| { | |
| "name": "BenchLMM", | |
| "release_date": "2024-07-01", | |
| "subsets": { | |
| "main": { | |
| "language": [ | |
| "en" | |
| ], | |
| "modalities": [ | |
| "single_image_start" | |
| ], | |
| "task_type": "short_answer_qa", | |
| "score_pipeline": [ | |
| "llm-judge" | |
| ], | |
| "score_params": { | |
| "rubric": "scale", | |
| "scale": [ | |
| 0, | |
| 1 | |
| ], | |
| "official_judge_model": "gpt-4-0613" | |
| }, | |
| "score_protocol": { | |
| "reference": "official@github.com/AIFEG/BenchLMM evaluate/gpt_evaluation_script.py — GPT judge prompt: 'Compare the ground truth and prediction from AI models, to give a correctness score for the prediction' ignoring case/grammar, '/' = multiple acceptable answers, similar meaning gets full marks; judge outputs a score in {0.0,0.1,...,1.0}; example model gpt-4-0613; final metric = mean score (avg_score.py).", | |
| "note": "Every sample is LLM-judged with fractional credit on a 0-1 scale (11 discrete values); headline is the mean per-sample score, not binary accuracy. The mm-eval copy has only 96 rows (test split) whereas official BenchLMM spans many style/domain files (CT, MRI, infrared, AD, RS, styles, Robots/Games, each with its own eval variant) — published copy appears to be a small subset." | |
| }, | |
| "prompt_template": "<image>{{ question }}\nAnswer the question using a single word or phrase.", | |
| "mapping_from_source": { | |
| "media": { | |
| "from": "image", | |
| "type": "list", | |
| "min_items": 1, | |
| "max_items": 1 | |
| }, | |
| "id": { | |
| "from": "id" | |
| }, | |
| "question": { | |
| "from": "question" | |
| }, | |
| "answer": { | |
| "from": "answer", | |
| "optional": true | |
| }, | |
| "source": { | |
| "format": "json", | |
| "url": { | |
| "test": "https://huggingface.co/datasets/AIFEG/BenchLMM" | |
| } | |
| } | |
| }, | |
| "prompt_template_source": { | |
| "origin": "official", | |
| "reference": "https://github.com/AIFEG/BenchLMM/blob/main/baseline/LLaVA/BenchGPT_LLaVA_model_vqa.py#L44-L49 (qs = line['text']; DEFAULT_IMAGE_TOKEN + '\\n' + qs); trailer '\\nAnswer the question using a single word or phrase.' is embedded in every official jsonl 'text' field, e.g. https://github.com/AIFEG/BenchLMM/blob/main/jsonl/Benchmark_style_cartoon.jsonl", | |
| "notes": "Tier 1: BenchLMM authors' own inference script feeds <image>\\n + question, and the official jsonl question text carries the '\\nAnswer the question using a single word or phrase.' trailer; the mm-eval template reproduces this (trailer re-appended because the mm-eval question column is trailer-free)." | |
| } | |
| } | |
| } | |
| } |