| { |
| "name": "LiveBench", |
| "release_date": "2024-09-30", |
| "subsets": { |
| "2024-05": { |
| "language": [ |
| "en" |
| ], |
| "modalities": [ |
| "single_image_start" |
| ], |
| "task_type": "open_ended_qa", |
| "score_pipeline": [ |
| "llm-judge" |
| ], |
| "score_params": { |
| "rubric": "scale", |
| "scale": [ |
| 0, |
| 10 |
| ], |
| "official_judge_model": "gpt-4o" |
| }, |
| "score_protocol": { |
| "reference": "lmms-eval@lmms_eval/tasks/live_bench/utils.py:66-68,82-92,160-186 (v1 judge; utils.py:164 reads doc['criteria']) — this is lmms-lab/LiveBench (live web-news screenshots), NOT Abacus/NYU text LiveBench nor OpenCompass LiveMMBench; GPT-4o rates every answer 0-10 against ground truth with the screenshot images attached; aggregate = mean(rating/10)*100 with subtask breakdown. NOTE: 2024-05 is not in the official live_bench.yaml task group (only 2406/2407/2409).", |
| "note": "Source config 2024-05 does not ship a `criteria` column at all (verified 2026-07-07: lmms-lab/LiveBench config 2024-05 features lack `criteria`; the column first appears in 2024-06), so the official judge prompt cannot be fully populated for this month and no `criteria` extra is carried. Official judge is multimodal (sees the screenshots). 0-10 rating averaged then scaled x10 to 0-100." |
| }, |
| "prompt_template": "{% for _ in range(n_images) %}<image>{% endfor %}{{ question }}", |
| "prompt_template_source": { |
| "origin": "official", |
| "reference": "https://huggingface.co/datasets/lmms-lab/LiveBench", |
| "notes": "Tier 1: LiveBench live-web VQA, monthly config 2024-05; single-prompt eval per dataset card." |
| }, |
| "mapping_from_source": { |
| "media": { |
| "from": "images", |
| "type": "list", |
| "min_items": 1, |
| "max_items": 1 |
| }, |
| "id": { |
| "from": "id" |
| }, |
| "question": { |
| "from": "question" |
| }, |
| "answer": { |
| "from": "answer", |
| "optional": true |
| }, |
| "extra": { |
| "subtask": { |
| "from": "subtask" |
| }, |
| "website": { |
| "from": "website" |
| }, |
| "data_generator": { |
| "from": "data_generator" |
| }, |
| "n_images": { |
| "from": "n_images" |
| } |
| }, |
| "source": { |
| "format": "json", |
| "url": { |
| "2024-05_test": "https://huggingface.co/datasets/lmms-lab/LiveBench" |
| } |
| } |
| } |
| }, |
| "2024-06": { |
| "language": [ |
| "en" |
| ], |
| "modalities": [ |
| "single_image_start" |
| ], |
| "task_type": "open_ended_qa", |
| "score_pipeline": [ |
| "llm-judge" |
| ], |
| "score_params": { |
| "rubric": "scale", |
| "scale": [ |
| 0, |
| 10 |
| ], |
| "official_judge_model": "gpt-4o", |
| "judge_inputs": [ |
| "criteria" |
| ] |
| }, |
| "score_protocol": { |
| "reference": "lmms-eval@lmms_eval/tasks/live_bench/live_bench.yaml + live_bench_2406.yaml (includes live_bench_template_yaml -> utils.py) — utils.py:66-68 judge prompt interpolates per-sample `criteria`, utils.py:164 reads doc['criteria']; GPT-4o rates every answer 0-10 against ground truth with the screenshot images attached; aggregate = mean(rating/10)*100.", |
| "note": "Per-sample `criteria` column is carried as extra (fixed 2026-07-07; the earlier mm-eval copy omitted it). Official judge is multimodal (sees the screenshots) — a text-only judge deviates from official. 0-10 rating averaged then scaled x10 to 0-100. Some rows are MCQ/fill-in-the-blank phrased inline but all are judged with the same 0-10 GPT rubric." |
| }, |
| "prompt_template": "{% for _ in range(n_images) %}<image>{% endfor %}{{ question }}", |
| "prompt_template_source": { |
| "origin": "official", |
| "reference": "https://huggingface.co/datasets/lmms-lab/LiveBench", |
| "notes": "Tier 1: LiveBench live-web VQA, monthly config 2024-06; single-prompt eval per dataset card." |
| }, |
| "mapping_from_source": { |
