lighteternal/biodecision-v2-data

BioDecision-4B v2: training data

The exact data BioDecision-4B v2 was trained and tested on: typed biomedical decisions (Choice, Yes/No, Score) in the Tev1 / Jev prompt format. The training pool (train) has 1,750,239 rows; one shuffled epoch processed 1,750,208 of them (the trainer drops the final incomplete batch of 31).

BioDecision family

	BioDecision v2 (this release)	BioDecision v1 (previous)
Full model: merged weights, load and run	biodecision-v2-4b	biodecision-tev1-4b
LoRA adapter for Qwen/Qwen3.5-4B	biodecision-v2-4b-lora	BioDecision v1 LoRA adapter
Training data, test sets and overlap masks	biodecision-v2-data	BioDecision v1 training data + part 2
Live demo	biodecision-demo

What changed from BioDecision v1

• Training pool: 1,750,239 decisions, 766,914,222 tokens (BioDecision v1: 1,100,763 and 433,417,407).
• 22 new sources: VAERS, FDA MAUDE, OnSIDES, PHEE (drug and device safety); FDA recalls, FDA warning letters, drug-label sections (regulatory); CDISC COSMoS, MedConceptsQA (standards and coding); ClinicalTrials.gov, TrialPanorama, Chia (trial design); CIViC, ClinVar, Open Targets (genomics); PrimeKG; DDXPlus; BioNLI, BC5CDR; MedInfo; CMExam, CMB (Chinese exams). BioDecision v1's second-stage sources (CT Open training questions, MedHallu, rule-built grounding checks) are counted with v1.
• BioDecision v1's data re-screened: detected matches to benchmark group IDs, NCT identifiers and normalised complete evidence states were removed during construction; further detected overlaps are excluded through the evaluation masks. DDI and DrugProt were rebuilt with every candidate pair; TrialBench label priors were capped.
• Robustness variants across existing and newly added sources (prompt injection inside the source, number and unit flips, long documents, not-enough-information cases), with labels derived by deterministic rules.

Tasks

Use case	BioDecision v1 decisions	BioDecision v2 decisions	share	sources
Medical exams (English)	321,305	325,451	18.6%	headqa, medmcqa, medqa, sciq, ultramedical_medqa_evol, ultramedical_textbookqa
Evidence QA over papers	282,050	290,954	16.6%	evidence_inference, medical_abstracts, pubmed_rct, pubmedqa_artificial, pubmedqa_labeled
Trial design and protocol screening	0	132,217	7.6%	chia (new), ctgov_usdm (new), trialpanorama (new)
Trial outcome prediction	139,817	114,375	6.5%	ct_open_train, trialbench_adverse_event, trialbench_mortality_rate, trialbench_patient_dropout, trialbench_trial_approval, trialbench_trial_duration, trialbench_trial_failure_reason
Patient drug reviews	108,882	112,868	6.4%	drug_reviews_condition, drug_reviews_rating
Drug and device safety	7,928	109,136	6.2%	ade_corpus, maude (new), onsides (new), phee (new), vaers (new)
Drug, protein and chemical relations	26,278	99,487	5.7%	bc5cdr (new), ddi_corpus, drugprot
Genomics and precision oncology	4,068	79,414	4.5%	civic (new), clinvar (new), gad, opentargets (new)
Judging answers against a source	71,044	73,417	4.2%	grounding_synth, mcq_noul, medhallu_artificial
Patient–trial matching	65,597	64,187	3.7%	trec_ct
Data standards and coding (CDISC, ICD, ATC)	0	60,802	3.5%	cosmos (new), medconceptsqa (new)
Diagnosis from symptoms	0	54,257	3.1%	ddxplus (new)
Regulatory and quality	0	53,279	3.0%	fda_enforcement (new), fda_warning_letters (new), spl_sections (new)
Biomedical knowledge graph	0	47,100	2.7%	primekg (new)
General decisions (Tev1 replay)	36,867	43,133	2.5%	tev1_general
Medical exams (Chinese)	0	37,353	2.1%	cmb (new), cmexam (new)
Claim checking and medical NLI	23,247	36,978	2.1%	bionli (new), medical_question_pairs, mediqa_rqe, nli4ct, pubhealth, scifact
Medical information questions	13,680	15,831	0.9%	medinfo_qtype (new), medquad
Total	1,100,763	1,750,239	100%	57 sources

Splits

Split	Decisions	Use
train	1,750,239	training (one epoch)
dev	23,847	model selection
calibration	21,407	temperature and confidence-gate fitting, after overlap screening
benchmark	46,199	processed rows from the official test/dev splits of the 28 original benchmarks
benchmark_new	21,798	held-out test rows of the 22 sources added in BioDecision v2

Overlap masks

masks/<split>.json records the rows flagged by the conservative union of two overlap-screening scans, with the reasons: group (shared non-empty group ID), input (same normalised question, evidence and options, in any order), nct (a shared NCT number), or neardup:<train id> (at least half of a sampled sketch of its distinctive 13-word phrases occurs in one training row); scan1 / scan2 record which scan flagged it. Unflagged rows may still share evidence with training. Reported results exclude flagged rows: 1,958 of 46,199 benchmark rows, 2,607 of 21,798 added-source rows, 2,666 calibration rows and 3,206 dev rows. The script is decontam.py.

