lighteternal/tissuedecision-data

TissueDecision-9B training data

The training and test data of TissueDecision-9B, an open decision vision-language model (VLM) for rat and mouse tissue pathology. It holds 211,807 questions about 27,496 rat and mouse H&E images, each with its answer, all from openly licensed sources. The model trained on it has the highest score on VIPER, a public expert-reviewed benchmark for the field, across its multiple-choice and true/false questions, and no VIPER image or question was used to build this data.

It spans drug and chemical safety studies (toxicologic pathology), mouse genetics (IMPC phenotyping) and published disease research, across more than 20 organs. Open, labelled images from these species are scarce; this collection brings together what can be shared.

Each item is a multiple-choice, yes/no, true/false-statement or severity question about one image, with its answer, where that answer came from, and the licence of the image. Every row the model could train on is here.

Model · Live demo · LoRA adapter

Full model: merged weights, load and run	tissuedecision-9b
LoRA adapter for Qwen/Qwen3.5-9B	tissuedecision-9b-lora
Training data, splits and provenance	this repo

Summary

Items	211,807: 207,294 decisions (73,493 multiple choice, 131,012 yes/no or true/false, 2,789 severity) and 4,513 free-text descriptions
Images	27,496 H&E images; by species, the items are rat 112,831, mouse 97,291 and not stated 1,685
Sources	PubMed Central open-access figures, NTP Nonneoplastic Lesion Atlas, Open TG-GATEs slides with Bayer expert annotations, IMPC mouse histopathology
Splits	train 190,054 · dev 13,791 · test_internal 6,843 · test 1,119, split by article, atlas page, study or gene
Licence	CC BY-SA 4.0 for the collection; each image keeps its own licence (license, see Licence)
Benchmark contamination	none by construction: VIPER images and text and VIPER's MMO image source were never stored or used

from datasets import load_dataset

items = load_dataset("lighteternal/tissuedecision-data", "items", split="train")          # questions and answers
images = load_dataset("lighteternal/tissuedecision-data", "images", split="images")       # 8.4 GB of JPEGs
img_by_id = {iid: i for i, iid in enumerate(images["image_id"])}
row = items[0]
image = images[img_by_id[row["image_id"]]]["image"]   # PIL image

Composition

Source bucket (source)	Items	Images	How the answer was obtained
vstyle: VIPER-format re-renders	49,006	11,823	labels of the items below, re-rendered in the benchmark's question format
qa: caption-grounded QA, round 1	25,312	5,487	an open LLM wrote questions from figure captions; evidence spans checked in code; no-image filtered
replay: format retention	24,586	11,964	no label; in training the base model's own answer distribution is the target
impc: IMPC mouse histopathology	22,306	3,069	pathologist annotations (MPATH diagnoses, severity)
literature: PMC figure findings	21,923	8,712	findings extracted from figure captions and mapped to SEND terms
probe: over-reading probes	18,317	5,703	"within normal limits" or specific lesion statements, from the image's label
bayer: Bayer expert polygons	15,477	3,749	computed from expert polygons on TG-GATEs rat liver slides
slides: TG-GATEs slide tiles	14,124	6,728	TG-GATEs study metadata: vehicle controls and weak positives
ntp: NTP atlas	12,611	1,995	expert diagnoses of the atlas images
claude: Claude-written items	5,688	640	written by Claude Opus from the image and its expert label, then audited
qa2: verified QA, round 2	2,457	1,711	open-LLM questions kept by an image-grounded verifier

Images are shared between buckets (an atlas image can carry label-derived items, VIPER-format re-renders and Claude-written items), so the image column does not add up to 27,496.

Organ group	Items	Share
liver	56,492	26.7%
kidney	33,138	15.6%
male reproductive organs	15,888	7.5%
lung	14,244	6.7%
GI tract	12,779	6.0%
heart	5,783	2.7%
salivary gland	4,711	2.2%
thyroid	2,423	1.1%
urinary bladder	2,014	1.0%
other organs (brain, bone, eye, skin, spleen, uterus/ovary, pancreas, ...)	61,717	29.1%
organ not stated	2,618	1.2%

VIPER category tag	Items
Pathology identification	40,715
Over-reading probes	26,355
Anatomy	21,535
Characterization	11,277
Localization	3,596
Quantification	2,153
Artifact	744
No VIPER category (presence, organ, severity, replay, descriptions)	105,432

All items on one image share its split. test_internal holds out whole articles, atlas pages and TG-GATEs studies; test holds out IMPC genes (every allele and centre of a gene stays in one split) and a few round-2 QA items. Bayer studies are split by study: the 34 test studies used for the model's held-out slide evaluation are not in this release's training split.