| "media": { |
| "from": "images", |
| "type": "list", |
| "min_items": 1, |
| "max_items": 1 |
| }, |
| "id": { |
| "from": "id" |
| }, |
| "question": { |
| "from": "question" |
| }, |
| "answer": { |
| "from": "answer", |
| "optional": true |
| }, |
| "extra": { |
| "subtask": { |
| "from": "subtask" |
| }, |
| "website": { |
| "from": "website" |
| }, |
| "data_generator": { |
| "from": "data_generator" |
| }, |
| "n_images": { |
| "from": "n_images" |
| }, |
| "criteria": { |
| "from": "criteria" |
| } |
| }, |
| "source": { |
| "format": "json", |
| "url": { |
| "2024-06_test": "https://huggingface.co/datasets/lmms-lab/LiveBench" |
| } |
| } |
| } |
| }, |
| "2024-07": { |
| "language": [ |
| "en" |
| ], |
| "modalities": [ |
| "single_image_start" |
| ], |
| "task_type": "open_ended_qa", |
| "score_pipeline": [ |
| "llm-judge" |
| ], |
| "score_params": { |
| "rubric": "scale", |
| "scale": [ |
| 0, |
| 10 |
| ], |
| "official_judge_model": "gpt-4o", |
| "judge_inputs": [ |
| "criteria" |
| ] |
| }, |
| "score_protocol": { |
| "reference": "lmms-eval@lmms_eval/tasks/live_bench/live_bench.yaml + live_bench_2407.yaml (includes live_bench_template_yaml -> utils.py) — utils.py:66-68 judge prompt interpolates per-sample `criteria`, utils.py:164 reads doc['criteria']; GPT-4o rates every answer 0-10 with the screenshot images attached; aggregate = mean(rating/10)*100.", |
| "note": "Per-sample `criteria` column is carried as extra (fixed 2026-07-07; the earlier mm-eval copy omitted it). Official judge is multimodal (sees the screenshots) — a text-only judge deviates from official. 0-10 rating averaged then scaled x10 to 0-100." |
| }, |
| "prompt_template": "{% for _ in range(n_images) %}<image>{% endfor %}{{ question }}", |
| "prompt_template_source": { |
| "origin": "official", |
| "reference": "https://huggingface.co/datasets/lmms-lab/LiveBench", |
| "notes": "Tier 1: LiveBench live-web VQA, monthly config 2024-07; single-prompt eval per dataset card." |
| }, |
| "mapping_from_source": { |
| "media": { |
| "from": "images", |
| "type": "list", |
| "min_items": 1, |
| "max_items": 1 |
| }, |
| "id": { |
| "from": "id", |
| "note": "Composite key {source_id}_q{occurrence_index} produced by walking rows in upstream order (string-sorted by id, stable) and counting per-source-id occurrences (1-indexed); the source config ships duplicate `id` values." |
| }, |
| "question": { |
| "from": "question" |
| }, |
| "answer": { |
| "from": "answer", |
| "optional": true |
| }, |
| "extra": { |
| "subtask": { |
| "from": "subtask" |
| }, |
| "website": { |
| "from": "website" |
| }, |
| "data_generator": { |
| "from": "data_generator" |
| }, |
| "n_images": { |
| "from": "n_images" |
| }, |
| "criteria": { |
| "from": "criteria" |
| }, |
| "source_id": { |
| "from": "id", |
| "note": "Original upstream id before composite-id formation." |
| } |
| }, |
| "source": { |
| "format": "json", |
| "url": { |
| "2024-07_test": "https://huggingface.co/datasets/lmms-lab/LiveBench" |
| } |
| } |
| } |
| }, |
| "2024-08": { |
| "language": [ |
| "en" |
| ], |
| "modalities": [ |
| "single_image_start" |
| ], |
| "task_type": "open_ended_qa", |
| "score_pipeline": [ |
| "llm-judge" |
| ], |
| "score_params": { |
| "rubric": "scale", |
| "scale": [ |
| 0, |
| 10 |
| ], |
| "official_judge_model": "gpt-4o", |
| "judge_inputs": [ |
| "criteria" |
| ] |
| }, |
| "score_protocol": { |
| "reference": "lmms-eval@lmms_eval/tasks/live_bench/utils.py:66-68,164 and utils_v2.py:67-69,163 — same GPT-4o 0-10 judge protocol consuming per-sample `criteria`; 2024-08 config is not listed in the live_bench.yaml task group.", |