Where labels come from

label_origin is recorded for 1,729,863 training rows; 20,376 second-stage rows from BioDecision v1 have no value, shown below as unset:

label_origin	train rows	meaning
gold	1,235,373	the source's own label
derived	182,927	follows deterministically from a source label by a rule: e.g. the yes/no form of an MCQ, and the robustness variants (prompt injection, number and unit flips, long documents)
synthetic_llm	132,929	LLM-written exam questions and answers from UltraMedical (85,190 textbook + 47,739 MedQA-evol)
gold_heuristic	109,065	a source's heuristic or machine-extracted label: PubMedQA's artificial labels (74,447), DDXPlus's simulated differentials (33,114), OnSIDES's model-extracted adverse events (1,504)
rule	49,235	built by a rule with a known answer
unset	20,376	no value recorded (BioDecision v1 second-stage rows: 8,979 CT Open, 7,712 MedHallu with LLM-written candidate answers, 3,685 rule-built grounding checks); labels from the source or a rule
augmented	20,334	earlier augmentations: 17,606 exam 'none of the listed' variants, 1,559 DrugProt other-relation variants, 1,169 Evidence Inference evidence substitutions

No Jev predictions were used as training labels. The pool does include upstream LLM-generated material: UltraMedical's questions and answers (written with GPT-4) and MedHallu's hallucinated answers.

Sources and licences

Research use only: several sources are non-commercial. commercial_ok is recorded where available; 20,376 training rows have no value, which does not establish commercial permission. Respect each source's licence; benchmark test sets belong to their authors.

Source	Train decisions	Licence	New in BioDecision v2
pubmed_rct	195,000	unknown
medmcqa	156,175	Apache-2.0
ultramedical_textbookqa	90,782	MIT
pubmedqa_artificial	76,467	MIT
trec_ct	64,187	CC BY-SA 4.0 (qrels)
mcq_noul	62,020	as parent MCQ
trialpanorama	61,373	Apache-2.0 (TrialPanorama); ClinicalTrials.gov public domain; PubMed abstracts (publisher copyright)	yes
drug_reviews_rating	57,875	CC BY 4.0
drug_reviews_condition	54,993	CC BY 4.0
ddxplus	54,257	CC-BY-4.0 (DDXPlus, Mila / Dialogue Health Technologies)	yes
drugprot	53,560	CC BY 4.0
ctgov_usdm	53,360	ClinicalTrials.gov (public domain, U.S. NLM)	yes
ultramedical_medqa_evol	51,072	MIT
primekg	47,100	CC0-1.0 (PrimeKG); upstream DrugCentral CC BY-SA 4.0; upstream HPO annotations (HPO licence, attribution); upstream DisGeNET CC BY-NC-SA 4.0	yes
tev1_general	43,133	mixed (Together Tev1)
medconceptsqa	38,919	Apache-2.0 (MedConceptsQA); ICD code sets are U.S. government works (public domain); ATC (c) WHOCC Oslo, no commercial copying or distribution	yes
vaers	37,901	Public domain (US Gov, CDC/FDA VAERS); via HF mirrors vaers-outcomes (Apache-2.0) and vaers-narrative-generation (CC0-1.0); MedDRA terms	yes
maude	35,786	CC0-1.0 (openFDA MAUDE, public domain)	yes
ddi_corpus	35,202	CC BY-NC 4.0
trialbench_trial_duration	30,623	unknown (TrialBench)
fda_enforcement	29,308	CC0-1.0 (openFDA, public domain)	yes
opentargets	28,383	CC0-1.0	yes
clinvar	23,668	ClinVar (NCBI): attribution requested; submitter-authored text may carry submitter rights; ClinVar (NCBI): US-government public domain facts; attribution requested	yes
civic	22,214	CC0-1.0	yes
trialbench_trial_approval	22,051	unknown (TrialBench)
cosmos	21,883	CDISC COSMoS (permission confirmed by CDISC)	yes
onsides	19,248	CC-BY-4.0 (OnSIDES v3.1.1; MedDRA term names)	yes
cmexam	19,120	Apache-2.0 (github.com/williamliujl/CMExam LICENSE) + README: academic/research use only	yes
spl_sections	18,980	CC0-1.0 (openFDA drug labels, public domain)	yes
cmb	18,233	Apache-2.0 (github.com/FreedomIntelligence/CMB LICENSE)	yes
chia	17,484	CC-BY-4.0 (Chia, Kury et al. 2020)	yes
trialbench_trial_failure_reason	15,010	unknown (TrialBench)
medquad	14,931	CC BY 4.0
trialbench_patient_dropout	14,799	unknown (TrialBench)
bionli	12,844	CC BY 4.0 (BioNLI README, github.com/StonyBrookNLP/BioNLI)	yes
sciq	12,765	CC BY-NC 3.0
trialbench_mortality_rate	12,179	unknown (TrialBench)
evidence_inference	11,243	MIT
trialbench_adverse_event	10,734	unknown (TrialBench)
bc5cdr	10,725	Public domain (NCBI/NLM US Government Work notice in CDR_Data/README.txt)	yes
mediqa_rqe	9,730	unknown
medqa	9,571	CC BY 4.0
ade_corpus	9,073	unknown
ct_open_train	8,979	see CT Open (BioDecision v1 training data, part 2)
medhallu_artificial	7,712	see MedHallu (BioDecision v1 training data, part 2)
medical_abstracts	7,160	CC BY-SA 3.0
phee	7,128	MIT (PHEE, github.com/ZhaoyueSun/PHEE)	yes
pubhealth	6,313	MIT
gad	5,149	CC BY 4.0
headqa	5,086	MIT
fda_warning_letters	4,991	Public domain (FDA warning letters, fda.gov) via HF mirror tim-a-wood/fda-warning-letters (card: other); eCFR public domain	yes
medical_question_pairs	3,822	unknown
grounding_synth	3,685	derived from licensed rows (BioDecision v1 training data, part 2)
nli4ct	2,469	CC BY-SA 4.0
scifact	1,800	CC BY-NC 2.0
pubmedqa_labeled	1,084	MIT
medinfo_qtype	900	CC-BY-4.0 (MedicationQA, Ben Abacha et al. 2019)	yes