| { |
| "name": "MEGA-Bench", |
| "release_date": "2024-10-18", |
| "subsets": { |
| "main": { |
| "language": [ |
| "en" |
| ], |
| "modalities": [ |
| "multi_image_interleave", |
| "text" |
| ], |
| "task_type": "short_answer_qa", |
| "score_pipeline": { |
| "unsupported": "per-task-metric", |
| "reason": "requires the benchmark's own evaluator (per-row metric dispatch)" |
| }, |
| "score_params": { |
| "metric_field": "task_name" |
| }, |
| "score_protocol": { |
| "reference": "lmms-eval@lmms_eval/tasks/megabench/metrics/metric_type.py:1-120 + metrics/scoring/* (45 per-task metric types: exact/near str match, set/dict equality+Jaccard, BLEU/GLEU, Levenshtein similarity, nbbox IoU, number_rel_diff_ratio, LaTeX equality, program_judge, symbolic planning, VLM-as-judge, ...) and tasks/megabench/README.md + evaluator.py — official protocol (adapted from TIGER-AI-Lab/MEGA-Bench) runs a stand-alone evaluator that dispatches each task to its own metric; core split = rule metrics, open split = GPT(-4o) judge; headline = macro mean over ~505 tasks.", |
| "note": "The mm-eval copy merges test_core and test_open into one subset and stores only task_name in extra — the per-task metric/aggregation/parsing config required by the official evaluator is not shipped, so no executable score_pipeline can reproduce the official protocol (declared unsupported). Many metrics are fractional per-sample; official headline is macro-averaged per task, not per-sample mean." |
| }, |
| "prompt_template": "{{ question }}", |
| "mapping_from_source": { |
| "media": { |
| "from": "images", |
| "type": "list", |
| "min_items": 0, |
| "max_items": 21 |
| }, |
| "id": { |
| "from": "id" |
| }, |
| "question": { |
| "from": "question" |
| }, |
| "answer": { |
| "from": "answer", |
| "optional": true |
| }, |
| "extra": { |
| "task_name": { |
| "from": "task_name" |
| } |
| }, |
| "source": { |
| "format": "json", |
| "url": { |
| "test_core": "https://huggingface.co/datasets/TIGER-Lab/MEGA-Bench", |
| "test_open": "https://huggingface.co/datasets/TIGER-Lab/MEGA-Bench" |
| } |
| } |
| }, |
| "prompt_template_source": { |
| "origin": "official", |
| "reference": "https://github.com/TIGER-AI-Lab/MEGA-Bench/blob/main/megabench/loader.py (task_description+example_text+query_text concatenated into question column at conversion time; template is bare {{ question }})", |
| "notes": "Tier 1: MEGA-Bench: per-task scaffolding (task_description, example_text, query_text) is concatenated into the question column at conversion time; template is intentionally bare to avoid double-scaffolding." |
| } |
| } |
| } |
| } |