| "note": "Only 8 rows in this config (source upload appears truncated) — results statistically meaningless. Per-sample `criteria` column is carried as extra (fixed 2026-07-07; the earlier mm-eval copy omitted it). Official judge is multimodal (sees the screenshots) — a text-only judge deviates from official. 0-10 rating averaged then scaled x10 to 0-100." |
| }, |
| "prompt_template": "{% for _ in range(n_images) %}<image>{% endfor %}{{ question }}", |
| "prompt_template_source": { |
| "origin": "official", |
| "reference": "https://huggingface.co/datasets/lmms-lab/LiveBench", |
| "notes": "Tier 1: LiveBench live-web VQA, monthly config 2024-08; single-prompt eval per dataset card." |
| }, |
| "mapping_from_source": { |
| "media": { |
| "from": "images", |
| "type": "list", |
| "min_items": 1, |
| "max_items": 1 |
| }, |
| "id": { |
| "from": "id" |
| }, |
| "question": { |
| "from": "question" |
| }, |
| "answer": { |
| "from": "answer", |
| "optional": true |
| }, |
| "extra": { |
| "subtask": { |
| "from": "subtask" |
| }, |
| "website": { |
| "from": "website" |
| }, |
| "data_generator": { |
| "from": "data_generator" |
| }, |
| "n_images": { |
| "from": "n_images" |
| }, |
| "criteria": { |
| "from": "criteria" |
| } |
| }, |
| "source": { |
| "format": "json", |
| "url": { |
| "2024-08_test": "https://huggingface.co/datasets/lmms-lab/LiveBench" |
| } |
| } |
| } |
| }, |
| "2024-09": { |
| "language": [ |
| "en" |
| ], |
| "modalities": [ |
| "single_image_start" |
| ], |
| "task_type": "open_ended_qa", |
| "score_pipeline": [ |
| "llm-judge" |
| ], |
| "score_params": { |
| "rubric": "scale", |
| "scale": [ |
| 0, |
| 10 |
| ], |
| "official_judge_model": "gpt-4o", |
| "judge_inputs": [ |
| "criteria" |
| ] |
| }, |
| "score_protocol": { |
| "reference": "lmms-eval@lmms_eval/tasks/live_bench/live_bench.yaml + live_bench_2409.yaml (includes live_bench_template_yaml_v2 -> utils_v2.py) — utils_v2.py:67-69 judge prompt interpolates per-sample `criteria`, utils_v2.py:163 reads doc['criteria']; GPT-4o rates 0-10 with the screenshot images attached; subtask taxonomy v2 (Concrete Recognition/Analytical Questions/Divergent Thinking/Real-world Assistance); aggregate = mean(rating/10)*100.", |
| "note": "Per-sample `criteria` column is carried as extra (fixed 2026-07-07; the earlier mm-eval copy omitted it). Official judge is multimodal (sees the screenshots) — a text-only judge deviates from official. 0-10 rating averaged then scaled x10 to 0-100." |
| }, |
| "prompt_template": "{% for _ in range(n_images) %}<image>{% endfor %}{{ question }}", |
| "prompt_template_source": { |
| "origin": "official", |
| "reference": "https://huggingface.co/datasets/lmms-lab/LiveBench", |
| "notes": "Tier 1: LiveBench live-web VQA, monthly config 2024-09; single-prompt eval per dataset card." |
| }, |
| "mapping_from_source": { |
| "media": { |
| "from": "images", |
| "type": "list", |
| "min_items": 1, |
| "max_items": 1 |
| }, |
| "id": { |
| "from": "id", |
| "note": "Composite key {source_id}_q{occurrence_index} produced by walking rows in upstream order (string-sorted by id, stable) and counting per-source-id occurrences (1-indexed); the source config ships duplicate `id` values." |
| }, |
| "question": { |
| "from": "question" |
| }, |
| "answer": { |
| "from": "answer", |
| "optional": true |
| }, |
| "extra": { |
| "subtask": { |
| "from": "subtask" |
| }, |
| "website": { |
| "from": "website" |
| }, |
| "data_generator": { |
| "from": "data_generator" |
| }, |
| "n_images": { |
| "from": "n_images" |
| }, |
| "criteria": { |
| "from": "criteria" |
| }, |
| "source_id": { |
| "from": "id", |
| "note": "Original upstream id before composite-id formation." |
| } |
| }, |
| "source": { |
| "format": "json", |
| "url": { |
| "2024-09_test": "https://huggingface.co/datasets/lmms-lab/LiveBench" |
| } |
| } |
| } |
| } |
| } |
| } |