Sources, licences and attribution

Source	Licence	Images	Items	Attribution
PubMed Central Open Access figures	CC BY (11,722 images), CC0 (74)	11,796	69,410	per image: article citation and DOI in the attribution and doi columns of the images
Open TG-GATEs rat liver and kidney slides	CC BY-SA 2.1 JP	10,573 (incl. Bayer tiles)	59,161	"Toxicogenomics Project and Toxicogenomics Informatics Project under CC Attribution-Share Alike 2.1 Japan". Licence and Additional License: licenses/OPEN_TG-GATEs_LICENSE.txt
Bayer expert annotations on 230 Open TG-GATEs rat liver slides	CC BY-SA 2.1 JP (as the slides)	(on the TG-GATEs tiles)	(in the row above)	M. Radbruch, B. Lawrenz, H. Pischon, V. Bertani, U. Bach, F. Schorsch, A. Poehlmann, S. Villalba: "Annotated necrosis and mitotic figures on 230 rat liver slides of the Open TG-GATEs dataset" (Bayer AG, with Aignostics GmbH)
NTP Nonneoplastic Lesion Atlas	US Government work, public domain	2,058	46,742	National Toxicology Program, NIEHS; atlas page URL per image
IMPC histopathology, BioImage Archive S-BIAD3040	CC BY 4.0	3,069	36,494	International Mouse Phenotyping Consortium; Groza et al. 2023, Nucleic Acids Res., doi:10.1093/nar/gkac972

Items take the licence of their image (license column). The image rows carry attribution, ref (source URL, PMC figure or slide reference) and, for articles, doi.

Left out of this release: 23,648 items on 3,177 images whose licence is non-commercial or no-derivatives (CC BY-NC 4,753 items, CC BY-NC-SA 3,108, CC BY-NC-ND 15,787). None of them was used to train TissueDecision-9B, and the no-derivatives terms do not allow sharing crops. The rule and the counts are in provenance/release_filter.json.

How the items were built

1. Label-derived items. Each image's expert or source label (NTP diagnosis, IMPC pathologist annotation, caption findings mapped to SEND terms, TG-GATEs study record) becomes presence questions ("Is necrosis present in this image?"), a primary-finding multiple choice, organ identification, severity grades and "any lesion?" questions.
2. Rendering in VIPER's format (vstyle, probe). VIPER was profiled only through aggregate statistics (option counts and lengths, stem patterns, keyword shares, statement lengths, category, organ and magnification shares, image colour statistics), never through its text. The labelled items were then re-rendered to match:
   • stems that name species, organ and stain;
   • exactly 5 options of the same kind and within ±20% length, drawn from a 2,050-entry phrase dictionary checked against rules that forbid added location, extent, severity or cause;
   • single-assertion true/false statements of 9 to 16 words, worded like VIPER's true/false sets;
   • "image-swap" false statements (a true statement about another image);
   • over-reading probes, with a normal answer in about 45% of them.
3. Bayer programmatic items (bayer). Structure position, structure identity, extent bins, counts, artifacts and low-power fields, with exact answers computed from the expert polygons. Lobular-zone items were dropped after an audit (blind agreement 3 of 12; the zone was not visible in the field).
4. Verified question synthesis (qa2, and re-rendered in vstyle). An open LLM (Qwen3.6-35B-A3B) wrote questions grounded in evidence quoted from captions and atlas narratives. An image-grounded verifier (the same model, multimodal), calibrated against expert-style audits, kept 4,365 items.
5. Round-1 caption questions (qa). The same open LLM wrote caption-grounded questions; every question had to quote a verbatim evidence span, checked in code, and questions about doses, compounds, time points or genotypes (not visible in an image) were rejected.
6. Claude-written items (claude). Claude Opus (Anthropic) viewed each image together with its expert label and wrote VIPER-style items. Pathology claims had to follow the label; anatomy and position claims were allowed only when clearly visible. 640 rat images (NTP atlas pages, and TG-GATEs control tiles of kidney and liver), 5,688 items.
7. Replay (replay). Generic decision items on the same images. They have no gold answer: during training the model is pulled towards the base model's own answer distribution on them, to keep its general behaviour.

Quality control

All audits were done by Claude Opus viewing each image under a strict one-image-at-a-time protocol, with hidden wrong-answer controls mixed in. Every control was caught.

