| { |
| "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\": <float|null>, \"full_text\": <string|null>}", |
| "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." |
| } |
| } |
| ] |
| } |
| } |
|
|
