Datasets:

erow
/

curated_seeing

data_source string	agent_name string	prompt list	images list	question string	ground_truth string	pairs list	reward_model dict	extra_info dict	env_kwargs dict
fla_medvqa	fla_agent	[{"role":"system","content":"You are a medical visual question answering assistant.\nFirst emit zero(...TRUNCATED)	[{"bytes":"iVBORw0KGgoAAAANSUhEUgAAAlgAAAD7CAIAAAA5E33iAAAA42lDQ1BJQ0MgUHJvZmlsZQAAeJxjYGCc4eji5Mokw(...TRUNCATED)	What is the name of the surface shown in the micrographs?	Endothelial surface	[{"concept":"Endothelial surface in panel A demonstrating confluent cellular pavement","region":[0.0(...TRUNCATED)	{ "style": "rule", "ground_truth": "Endothelial surface" }	{"split":"train","index":0,"id":"medix-rl-data_16535","question":"What is the name of the surface sh(...TRUNCATED)	{ "tools_kwargs": { "ground_truth": "Endothelial surface" } }
fla_medvqa	fla_agent	[{"role":"system","content":"You are a medical visual question answering assistant.\nFirst emit zero(...TRUNCATED)	[{"bytes":"iVBORw0KGgoAAAANSUhEUgAAAxgAAAIQCAIAAACNO1/qAAEAAElEQVR4nMT967MkN5IviP3cgYjM86gii8Vns19zd(...TRUNCATED)	What medical condition is currently present?	The currently present medical condition is petechiae.	[{"concept":"Scattered petechiae upper right","region":[0.55,0.12,0.95,0.42]},{"concept":"Dense pete(...TRUNCATED)	{ "style": "rule", "ground_truth": "The currently present medical condition is petechiae." }	{"split":"train","index":2,"id":"medix-rl-data_4270","question":"What medical condition is currently(...TRUNCATED)	{ "tools_kwargs": { "ground_truth": "The currently present medical condition is petechiae." } }
fla_medvqa	fla_agent	[{"role":"system","content":"You are a medical visual question answering assistant.\nFirst emit zero(...TRUNCATED)	[{"bytes":"iVBORw0KGgoAAAANSUhEUgAAAfYAAADGCAIAAACvhDXyAAEAAElEQVR4nIz9SYylW3bWje/9NqeJ9kSfmTdv5b1V1(...TRUNCATED)	What body part did the angiography examination examine?	The spleen.	[{"concept":"Splenic artery and hilar branches in early arterial phase","region":[0.05,0.4,0.48,0.98(...TRUNCATED)	{ "style": "rule", "ground_truth": "The spleen." }	{"split":"train","index":4,"id":"medix-rl-data_32573","question":"What body part did the angiography(...TRUNCATED)	{ "tools_kwargs": { "ground_truth": "The spleen." } }
fla_medvqa	fla_agent	[{"role":"system","content":"You are a medical visual question answering assistant.\nFirst emit zero(...TRUNCATED)	[{"bytes":"iVBORw0KGgoAAAANSUhEUgAAAcUAAAijCAIAAABHltxHAAEAAElEQVR4nOz9aZpc53Xlj66133MiMgEQIEiqs10uF(...TRUNCATED)	What does the green stain represent?	Regeneration marker	[{"concept":"dMHC-positive regenerating fiber in panel (a)","region":[0.52,0.09,0.88,0.19]},{"concep(...TRUNCATED)	{ "style": "rule", "ground_truth": "Regeneration marker" }	{"split":"train","index":5,"id":"medix-rl-data_9526","question":"What does the green stain represent(...TRUNCATED)	{ "tools_kwargs": { "ground_truth": "Regeneration marker" } }
fla_medvqa	fla_agent	[{"role":"system","content":"You are a medical visual question answering assistant.\nFirst emit zero(...TRUNCATED)	[{"bytes":"iVBORw0KGgoAAAANSUhEUgAAAxAAAAImCAIAAABMz/uuAAEAAElEQVR4nET9244kzZKlia0lIqpm7h4RmfnvQ3V19(...TRUNCATED)	What is the name of the type of cancer shown in the image?	Breast Cancer	[{"concept":"Infiltrating malignant epithelial nests","region":[0.03,0.04,0.36,0.38]},{"concept":"In(...TRUNCATED)	{ "style": "rule", "ground_truth": "Breast Cancer" }	{"split":"train","index":6,"id":"medix-rl-data_27065","question":"What is the name of the type of ca(...TRUNCATED)	{ "tools_kwargs": { "ground_truth": "Breast Cancer" } }
