Datasets:

lighteternal
/

biodecision-sft-v2.2

Parquet error: Scan size limit exceeded: attempted to read 944891676 bytes, limit is 300000000 bytes Make sure that 1. the Parquet files contain a page index to enable random access without loading entire row groups2. otherwise use smaller row-group sizes when serializing the Parquet files

Error code:   TooBigContentError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

BioDecision SFT v2.2

1.1M biomedical decisions in one format: a source text, a question, lettered options, one correct letter. Built to train BioDecision-4B with Together AI's Tev1 recipe; usable with any classifier or LLM that scores options.

Split	Rows	Tokens	Use
train	1,080,373	420,677,060	training
dev	12,235	4,768,045	model selection
calibration	12,267	4,806,484	temperature fitting only
benchmark	46,199	21,256,443	official test/dev sets of 28 benchmarks; never train on these

Format

tev1 config: ready-to-train rows. prompt is the full chat-templated prompt (Qwen3.5, non-thinking); completion is the answer letter plus <|im_end|>. The user message is JSON:

{"state": "Concomitant administration of ketoconazole increased the plasma concentrations of simvastatin more than tenfold.",
 "question": "What does this text say about the interaction between the drugs \"ketoconazole\" and \"simvastatin\"?",
 "options": [{"label": "A", "key": "effect", "description": "A pharmacodynamic effect or clinical consequence of combining them."},
             {"label": "B", "key": "mechanism", "description": "A pharmacokinetic mechanism: one drug changes the absorption, distribution, metabolism or excretion of the other."},
             {"label": "C", "key": "no_interaction", "description": "The text describes no interaction between these two drugs."}, ...]}

Columns: id, prompt, completion, answer_letter, answer_key, option_keys, source, domain, kind (choice, noul, score), label_origin, commercial_ok, n_tokens. records holds the un-rendered decisions (state, question, options, answer, provenance, licence) for re-rendering in other formats.

from datasets import load_dataset
ds = load_dataset("lighteternal/biodecision-sft-v2.2", "tev1")
ok = ds["train"].filter(lambda r: r["commercial_ok"])  # permissive sources only

Sources

Source	Task	Licence	Train rows	Benchmark rows
pubmed_rct	sentence role in abstract	unknown	190,610	1,017
medmcqa	exam MCQ	Apache-2.0	153,251	4,162
ultramedical_textbookqa	textbook MCQ (synthetic)	MIT	90,782	0
pubmedqa_artificial	abstract → yes/no	MIT	74,469	0
trec_ct	patient vs trial eligibility	CC BY-SA 4.0 (qrels)	65,597	2,999
mcq_noul	is this MCQ answer correct? (derived)	as parent MCQ	59,643	0
drug_reviews_rating	rating from review	CC BY 4.0	55,913	583
drug_reviews_condition	condition from review	CC BY 4.0	52,969	575
ultramedical_medqa_evol	exam MCQ (synthetic)	MIT	51,072	0
tev1_general	general decisions (replay)	mixed (Together Tev1)	36,867	0
trialbench_trial_duration	trial trial duration from design	unknown (TrialBench)	33,558	1,500
trialbench_patient_dropout	trial patient dropout from design	unknown (TrialBench)	29,924	1,500
trialbench_trial_approval	trial trial approval from design	unknown (TrialBench)	23,877	1,500
drugprot	chemical–protein relation	CC BY 4.0	20,906	2,700
trialbench_trial_failure_reason	trial trial failure reason from design	unknown (TrialBench)	16,033	1,500
trialbench_adverse_event	trial adverse event from design	unknown (TrialBench)	13,718	1,500
trialbench_mortality_rate	trial mortality rate from design	unknown (TrialBench)	13,718	1,500
medquad	does the answer fit the question	CC BY 4.0	13,680	0
sciq	science MCQ	CC BY-NC 3.0	11,575	991
evidence_inference	trial result direction	MIT	10,018	1,216
medqa	exam MCQ	CC BY 4.0	9,511	1,273
mediqa_rqe	question entailment	unknown	8,563	230
ade_corpus	adverse drug event (yes/no)	unknown	7,928	0
medical_abstracts	disease class of abstract	CC BY-SA 3.0	6,026	2,888
pubhealth	health claim veracity	MIT	5,989	747
ddi_corpus	drug–drug interaction type	CC BY-NC 4.0	5,372	1,528
headqa	exam MCQ	MIT	5,114	2,663
nli4ct	statement vs trial report	CC BY-SA 4.0	4,365	5,500
gad	gene–disease association	CC BY 4.0	4,068	534
medical_question_pairs	same medical question?	unknown	2,768	0
scifact	claim vs evidence	CC BY-NC 2.0	1,562	339
pubmedqa_labeled	abstract → yes/no/maybe	MIT	927	500
mmlu_medical	evaluation only	see source	0	1,868
medxpertqa_text	evaluation only	see source	0	2,448
lab_bench	evaluation only	see source	0	1,424
trialgpt_criteria	evaluation only	see source	0	1,014

Train rows include replicas: HEAD-QA ×2, SciFact ×2, NLI4CT ×3 and labelled PubMedQA ×3, each replica with a new option order.

Construction

Text repair. 20,182 MedMCQA questions had a systematic deletion of the letters "rt" ("hea disease"); repaired against a vocabulary of 77,966 words.
Filters. Dropped image-dependent questions (1,101), multi-select questions, malformed options, and MedMCQA items whose explanation contradicts the key (99). PUBHEALTH claims that leak their verdict were removed (112).
Duplicates. Exact and MinHash near-duplicates were clustered (32,841 verified pairs). Removed: 24,342 duplicate exam questions, 2,913 exam questions whose duplicates disagree on the answer, and label conflicts in other sources.
Decontamination. Training rows matching any benchmark row were removed: exact content, question-and-options, 13-gram overlap and near-duplicate clusters (1,719 rows). NLI4CT, SciFact, TREC CT and TrialBench reuse the same trial reports or abstracts across their official splits; for them, matching uses the question, and the official splits are kept.
Option shuffling. Options are shuffled per row with a seeded order; keys stay semantic (not_included, not B).
Derived rows. 59,643 yes/no rows ask whether a given MCQ answer is correct; 35,132 rows add a "none of the listed" option, half of them with the correct answer removed.
Clipping. Long evidence is clipped to keep the answer-bearing text: a ±3-sentence window for PubMed RCT, 350 words of eligibility criteria and 80 of summary for TREC CT, 100 and 150 words for TrialBench. Every training row fits in 3,584 tokens.
Reproducibility. Every source is pinned to a Hub revision; builder hashes and all counts are in report.json.

Licences

Licences are recorded per row. commercial_ok is false for non-commercial sources (DDI, SciFact, SciQ) and for sources with no stated licence (PubMed RCT, ADE, MEDIQA-RQE, medical question pairs, TrialBench, the Together mix). Filter on it before any commercial use. Benchmarks keep their original licences.