metadata: migrate score_type -> score_pipeline (atomic stage contract; see mm-eval scorer docs/en/SCORING.md)
Browse files- metadata.json +64 -61
metadata.json
CHANGED
|
@@ -1,64 +1,67 @@
|
|
| 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 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 60 |
}
|
| 61 |
-
}
|
| 62 |
}
|
| 63 |
-
|
| 64 |
-
}
|
|
|
|
| 1 |
{
|
| 2 |
+
"name": "MathVision",
|
| 3 |
+
"release_date": "2026-07-07",
|
| 4 |
+
"subsets": {
|
| 5 |
+
"main": {
|
| 6 |
+
"language": [
|
| 7 |
+
"en"
|
| 8 |
+
],
|
| 9 |
+
"modalities": [
|
| 10 |
+
"single_image_start"
|
| 11 |
+
],
|
| 12 |
+
"task_type": "multiple_choice_qa",
|
| 13 |
+
"score_pipeline": [
|
| 14 |
+
"exact-match",
|
| 15 |
+
"rule-match"
|
| 16 |
+
],
|
| 17 |
+
"score_protocol": {
|
| 18 |
+
"reference": "official mathllm/MATH-V@evaluation/evaluate.py — purely rule-based: boxed/'the answer is' extraction, find_math_answer, is_equal (latex2sympy symbolic compare), no LLM call; identical logic vendored at lmms-eval@lmms_eval/tasks/mathvision/utils.py:79-137 (mathvision_standard_eval) + eval_utils.py.",
|
| 19 |
+
"note": "Official is_equal does latex2sympy symbolic equivalence and also accepts the option TEXT for MCQ rows (gt letter or option value); the scorer's exact/template chain reproduces letter/number matches but not symbolic LaTeX equivalence. VLMEvalKit disagrees with the official protocol (utils/mathv.py:116-160: rule prefetch then GPT answer extraction = llm_extract); official rule protocol preferred. The prompt is the official MATH-V boxed prompt (models/GPT4V.py process_example_mergedimg), so boxed template extraction applies. Rows whose options are the placeholder list ['A','B','C','D','E'] have the option contents printed inside the image; the official prompt suppresses the options block for them and the options dict {A:'A',...,E:'E'} is carried verbatim from the source."
|
| 20 |
+
},
|
| 21 |
+
"prompt_template": "Please solve the problem and put your answer in one \"\\boxed{}\". If it is a multiple choice question, only one letter is allowed in the \"\\boxed{}\".\n<image>{{ question }}\n{% if options and options.values()|join('') != 'ABCDE' %}{% for k, v in options.items() %}({{ k }}) {{ v }}\n{% endfor %}{% endif %}",
|
| 22 |
+
"prompt_template_source": {
|
| 23 |
+
"origin": "official",
|
| 24 |
+
"reference": "https://github.com/mathllm/MATH-V/blob/main/models/GPT4V.py#L109-L117 (process_example_mergedimg, the __main__ inference path: boxed instruction + question + '(A) ...' options block only when options are non-empty and not the in-image placeholder list ABCDE; same suppression in geninput L22-L30)",
|
| 25 |
+
"notes": "Tier 1: official MATH-V inference code. Byte-identical to the official prompt except a single <image> placeholder inserted before the question (the HF dataset ships one merged image per row; the official code attaches it separately). lmms-eval (lmms_eval/tasks/mathvision/utils.py L33-L48) equivalently emits a 'Choices:' block only when options are present."
|
| 26 |
+
},
|
| 27 |
+
"mapping_from_source": {
|
| 28 |
+
"source": {
|
| 29 |
+
"format": "huggingface",
|
| 30 |
+
"url": {
|
| 31 |
+
"test": "https://huggingface.co/datasets/MathLLMs/MathVision",
|
| 32 |
+
"testmini": "https://huggingface.co/datasets/MathLLMs/MathVision"
|
| 33 |
+
}
|
| 34 |
+
},
|
| 35 |
+
"id": {
|
| 36 |
+
"from": "id"
|
| 37 |
+
},
|
| 38 |
+
"question": {
|
| 39 |
+
"from": "question"
|
| 40 |
+
},
|
| 41 |
+
"options": {
|
| 42 |
+
"from": "options",
|
| 43 |
+
"optional": true,
|
| 44 |
+
"note": "list source values are normalized to {A,B,...} dict; empty list stays {}. 384/3040 test (73/304 testmini) MCQ rows carry the placeholder list ['A','B','C','D','E'] meaning the option contents are printed inside the image; the prompt template suppresses the options block for exactly those rows, mirroring official handling."
|
| 45 |
+
},
|
| 46 |
+
"answer": {
|
| 47 |
+
"from": "answer",
|
| 48 |
+
"optional": true
|
| 49 |
+
},
|
| 50 |
+
"media": {
|
| 51 |
+
"from": "decoded_image",
|
| 52 |
+
"type": "list",
|
| 53 |
+
"min_items": 1,
|
| 54 |
+
"max_items": 1
|
| 55 |
+
},
|
| 56 |
+
"extra": {
|
| 57 |
+
"level": {
|
| 58 |
+
"from": "level"
|
| 59 |
+
},
|
| 60 |
+
"subject": {
|
| 61 |
+
"from": "subject"
|
| 62 |
+
}
|
| 63 |
+
}
|
| 64 |
+
}
|
| 65 |
}
|
|
|
|
| 66 |
}
|
| 67 |
+
}
|
|
|