ZTWHHH commited on
Commit
457133c
·
verified ·
1 Parent(s): aed3e52

metadata: add score_type/score_params/score_protocol (official grading protocol taxonomy); task_type is now format-only (judge_scored migrated to vqa + llm_judge). See dataset-upload skill references/metadata-json.md.

Browse files
Files changed (1) hide show
  1. metadata.json +5 -0
metadata.json CHANGED
@@ -10,6 +10,11 @@
10
  "single_image_start"
11
  ],
12
  "task_type": "multiple_choice_vqa",
 
 
 
 
 
13
  "prompt_template": "<image>{{ question }}\n{% for k, v in options.items() %}{{ k }}. {{ v }}\n{% endfor %}Answer with the option's letter from the given choices directly.",
14
  "mapping_from_source": {
15
  "media": {
 
10
  "single_image_start"
11
  ],
12
  "task_type": "multiple_choice_vqa",
13
+ "score_type": "llm_extract",
14
+ "score_protocol": {
15
+ "reference": "vlmevalkit@vlmeval/dataset/image_mcq.py:249-324 (ImageMCQDataset.evaluate_heuristic) -> vlmevalkit@vlmeval/dataset/utils/multiple_choice.py:475-508 mcq_vanilla_eval: can_infer rule prefetch, then GPT extractor (extract_answer_from_item @vlmeval/dataset/utils/multiple_choice.py:359-407) maps verbose output to a letter, rule-compare vs GT; AI2D_TEST registered @vlmeval/dataset/image_mcq.py:100. Cross-check: lmms-eval@lmms_eval/tasks/ai2d/ai2d.yaml:41-55 (MultiChoiceRegexFilter + exact_match, pure rule)",
16
+ "note": "mm-eval prompt follows the lmms-eval trailer; lmms-eval grades AI2D with deterministic regex extraction only, VLMEvalKit adds the GPT extraction fallback. Both are plain per-sample accuracy — no aggregation caveat."
17
+ },
18
  "prompt_template": "<image>{{ question }}\n{% for k, v in options.items() %}{{ k }}. {{ v }}\n{% endfor %}Answer with the option's letter from the given choices directly.",
19
  "mapping_from_source": {
20
  "media": {