| {
|
| "meta_info": {
|
| "case_id": "SEED-V-04.cnt",
|
| "bench_subset": "NeuroBench-Core",
|
| "difficult": 2.0,
|
| "original_dataset": "SEED-V"
|
| },
|
| "agent_input": {
|
| "data_path": "data/core/SEED-V-04.cnt",
|
| "instruction": "Please select standard EEG channels and apply an 8-13Hz zero-phase FIR bandpass filter to the selected continuous signals. Then use the first 60 seconds of the filtered data, extract instantaneous phase with the Hilbert transform, compute PLV for every unique channel pair, report the global mean PLV, and report the top-3 channel pairs ranked from highest to lowest PLV."
|
| },
|
| "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 outputs from the agent report:\n1) global mean PLV\n2) top-3 channel pairs ranked by PLV\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 must be EXACTLY \"global_mean_plv\" and \"top3_plv_pairs\".\n5. \"global_mean_plv\" must be a FLOAT or null.\n6. \"top3_plv_pairs\" must be an array of up to 3 strings or null. Normalize EACH reported pair independently.\n7. 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.\n8. 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.\n9. Convert legacy aliases after removing the reference suffix: T3->T7, T4->T8, T5->P7, T6->P8.\n10. 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.\n11. 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\".\n12. Preserve the pairs in descending PLV order as reported. Do not sort the array lexicographically. If any pair cannot be resolved into exactly two non-reference channels, set \"top3_plv_pairs\" to null; do not guess.\n\n### OUTPUT TEMPLATE\n{\"global_mean_plv\": <float|null>, \"top3_plv_pairs\": [\"CH1-CH2\",\"CH3-CH4\",\"CH5-CH6\"]}",
|
| "metrics": [
|
| {
|
| "metric_id": "global_mean_plv_accuracy",
|
| "type": "numeric_check",
|
| "target_key": "global_mean_plv",
|
| "weight": 40,
|
| "params": {
|
| "gt_value": 0.35428962170927186,
|
| "tolerance": 0.017714481085463595
|
| }
|
| },
|
| {
|
| "metric_id": "top3_plv_pair_set_accuracy",
|
| "type": "set_match_check",
|
| "target_key": "top3_plv_pairs",
|
| "weight": 60,
|
| "params": {
|
| "gt_value": [
|
| "O2-PO4",
|
| "PO6-PO8",
|
| "P4-P6"
|
| ],
|
| "match_mode": "per_hit_fraction",
|
| "top_k": 3,
|
| "score_per_hit": 0.3333333333
|
| }
|
| }
|
| ]
|
| }
|
| }
|
|
|