fla_medvqa	fla_agent	[{"role":"system","content":"You are a medical visual question answering assistant.\nFirst emit zero(...TRUNCATED)	[{"bytes":"iVBORw0KGgoAAAANSUhEUgAAAbcAAAGFCAIAAADxRyoaAAEAAElEQVR4nIz9W3MdOZKuCTsQQYmSsqq79+4Zs21jN(...TRUNCATED)	What imaging modality is used in this diagnostic study?	X-ray	[{"concept":"L marker typical of plain film radiography","region":[0.82,0.04,0.87,0.09]},{"concept":(...TRUNCATED)	{ "style": "rule", "ground_truth": "X-ray" }	{"split":"train","index":9,"id":"medix-rl-data_47271","question":"What imaging modality is used in t(...TRUNCATED)	{ "tools_kwargs": { "ground_truth": "X-ray" } }
fla_medvqa	fla_agent	[{"role":"system","content":"You are a medical visual question answering assistant.\nFirst emit zero(...TRUNCATED)	[{"bytes":"iVBORw0KGgoAAAANSUhEUgAAAxgAAAIQCAIAAACNO1/qAAEAAElEQVR4nHz9y5Iky7IshqmauUdkVvV67LXPATCCk(...TRUNCATED)	"Is the image provided a good but not the best example of high new bone formation involving osteobla(...TRUNCATED)	"Yes, the image is a good but not the best example of high new bone formation with osteoblasts and o(...TRUNCATED)	[{"concept":"Osteoclasts in left resorption pit","region":[0.13,0.14,0.4,0.58]},{"concept":"Osteocla(...TRUNCATED)	{"style":"rule","ground_truth":"Yes, the image is a good but not the best example of high new bone f(...TRUNCATED)	{"split":"train","index":11,"id":"medix-rl-data_32959","question":"Is the image provided a good but (...TRUNCATED)	{"tools_kwargs":{"ground_truth":"Yes, the image is a good but not the best example of high new bone (...TRUNCATED)
fla_medvqa	fla_agent	[{"role":"system","content":"You are a medical visual question answering assistant.\nFirst emit zero(...TRUNCATED)	[{"bytes":"iVBORw0KGgoAAAANSUhEUgAAAlgAAAHDCAIAAAB3TfOWAAEAAElEQVR4nHz9149tSZbmiZncWh0tXYur7w2RkbKys(...TRUNCATED)	What is the structure being examined in the image?	Appendix tissue	[{"concept":"Mucosal crypts within the lumen","region":[0.2,0.1,0.5,0.4]},{"concept":"Prominent lymp(...TRUNCATED)	{ "style": "rule", "ground_truth": "Appendix tissue" }	{"split":"train","index":14,"id":"medix-rl-data_8016","question":"What is the structure being examin(...TRUNCATED)	{ "tools_kwargs": { "ground_truth": "Appendix tissue" } }
fla_medvqa	fla_agent	[{"role":"system","content":"You are a medical visual question answering assistant.\nFirst emit zero(...TRUNCATED)	[{"bytes":"iVBORw0KGgoAAAANSUhEUgAAAxgAAAIQCAIAAACNO1/qAAEAAElEQVR4nLT925IkSZIliB2+iIiqmblHRGZWdTeqG(...TRUNCATED)	Is eosinophilic adenoma present in the sample?	Yes, eosinophilic adenoma is present.	[{"concept":"Eosinophilic adenoma cells with abundant pink granular cytoplasm","region":[0.1,0.08,0.(...TRUNCATED)	{ "style": "rule", "ground_truth": "Yes, eosinophilic adenoma is present." }	{"split":"train","index":15,"id":"medix-rl-data_30895","question":"Is eosinophilic adenoma present i(...TRUNCATED)	{ "tools_kwargs": { "ground_truth": "Yes, eosinophilic adenoma is present." } }
fla_medvqa	fla_agent	[{"role":"system","content":"You are a medical visual question answering assistant.\nFirst emit zero(...TRUNCATED)	[{"bytes":"iVBORw0KGgoAAAANSUhEUgAAAvQAAAI4CAIAAACKlOqOAAEAAElEQVR4nOz92ZbjSJIsioqoGkC6x5RTVQ/nPtyH/(...TRUNCATED)	Where is the irregular lesion located?	Left vocal cord	[ { "concept": "Left vocal cord with irregular whitish-yellow lesion", "region": [ 0.25, 0.38, 0.5, 0.98 ] } ]	{ "style": "rule", "ground_truth": "Left vocal cord" }	{"split":"train","index":16,"id":"medix-rl-data_34068","question":"Where is the irregular lesion loc(...TRUNCATED)	{ "tools_kwargs": { "ground_truth": "Left vocal cord" } }

