metadata: migrate score_type -> score_pipeline (atomic stage contract; see mm-eval scorer docs/en/SCORING.md)
Browse files- metadata.json +65 -62
metadata.json
CHANGED
|
@@ -1,65 +1,68 @@
|
|
| 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 |
-
|
| 65 |
-
}
|
|
|
|
| 1 |
{
|
| 2 |
+
"name": "Mementos",
|
| 3 |
+
"release_date": "2024-01-17",
|
| 4 |
+
"subsets": {
|
| 5 |
+
"default": {
|
| 6 |
+
"language": [
|
| 7 |
+
"en"
|
| 8 |
+
],
|
| 9 |
+
"modalities": [
|
| 10 |
+
"single_image_start"
|
| 11 |
+
],
|
| 12 |
+
"task_type": "captioning",
|
| 13 |
+
"score_pipeline": {
|
| 14 |
+
"unsupported": "keyword-set-f1",
|
| 15 |
+
"reason": "requires keyword extraction + set-F1 grading (Mementos); use the benchmark's own evaluator"
|
| 16 |
+
},
|
| 17 |
+
"score_params": {
|
| 18 |
+
"compare": "keyword_set_f1"
|
| 19 |
+
},
|
| 20 |
+
"score_protocol": {
|
| 21 |
+
"reference": "Official repo https://github.com/umd-huang-lab/Mementos GPT-4-assisted_evaluation.ipynb — \"Generate keywords list\" cell: GPT-4 (gpt-4-1106-preview, temperature 0) jointly extracts object and action keyword lists from the human reference description AND the model description in a single call, normalizing synonyms across the four lists; \"Object evaluation\" / \"Behavior evaluation\" cells: deterministic set precision/recall/F1 of the extracted lists after synonym replacement using per-domain pickled graphs from sym_graphs/. Not implemented in VLMEvalKit or lmms-eval.",
|
| 22 |
+
"note": "Official metric is LLM keyword extraction followed by rule-based set-F1 (separate object-F1 and behavior-F1 headlines, averaged per domain), NOT a judge rating. Static ground-truth keyword annotations do NOT exist in the official release: keyword lists are produced at eval time by the joint GPT-4 extraction over (reference description, model description), so per-row object/behavior keyword lists cannot be shipped without fabrication. This copy ships the exact official join keys instead: extras image_name and domain match the official annotation CSVs on github.com/umd-huang-lab/Mementos (dl_description.csv / robo_description.csv / cmc_description.csv; 697/697 filename join on the test split) and select the per-domain synonym graph. synonym_graph_ref: https://github.com/umd-huang-lab/Mementos/tree/main/sym_graphs — {dl,cmc,robo}_{object,action}_graph_v1.pkl (dl=dailylife, cmc=comics, robo=robotics), loadable via load_graph() in sym_graphs/build_action_graph.ipynb; graphs are intentionally not embedded in this dataset. Official annotations cover only the test split (697 rows); the train split (4060 rows) carries reference descriptions from the source metadata.csv but has no official eval coverage. Each row has exactly one image because the official furonghuang-lab/Mementos release stitches every image sequence into a single composite PNG (no per-frame files are published); n_images=1 is faithful to the source, not a flattening defect. This re-conversion also fixed a truncation defect: the previous mm-eval copy carried only the first 1826 dailylife train rows; train now carries all 4060 source rows (previously published ids preserved verbatim; recovered rows appended as mementos_train_1826..4059 in dailylife->robotics->comics metadata.csv order)."
|
| 23 |
+
},
|
| 24 |
+
"prompt_template": "<image>{{ question }}",
|
| 25 |
+
"prompt_template_source": {
|
| 26 |
+
"origin": "official",
|
| 27 |
+
"reference": "Mementos paper arXiv:2401.10529, Appendix C Table 6 (\"Generate description for the given image sequence\") — prompt reproduced verbatim in the question field; official repo https://github.com/umd-huang-lab/Mementos",
|
| 28 |
+
"notes": "Tier 1: dataset authors publish the exact model-input prompt in the paper appendix; every row uses it verbatim."
|
| 29 |
+
},
|
| 30 |
+
"mapping_from_source": {
|
| 31 |
+
"media": {
|
| 32 |
+
"from": "image",
|
| 33 |
+
"type": "list",
|
| 34 |
+
"min_items": 1,
|
| 35 |
+
"max_items": 1
|
| 36 |
+
},
|
| 37 |
+
"id": {
|
| 38 |
+
"from": "id"
|
| 39 |
+
},
|
| 40 |
+
"question": {
|
| 41 |
+
"from": "question"
|
| 42 |
+
},
|
| 43 |
+
"answer": {
|
| 44 |
+
"from": "answer",
|
| 45 |
+
"optional": true
|
| 46 |
+
},
|
| 47 |
+
"extra": {
|
| 48 |
+
"reference_response": {
|
| 49 |
+
"from": "reference_response"
|
| 50 |
+
},
|
| 51 |
+
"image_name": {
|
| 52 |
+
"from": "image_name"
|
| 53 |
+
},
|
| 54 |
+
"domain": {
|
| 55 |
+
"from": "domain"
|
| 56 |
+
}
|
| 57 |
+
},
|
| 58 |
+
"source": {
|
| 59 |
+
"format": "huggingface",
|
| 60 |
+
"url": {
|
| 61 |
+
"test": "https://huggingface.co/datasets/furonghuang-lab/Mementos",
|
| 62 |
+
"train": "https://huggingface.co/datasets/furonghuang-lab/Mementos"
|
| 63 |
+
}
|
| 64 |
+
}
|
| 65 |
+
}
|
| 66 |
}
|
|
|
|
| 67 |
}
|
| 68 |
+
}
|
|
|