vnytht/cancer-screening-evidence-reasoner

Cancer Screening Evidence Reasoner (AutoScientist Challenge)

Fine-tuning dataset for teaching a language model to answer cancer screening eligibility and evidence questions with exact, verifiable citations — not hedged guesses.

Motivation

Base models know screening guidelines roughly but invent citations and get exact statistics wrong. Every completion in this dataset is computed by a rule engine from verified USPSTF and SEER ground truth — not LLM-generated.

Format

Files: train.jsonl (recommended) or train.csv — 3,714 rows.

Row mix

• —~76% medical: screening eligibility, red flags, BRCA/Lynch, grade-D refusals, version traps
• —~24% general instruction buffer (Databricks Dolly-15k, CC-BY-SA-3.0) to prevent catastrophic forgetting

Sources (all public domain or open license)

Usage with Adaption Adaptive Data

column_mapping = {"prompt": "prompt", "completion": "completion"}

Import URL: https://huggingface.co/datasets/<your-username>/cancer-screening-evidence-reasoner

File to import: train.jsonl

Limitations

• —Population-level screening guidance only — not medical advice
• —US-focused (USPSTF / SEER)
• —Does not diagnose; includes refusal rows for diagnosis requests

Citation

@dataset{cancer_screening_evidence_reasoner_2026,
  author = {Vinay Thorat},
  title = {Cancer Screening Evidence Reasoner},
  year = {2026},
  publisher = {Hugging Face},
  note = {AutoScientist Challenge Part 2, Science category}
}