSG, whereas PLMD does not.\nC. RLS and PLMD are synonymous.\nD. RLS is a clinical diagnosis, whereas PLMD diagnosis requires PSG evidence together with relevant sleep disturbance or daytime impairment." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "D" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_29.json b/sleep_assessment/cases/case43/case43_29.json new file mode 100644 index 0000000000000000000000000000000000000000..74c842431062c503590c54758e2fa8d53acd600c --- /dev/null +++ b/sleep_assessment/cases/case43/case43_29.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q29", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which statement most accurately describes the diagnosis of REM sleep behavior disorder (RBD)?\nA. A transient burst of chin EMG activity alone confirms RBD.\nB. Frequent dreaming and a high REM percentage confirm RBD.\nC. There must be recurrent sleep-related vocalization or complex motor behavior occurring or presumed to occur during REM, PSG-confirmed REM sleep without atonia, and no better explanation by another sleep disorder, mental disorder, medication, or substance.\nD. A patient's report of frequent dream-related movement alone confirms RBD." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "C" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_30.json b/sleep_assessment/cases/case43/case43_30.json new file mode 100644 index 0000000000000000000000000000000000000000..0e9a14f7830d4a890c9a4e9e0c47055cf6eb0775 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_30.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q30", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which approach can improve diagnostic yield during PSG for suspected RBD?\nA. Retain only a single chin EMG channel and remove limb EMG channels.\nB. Use an expanded SINBAR EMG montage, such as simultaneous mentalis, flexor digitorum superficialis, and extensor digitorum brevis recordings.\nC. Remove video monitoring to reduce interference.\nD. Assess only changes in SpO2." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "B" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_31.json b/sleep_assessment/cases/case43/case43_31.json new file mode 100644 index 0000000000000000000000000000000000000000..6150b198b514a2edf9d5d29145d1dea5f5eff348 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_31.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q31", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which statement about REM muscle activity and RBD is correct?\nA. Brief, small muscle bursts may occur in normal REM; excessive tonic or phasic EMG activity during REM constitutes RWA/RSWA and must be interpreted with the clinical presentation when assessing RBD.\nB. Low chin muscle tone is absolutely incompatible with RBD.\nC. Normal REM never contains any brief muscle burst.\nD. Any anterior tibialis EMG activity proves RBD." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "A" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_32.json b/sleep_assessment/cases/case43/case43_32.json new file mode 100644 index 0000000000000000000000000000000000000000..6f5d88cb0a9ca9e33823eececefabc618a5b8f28 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_32.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q32", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which statement best matches current commonly used interpretation of the MSLT for narcolepsy?\nA. A mean sleep latency of 8 minutes or less alone diagnoses narcolepsy.\nB. A mean sleep latency of 8 minutes or less plus at least 2 SOREMPs supports narcolepsy; these may be 2 on the MSLT, or 1 on the MSLT plus 1 on the preceding-night PSG.\nC. A mean latency greater than 8 minutes plus 1 SOREMP diagnoses narcolepsy.\nD. The MSLT requires only one nap trial." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "B" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_33.json b/sleep_assessment/cases/case43/case43_33.json new file mode 100644 index 0000000000000000000000000000000000000000..30a8581cf41f897facc84e8c71be9f368c7bc13a --- /dev/null +++ b/sleep_assessment/cases/case43/case43_33.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q33", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which statement correctly describes the two-process model of sleep regulation?\nA. Both processes describe only REM regulation.\nB. Process S is circadian rhythm and Process C is sleep debt.\nC. The model assumes that sleep is completely independent of the biological clock.\nD. Process S is homeostatic sleep pressure that accumulates during wakefulness, while Process C is controlled by the circadian pacemaker." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "D" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_34.json b/sleep_assessment/cases/case43/case43_34.json new file mode 100644 index 0000000000000000000000000000000000000000..1dbb60f6b5c117a700b79a02b30098b05ee4c234 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_34.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q34", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which measure is widely regarded as the gold standard for assessing human circadian phase?\nA. Dim light melatonin onset (DLMO).\nB. Total PSQI score.\nC. Nightly REM percentage.\nD. Sleep-onset latency." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "A" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_35.json b/sleep_assessment/cases/case43/case43_35.json new file mode 100644 index 0000000000000000000000000000000000000000..784d347d8062d9632585bfd646a61d586fd882b2 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_35.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q35", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which statement about adenosine and sleep homeostasis is correct?