Item set (audit sample)	Judged correct	Action
Bayer programmatic (36)	extent 6/6; field 1/6; zone 3/12	redesigned; re-audit 6/6 on ink and position
IMPC primary findings (53)	66% (male reproductive 12/13)	kept; images with several findings lose the primary-finding item
QA synthesis without the verifier (142)	61%	image-grounded verifier added
QA synthesis with the verifier (54)	87%	kept; IMPC multiple-choice items dropped
VIPER-format re-renders (30)	77%	distractors with overlapping terms excluded
Claude-written items (136)	98.5%	scaled to 640 images

Other filters:

• No-image filter. A text-only model (Qwen3.5-0.8B) answered the items without the image; 11,024 round-1 multiple-choice items that it could answer were dropped (42% of that set). Its flag and score are kept as blind_flag and blind_p where computed.
• Panel check. An image check removed 27,016 items on 4,876 composite, non-H&E or wrong-organ figure panels.
• Image exclusions. Writers' flags removed 5 non-H&E slides and relabelled 14 NTP "testis" images that showed epididymis.

Decontamination

• VIPER's images, questions and answers were never downloaded into this data, stored or trained on. VIPER was used only as aggregate statistics (counts and shares) to align the question format.
• MMO, the Roche multi-organ set that supplies most VIPER images (CC BY-NC), was never used.
• VIPER also draws 63 images from Open TG-GATEs rat liver. Slides whose regions matched VIPER images in a template and perceptual-hash (pHash) check were removed; this removed 3 TG-GATEs test slides.
• Held-out evaluation slides (34 Bayer test studies, the TG-GATEs liver probe slides) are not in the training split.

Columns

items

Column	Meaning
image_id	key into the images config
split	train / dev / test_internal / test
mode	decision (single-pass answer) or describe (free-text description; not used to train TissueDecision-9B)
kind	choice (multiple choice), noul (yes/no or true/false; gold true/false), score (ordered severity levels), text (description)
question	the question, or for statements: stem + Statement: "..." + Is this statement true?
criteria	JSON object of the answer options in display order (choice, score)
gold	the correct option, true/false, or the description
soft_true	soft target for yes/no items when the label is uncertain (e.g. 0.95)
state	the context line shown with the image, e.g. "Rat liver, H&E" (training drops it for 40% of samples, as VIPER prompts have none)
source	source bucket (table above); image_source: where the image comes from
qtype, orig_qtype, style	item template, the template it was re-rendered from, and the stem style
category, viper_category	VIPER-style category tags where assigned
mag	magnification when known (2.5×, 5×, 10×, 20×)
license, label_source	licence of the image; how the answer was obtained (e.g. ntp_atlas_expert, bayer_polygon_area, claude_opus_image_grounded_expert_anchored, base_model for replay)
replay, weight, ctx, blind_flag, blind_p, knowledge	build fields: replay flag, sampling weight, context flag, no-image filter flag and score, and knowledge = kept although answerable from text knowledge (set on some buckets only)

images: image_id, image (JPEG), image_source, source, license, attribution, ref, doi, species, organ, width, height, findings (source diagnoses), normal, meta_json (other source fields, e.g. IMPC gene, sex, zygosity, MPATH terms and the pathologist's description).

Licence

The collection is released under CC BY-SA 4.0. Each row also keeps the licence of its image (license):

• CC BY-SA 2.1 JP (Open TG-GATEs slides and the Bayer annotations): share-alike. Redistribution must carry the attribution string above and the database's Additional License (licenses/OPEN_TG-GATEs_LICENSE.txt), and derivatives must stay under that licence (the CC 2.x terms also allow a later version with the same licence elements).
• CC BY / CC BY 4.0 (PMC figures, IMPC): attribution, given per image.
• CC0 and US Government public domain (NTP): no conditions.

Limitations

• Label noise. Caption-derived labels can inherit a mismatch between a caption and the panel it was assigned to, despite the panel check. Round-1 caption questions were filtered for image dependence but not verified item by item.
• Audits by one model family. The Claude-written items were audited by the same model family, not by a human pathologist.
• Coverage. Rat thyroid, urinary bladder, salivary gland and male accessory glands are scarce in open data (thyroid 1.1% and bladder 1.0% of items). Nearly half of the items are mouse, while VIPER is rat only.
• Replay items have no gold answer and are only meaningful as distillation targets.
• TG-GATEs tiles come from one laboratory, scanner and stain, as do the held-out Bayer and probe tests.

Intended use

Training and evaluating decision models for rat and mouse tissue pathology (safety studies, disease models, mouse phenotyping), studying calibration and over-reading, and building question sets for pathology education. Not for diagnosis or for regulatory decisions about individual studies.

Acknowledgements

Open TG-GATEs (Toxicogenomics Project and Toxicogenomics Informatics Project; NIBIOHN, Japan), the Bayer pathologists who annotated the TG-GATEs liver slides, the National Toxicology Program (NIEHS) for the Nonneoplastic Lesion Atlas, the International Mouse Phenotyping Consortium and the BioImage Archive, the authors of the PubMed Central open-access articles whose figures are used, and the VIPER authors (Mahmood lab) for the benchmark and its documentation.