From completion to conversation.

Live generation — GPU inference. Greedy decoding. This experimental model can invent facts and miss instructions. Responses appear when generation finishes.

Inspect the evidence

These are recorded outputs, not live comparisons. Both models received the same native chat-formatted prompt, greedy decoding, and a 256-new-token cap. The base is a completion model; this is not its best possible completion-prompt baseline. Prompts were visible during development.

The saved fine-tune gives 11 cm; the base gives 15 cm. The fine-tune still misses the requested final-answer format.

Base: 99 new tokens; ended at EOS · Instruct: 226 new tokens; ended at EOS

Better fit to held-out conversations

On the same 511 development-validation conversations, assistant prediction loss went from 0.960699 → 0.516619, and teacher-forced token accuracy from 76.16% → 84.41%. These are development metrics, not independent benchmark scores or answer accuracy.

How it was trained

Full-model SFT with TRL, 8,192-token packing, assistant-response loss, learning rate 1e-4, global batch 16, and 15,904 optimizer updates (approximately two epochs). This is the full fine-tuned model, not a LoRA adapter.

What this does not establish

The model can hallucinate, violate output constraints, and produce incorrect code or math. No independent IFEval, GSM8K, MATH, HumanEval, MBPP, or multilingual improvement is claimed. No reasoning-specific post-training was performed. Generated tool-call text does not execute tools.

Attribution & use

Independent community derivative, not an official Sarvam AI release. The model is governed by the Sarvam AI Research License, including its commercial-use restrictions. Full license · Training details and provenance