\nA. Adenosine decreases during prolonged wakefulness and rises during sleep.\nB. Adenosine is unrelated to sleep, and caffeine acts mainly by increasing melatonin.\nC. Adenosine accumulates during wakefulness and promotes sleep pressure, while caffeine alters sleep-wake regulation by antagonizing adenosine signaling.\nD. Adenosine regulates only REM and has no effect on NREM." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "C" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_36.json b/sleep_assessment/cases/case43/case43_36.json new file mode 100644 index 0000000000000000000000000000000000000000..c26dbbea6b792b26da2548da001d5c15691829f7 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_36.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q36", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which statement most accurately describes the role of the VLPO in sleep-wake regulation?\nA. The VLPO is a wake-promoting nucleus that primarily releases dopamine.\nB. The VLPO controls only REM and has no role in NREM.\nC. The VLPO regulates respiration only, not sleep.\nD. Sleep-active VLPO neurons express GABA and galanin and primarily promote NREM sleep." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "D" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_37.json b/sleep_assessment/cases/case43/case43_37.json new file mode 100644 index 0000000000000000000000000000000000000000..79a0415fa056a7c05bacec24896a1a444ee53a90 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_37.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q37", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which statement about orexin (hypocretin) and narcolepsy is correct?\nA. Orexin affects appetite only and does not stabilize sleep-wake states.\nB. Narcolepsy type 1 is primarily associated with loss of hypothalamic orexin-producing neurons or orexin deficiency.\nC. Orexin deficiency causes REM sleep to disappear completely.\nD. Orexin deficiency usually causes insomnia rather than sleepiness." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "B" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_38.json b/sleep_assessment/cases/case43/case43_38.json new file mode 100644 index 0000000000000000000000000000000000000000..c9f69c094b281638066e6adca62c639c4ab5f783 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_38.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q38", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which statement best describes the overall autonomic pattern during normal sleep?\nA. HRV changes occur only during wakefulness; sleep stage has no effect on HRV.\nB. NREM, especially slow-wave sleep, is relatively parasympathetic-dominant, whereas REM shows greater sympathetic fluctuation and autonomic instability.\nC. NREM, especially slow-wave sleep, is sympathetic-dominant, whereas REM is parasympathetic-dominant.\nD. NREM and REM have identical autonomic patterns." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "B" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_39.json b/sleep_assessment/cases/case43/case43_39.json new file mode 100644 index 0000000000000000000000000000000000000000..2b48083bd67c0cf36c4ff8964c5c0989c7e890f1 --- /dev/null +++ b/sleep_assessment/cases/case43/case43_39.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q39", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which statement about the function of sleep spindles is best supported by current evidence?\nA. Sleep spindles are used only to define N2 and have no relation to memory.\nB. Fewer sleep spindles necessarily improve memory.\nC. Sleep spindles are associated with brain plasticity and consolidation of declarative and procedural memory.\nD. Sleep spindles occur only in epilepsy." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "C" + } + } + ] + } +} diff --git a/sleep_assessment/cases/case43/case43_40.json b/sleep_assessment/cases/case43/case43_40.json new file mode 100644 index 0000000000000000000000000000000000000000..bc7bfad359f19b5f65451e8df6de609326c242af --- /dev/null +++ b/sleep_assessment/cases/case43/case43_40.json @@ -0,0 +1,26 @@ +{ + "meta_info": { + "case_id": "Q40", + "bench_subset": "NeuroBench-Sleep", + "difficult": 1, + "original_dataset": "Sleep_QA" + }, + "agent_input": { + "data_path": "", + "instruction": "This is a standalone sleep-medicine multiple-choice question. No sleep data file is required. Select the single correct answer from A, B, C, or D, and clearly state your final option.\n\nQuestion: Which statement best reflects current evidence on sleep, slow-wave activity, and brain CSF or glymphatic clearance?\nA. Glymphatic clearance is most active only during full wakefulness.\nB. The glymphatic system has no established association with sleep stage.\nC. Sleep and slow-wave activity are associated with CSF dynamics and the glymphatic hypothesis, but the direction and magnitude of clearance effects and their causal role in humans remain under investigation and debate.\nD. Only REM sleep is related to the glymphatic system." + }, + "eval_config": { + "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep-medicine multiple-choice responses.\n\n### TASK\nExtract only the agent's final selected option.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. Do not include Markdown or explanatory text.\n3. The key must be exactly \"selected_option\".\n4. The value must be one of \"A\", \"B\", \"C\", \"D\", or null.\n5. If multiple options are mentioned, extract the explicitly identified final answer.\n6. If the final option is missing or ambiguous, return null.\n\n### OUTPUT TEMPLATE\n{\"selected_option\": }", + "metrics": [ + { + "metric_id": "sleep_qa_option_accuracy", + "type": "categorical_check", + "target_key": "selected_option", + "weight": 100, + "params": { + "gt_value": "C" + } + } + ] + } +}