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- metadata.json +8 -0
metadata.json
CHANGED
|
@@ -10,6 +10,14 @@
|
|
| 10 |
"single_image_start"
|
| 11 |
],
|
| 12 |
"task_type": "vqa",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 13 |
"prompt_template": "<image>{{ question }}\nAnswer the question using a single word or phrase.",
|
| 14 |
"mapping_from_source": {
|
| 15 |
"media": {
|
|
|
|
| 10 |
"single_image_start"
|
| 11 |
],
|
| 12 |
"task_type": "vqa",
|
| 13 |
+
"score_type": "rule",
|
| 14 |
+
"score_params": {
|
| 15 |
+
"string_match": "exact"
|
| 16 |
+
},
|
| 17 |
+
"score_protocol": {
|
| 18 |
+
"reference": "lmms-eval@lmms_eval/tasks/websrc/utils.py:133-156 — per-sample token-set F1 (lowercase, word/punct tokenization), averaged per domain then example-weighted overall (websrc_val.yaml metric websrc_squad_f1); official WebSRC (X-LANCE/WebSRC, arXiv:2101.09465 §5) reports Exact Match, F1 and POS.",
|
| 19 |
+
"note": "Official headline includes SQuAD-style F1 and POS (path overlap) which a per-sample binary exact matcher will not reproduce — only the EM component is reproduced; answers are short spans/yes-no so EM is the dominant signal. Published copy ships train+dev (test answers are withheld on the official leaderboard); domain field kept in extras but per-domain aggregation not reproduced per-sample."
|
| 20 |
+
},
|
| 21 |
"prompt_template": "<image>{{ question }}\nAnswer the question using a single word or phrase.",
|
| 22 |
"mapping_from_source": {
|
| 23 |
"media": {
|