Flickr8k / metadata.json
ZTWHHH's picture
metadata: migrate score_type -> score_pipeline (atomic stage contract; see mm-eval scorer docs/en/SCORING.md)
a347b6e verified
Raw
History Blame Contribute Delete
2.3 kB
{
"name": "Flickr8k",
"release_date": "2026-05-25",
"subsets": {
"main": {
"language": [
"en"
],
"modalities": [
"single_image_start"
],
"task_type": "captioning",
"score_pipeline": {
"unsupported": "caption-metrics",
"reason": "requires corpus-level caption metrics (CIDEr/BLEU/ROUGE/BERTScore); use the benchmark's own evaluator"
},
"score_params": {
"metrics": [
"cider",
"bleu4",
"meteor",
"rouge_l"
]
},
"score_protocol": {
"reference": "lmms-eval@lmms_eval/tasks/flickr30k/utils.py:5,42 — nearest framework task (Flickr30k) uses the same pycocoevalcap corpus scorers (Bleu1-4, METEOR, ROUGE_L, CIDEr; SPICE commented out); standard Karpathy-split captioning protocol. Neither framework ships Flickr8k itself.",
"note": "Corpus-level metrics; no per-sample correctness. mm-eval copy stores 5 reference captions per image (rows verified on all three splits). Published dataset includes train/validation/test splits; official captioning eval is on the Karpathy test split (1000 images — matches the published test split size) — evaluation runs should use test only."
},
"prompt_template": "<image>Provide a one-sentence caption for the provided image.",
"mapping_from_source": {
"media": {
"from": "image",
"type": "list",
"min_items": 1,
"max_items": 1
},
"id": {
"from": "id"
},
"answer": {
"from": "answer",
"optional": true
},
"source": {
"format": "huggingface",
"url": {
"train": "https://huggingface.co/datasets/jxie/flickr8k",
"validation": "https://huggingface.co/datasets/jxie/flickr8k",
"test": "https://huggingface.co/datasets/jxie/flickr8k"
}
}
},
"prompt_template_source": {
"origin": "official",
"reference": "https://github.com/open-compass/VLMEvalKit/blob/main/vlmeval/dataset/image_caption.py (ImageCaptionDataset canonical caption prompt)",
"notes": "Tier 3: VLMEvalKit ImageCaptionDataset canonical Flickr8k captioning prompt."
}
}
}
}