End of preview.

curated_seeing

Quality-filtered medical VQA + visual grounding split for localize-then-answer training.

Teacher ROIs come from erow/Medical-ROIs-K2.6 (moonshotai/Kimi-K2.6). Upstream images and QA come from MBZUAI/medix-rl-data. This release keeps only rows whose teacher-text audit scores quality = 1.0, and keeps the original image-disjoint train / validation / test membership.

How the data is curated

MBZUAI/medix-rl-data
        │
        │  problem / solution / image / source / id
        ▼
Teacher: moonshotai/Kimi-K2.6
        │  ROI evidence + reasoning
        ▼
seeing/full  (image-disjoint train / validation / test)
        │
        │  LLM rubric audit of teacher text
        │  keep quality == 1.0
        ▼
curated_seeing  (this repo)

Quality is the mean of five teacher-text checks (the judge does not see pixels):

rubric	pass when
`support`	region titles are evidence for the gold answer
`no_hallucination`	no extra findings beyond what the gold answer needs
`non_circular`	reasoning cites visual cues, not only the gold answer
`gold_conflict`	inverted: reasoning does not reject the gold answer
`title_specificity`	titles name a concrete structure or finding

Splits

split	kept	source pool	keep rate
`train`	14,645	21,940	0.6675
`validation`	1,740	2,572	0.6765
`test`	1,770	2,622	0.6751
total	18,155	27,134	0.6691

Source mix

source	train	validation	test
`RadGenome/PMC-VQA`	7,228	913	893
`flaviagiammarino/path-vqa`	5,310	610	620
`BoKelvin/SLAKE`	1,528	144	183
`flaviagiammarino/vqa-rad`	579	73	74

Schema

Agent-R1 / FLA parquet rows. Images are stored as bytes inside each split file.

field	description
`data_source`	`fla_medvqa`
`agent_name`	`fla_agent`
`prompt`	system + user chat turns (`<evidence>` then `<answer>`)
`images`	one-image list (`{"bytes": ...}`)
`question` / `ground_truth`	clinical question and gold answer
`pairs`	teacher evidence `{concept, region}` list; `region` is `[x1, y1, x2, y2]` in `[0, 1]`
`reward_model`	`{style, ground_truth}`
`extra_info.id`	sample id (matches `medix-rl-data`, e.g. `medix-rl-data_16535`)
`extra_info.source`	upstream VQA dataset
`extra_info.gt_evidence`	same list as `pairs` (FLA training field)
`extra_info.has_gt_box`	whether teacher boxes are present
`extra_info.teacher_model`	`moonshotai/Kimi-K2.6`
`extra_info.image_group`	hash used to keep splits image-disjoint
`env_kwargs`	tool kwargs with the gold answer

Load

from datasets import load_dataset

ds = load_dataset("erow/curated_seeing")
row = ds["train"][0]
print(row["question"], row["ground_truth"])
print(row["pairs"])
row["images"][0]  # PIL.Image

Intended use

Supervised or RL training of medical VLMs that first emit answer-supporting regions, then an answer. Complementary to the unfiltered teacher-ROI SFT set erow/Medical-ROIs-K2.6.

Limitations

Quality filtering is a text-only audit of teacher reasoning and titles; boxes are not human-verified on every sample.
Coverage is a subset of seeing/full (~67%).
Not a clinical decision-support product; research use only.

Citation / provenance

Upstream pool: MBZUAI/medix-rl-data and its constituent VQA datasets (PMC-VQA, PathVQA, SLAKE, VQA-RAD).
Teacher ROIs: erow/Medical-ROIs-K2.6 (moonshotai/Kimi-K2.6).
Filter: keep rows whose Kimi teacher-text quality is 1.0; split membership is unchanged.

Downloads last month: 75

Total file size:

7.33 GB