File size: 3,351 Bytes
d846fa6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
{
    "meta_info": {
        "case_id": "Mumtaz2016_02.edf",
        "bench_subset": "NeuroBench-Core",
        "difficult": 1.5,
        "original_dataset": "Mumtaz2016"
    },
    "agent_input": {
        "data_path": "data/core/Mumtaz2016_02.edf",
        "instruction": "Please select EEG channels from the raw signal, apply an 8-13Hz alpha-band FIR (linear-phase) bandpass filter, and then use the first 30 seconds of the filtered signal to compute inter-channel correlations using the Pearson correlation coefficient. Then report: (1) the global synchronization value defined as the mean absolute correlation across all channel pairs, and (2) the single channel pair with the highest correlation coefficient."
    },
    "eval_config": {
        "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for neurophysiological data. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract two final outputs from the agent report:\n1) global synchronization value in alpha band\n2) highest-correlation channel pair\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include any conversational text or explanations.\n4. Keys in JSON must be EXACTLY \"global_sync\" and \"top_pair\".\n5. \"global_sync\" must be a FLOAT.\n6. To extract \"top_pair\", first identify the two complete source channel labels. A referenced label such as F4-A1 is one complete channel label, and its hyphen must not be treated as the pair separator.\n7. For each source channel separately, uppercase it and remove one terminal reference suffix if present: -A1, -A2, -LE, -REF, -M1, -M2, or -AVG. Never output a reference electrode (A1/A2/LE/REF/M1/M2/AVG) as a pair member.\n8. Convert legacy aliases after removing the reference suffix: T3->T7, T4->T8, T5->P7, T6->P8.\n9. Sort the two normalized channel labels in ascending lexicographic order, then join them with exactly one hyphen using format \"CH1-CH2\" (uppercase, no spaces). If the report already uses a canonical pair such as F4-C4, normalize it to C4-F4. Each final label may contain only uppercase A-Z letters and digits; mixed-case forms such as Fp1 or Fp2 are forbidden and must become FP1 or FP2.\n10. Required examples: \"F4-A1 and C4-A1\" -> \"C4-F4\"; \"F4-C4\" -> \"C4-F4\"; \"T4-A1 and T3-A2\" -> \"T7-T8\"; \"Fp2 and Fp1\" -> \"FP1-FP2\".\n11. If the pair cannot be resolved into exactly two non-reference channels, set \"top_pair\" to null; do not guess. If global_sync is missing, set only global_sync to null.\n\n### OUTPUT TEMPLATE\n{\"global_sync\": <float|null>, \"top_pair\": <string|null>}",
        "metrics": [
            {
                "metric_id": "alpha_global_sync_accuracy",
                "type": "numeric_check",
                "target_key": "global_sync",
                "weight": 50,
                "params": {
                    "gt_value": 0.495031,
                    "tolerance": 0.02475155
                }
            },
            {
                "metric_id": "highest_corr_pair_accuracy",
                "type": "categorical_check",
                "target_key": "top_pair",
                "weight": 50,
                "params": {
                    "gt_value": "F3-FP1"
                }
            }
        ]
    }
}