Sidharth1743/tamil-med-sft-cot-600

tamil-med-sft-600

A small, high-quality supervised fine-tuning (SFT) dataset of 559 Tamil-language medical instruction–reasoning–answer triples. Every item is written in natural Tamil script, shows explicit step-by-step clinical reasoning before the answer, and was independently audited by an LLM judge and decontaminated against public Tamil medical benchmarks.

The goal is to teach a model how a clinician reasons in Tamil — across frontline care, basic-science mechanisms, and evidence appraisal — not just to recall facts.

⚠️ Not medical advice. This is synthetic data for ML research and education only. It must not be used for real clinical decision-making, diagnosis, or treatment.

Dataset summary

Items	559
Language	Tamil (ta); English retained for medical/drug terms, units, acronyms
Format	JSONL, one item per line
Task	Instruction tuning (SFT) for medical reasoning
Generator	Sarvam-105B (best-of-2 candidates per item)
Judge	Grok-4.3 (6-dimension rubric; verified)
Cell accept rate	93.2% (559 / 600 generation cells)
Benchmark contamination	0 (13-gram filtered vs 933 benchmark questions)

Languages

Prose (instruction_ta, reasoning_ta, answer_ta) is in Tamil (Tamil Unicode script). Established English medical/drug terminology, units, lab values, acronyms, and structural control tokens (format codes, safety labels) are deliberately kept in English/Latin script, mirroring real Indian clinical usage.

Dataset structure

Data fields

field	type	description
instruction_ta	string	The question / clinical scenario in Tamil (the prompt).
reasoning_ta	string	Step-by-step clinical reasoning in Tamil (chain-of-thought).
answer_ta	string	The final, concise, actionable answer in Tamil.
format	string	Item type — see format legend below.
subject	string	Medical subject (e.g. Internal Medicine, Pharmacology, Anatomy).
clinical_setting	string	Care setting / context for the scenario.
safety_note	string	referral_advised | within_scope | patient_education | not_applicable.
key_concept	string	Short label of the concept the item teaches.
domain	string	frontline_clinical | academic | clinical_trial.
_meta	object	Provenance: cell_id, cand_id, subject, topic, setting, format, seed, generated_at, hash.
_judge	object	The judge's 6-dimension scores, final_verdict, and (if revised) fix notes.

Format legend: A diagnostic · B management/triage · C patient-explanation · D drug reasoning · E open Q&A · K academic/mechanistic reasoning · T evidence appraisal.

Building SFT pairs

The dataset stores reasoning and answer separately so you control the target format. For chain-of-thought SFT, concatenate them as the completion:

from datasets import load_dataset

ds = load_dataset("<your-username>/tamil-med-sft-600", split="train")

def to_sft(ex):
    prompt = ex["instruction_ta"]
    completion = ex["reasoning_ta"].strip() + "\n\n" + ex["answer_ta"].strip()
    return {"prompt": prompt, "completion": completion}

sft = ds.map(to_sft, remove_columns=ds.column_names)

For answer-only SFT, use answer_ta as the completion. You can drop _meta / _judge before training (they are provenance, not training signal).

Splits

A single train split of 559 items. (No held-out test split — evaluate on the public Tamil medical benchmarks below, which this set is decontaminated against.)

Composition

By domain

domain	items	teaches
frontline_clinical	345	ASHA/PHC management, triage, counseling, drug reasoning (rural realism)
academic	159	basic science → mechanism → clinical correlation
clinical_trial	55	evidence appraisal / research literacy

By format: A 100 · B 98 · K 126 · D 78 · E 75 · T 55 · C 27.

By safety note: referral_advised 222 · not_applicable 213 · within_scope 119 · patient_education 5.

Subjects span 27 areas, deliberately covering gaps in public Tamil medical benchmarks: Internal Medicine, Pharmacology, Community Medicine, Pediatrics, Obstetrics & Gynaecology, Emergency Medicine, Microbiology, Pathology, Surgery, Dermatology, Nutrition & Metabolism, Anatomy, Physiology, Medical Genetics, Geriatrics, Biochemistry, Psychiatry, Biostatistics, Study Design & Epidemiology, and more.

Dataset creation

Pipeline

1. Sampling. A deterministic, seeded sampler expands a medical taxonomy (domain → subject → topic → setting → format) into generation "cells".
2. Generation (best-of-N). Sarvam-105B writes 2 candidates per cell (candidate 1 at low temperature for accuracy, candidate 2 at higher temperature for diversity), guided by domain-specific accuracy guardrails. Outputs are pre-validated for schema, Tamil-script ratio, and reasoning length.
3. Judging. Grok-4.3 scores every candidate on 6 dimensions: content accuracy, reasoning soundness, safety/scope, Tamil quality, answer usability, instruction realism → ACCEPT / REVISE / REJECT.
4. Best-of-N selection + salvage. The judge's best candidate per cell is kept. Salvageable REVISE items are corrected by Grok (the more knowledgeable model) from the judge's fix instructions and independently re-judged before promotion.
5. Decontamination. Any item with 13-gram overlap against the public benchmarks is dropped (0 dropped here).

Quality (judge scores on the accepted set; 1–5)

C1 content	C2 reasoning	C3 safety	C4 Tamil	C5 usability	C6 realism
4.74	4.93	4.93	4.60	4.89	4.99

Median is 5 on every dimension. Best-of-N + Grok-authored revision raised the per-cell accept rate from ~61% to 93.2% and content accuracy (C1) from 3.79 to 4.74, without weakening the judge.

Source data

Fully synthetic — generated by Sarvam-105B and audited by Grok-4.3. No patient data or copyrighted clinical text was ingested. The taxonomy and prompts encode Indian frontline-care realism (PHC/CHC settings, common presentations).

Decontamination

Filtered (13-gram overlap) against 933 questions from the public Tamil medical benchmarks IndicMMLU-Pro (Health) and TamilMedicalMILU, so the set is safe to fine-tune on before evaluating against those benchmarks.

Considerations for use

Intended use

• Supervised fine-tuning of Tamil/multilingual LLMs for medical reasoning and instruction following; research on low-resource clinical NLP and chain-of-thought.

Out-of-scope / limitations

• Not for clinical use. Synthetic, educational data — not a source of medical truth.
• Generator knowledge ceiling. Content reflects Sarvam-105B's medical knowledge (benchmark baseline ~0.66 on Tamil medical MCQs). Despite judging + Grok-authored fixes, residual factual errors are possible — verify before any downstream reliance.
• Distillation. Corrections were authored/audited by Grok-4.3; the data partially reflects that model's knowledge and style.
• Coverage. 559 items skew toward frontline clinical scenarios and the listed subjects; rare diseases and many specialties are underrepresented.
• Single language register. Indian Tamil clinical register with English code-mixing; may not transfer to other Tamil dialects/registers.

Biases

Topic mix, settings, and phrasing follow the curated taxonomy and the generator's tendencies; it is not a representative sample of all medical practice.

License

Released under MIT.

Citation

@misc{tamil_med_sft_600,
  title  = {tamil-med-sft-600: A Tamil Medical Reasoning SFT Dataset},
  author = {Sidharthan S},
  year   = {2026},
  note   = {Synthetic SFT data generated with Sarvam-105B and audited with Grok-4.3;
            best-of-N selection, judge-verified, benchmark-decontaminated.}
}