← TinyGPT · docs · devlog · roadmap · speedup
source: docs/prds/B8-multilingual-specialist.md · view on GitHub ↗

PRD — Multilingual specialist on top of Sarvam-Edge / Airavata base

Goal

Train a multilingual SFT/LoRA specialist on a Sarvam-Edge or Airavata base (both are Indic-LLM bases optimized for English + 10+ Indian languages) for a Mac-runnable Indic-capable agent. Ship gate: beats the 0-shot base on the MILU eval (tinygpt eval-indic) by ≥ 3pp average across at least 3 Indian languages.

Different from A1/B1: A1/B1 are domain specialists; B8 is a language-coverage specialist. Same recipe shape; different acceptance axis.

Why now

Scope — in

Scope — out

Files to touch

FileChange
scripts/recipes/b8-indic.shnew — recipe
Sources/TinyGPT/EvalIndic.swiftalready exists; add per-language breakdown if missing
docs/specialists/b8-indic.mdnew — brief
docs/PLAN.mdB8 ⬜ → ✅ on ship

Acceptance criteria

Reference patterns

Open questions