Stage-3 training mix (181k rows, 53.9M tokens, 836 steps from -v2)
- ~45k brand↔generic MCQA template-generated from the MIT-licensed junioralive/Indian-Medicine-Dataset (254k real Indian medicine records: brand, composition, manufacturer). Both directions: brand→generic composition and generic→brand.
- ~34k MedMCQA pharmacology questions (x2 upsampled) from the Apache-2.0 train split.
- ~4-6k concept↔drug MCQA from ekacare/BODHI-M TREATED_BY triples (CC-BY-NC-4.0 — non-commercial use only applies to this component's source data).
- 120k general MCQA replay + 10k lab-report replay (from eka-clinical-sft-train) so no skill was lost.
Eval-set hygiene: no question text from the eval sets was used — brand knowledge comes from an independent product corpus.
Results (ekacare MCQA battery, greedy letter-parse, full test splits)
Table with columns: Set, base, v1, v2, v3| Set | base | v1 | v2 | v3 |
|---|
| indian_drug_mcqa | 0.380 | 0.397 | 0.398 | 0.523 |
| Eka_NFI_MCQA | 0.428 | 0.455 | 0.458 | 0.480 |
| MedMCQA-Indic en | 0.266 | 0.438 | 0.435 | 0.435 |
| MedMCQA-Indic hi | 0.093 | 0.348 | 0.348 | 0.336 |
| MMLU-medical (8 configs) | 0.118–0.531 | 0.467–0.671 | 0.485–0.665 | 0.436–0.664 |
Drug MCQA +12.6 points over v1/v2 (0.397→0.523) and NFI +2.5 (0.455→0.480) — the knowledge injection worked. General MCQA held flat (en) or within noise (MMLU band). The lab-report skill was verified intact post-stage-3: structured summaries with correct HIGH/LOW flags on a held-out synthetic check.
Remaining gap: drug MCQA 0.523 still has headroom (distractors are hard close-combos), and MedMCQA-Indic hi drifts slightly down (-1.2 pts, parse 0.976→0.945). Next levers: more brand-brand distractor contrast, MedCritical-style CoT distillation (arXiv:2509.23368) for the general-MCQA ceiling.
Limitations
- Template-generated lab summaries; not clinician text. Not a diagnosis engine.
- Brand↔generic knowledge reflects the source catalog (1mg-style retail scrape, MIT); brand names change and dosing details are not modeled.
- No Hindi lab-report training data; Hindi MCQA degradation unmeasured for report inputs.
Training
TRL 1.18 SFTTrainer, lr 8e-6 cosine, 1 epoch, packed 4096, batch 8×2, bf16, flash-attn2, 1×H200, 39m56s. eval_results_v3.json is in this repo.