navyam-ai/navya-1c-sft-dpo-grpo-reasoning

Navya-1c-sft-dpo-grpo-reasoning

A reasoning-distilled variant of the India-first Navya-1c chat model (1.31B params).

This is the 1.31B Navya-1c lineage — from-scratch base (~103B tokens + WSD cooldown on 8×H100) → SFT → DPO → GRPO → reasoning distillation — taught to answer with an explicit <think>…</think> chain-of-thought before its final answer. The reasoning teacher was a locally-run deepseek-r1:14b (R1-Distill) over India personal-finance prompts.

Output format:

<think>
step-by-step reasoning about the question
</think>

final answer

⚠️ Status: research proof-of-concept

This checkpoint demonstrates the <think> reasoning habit but was distilled on a small purpose-built corpus and overfits it (train loss → ~0, validation loss rising). It reliably produces the reasoning format, but is not a production-grade reasoning model — treat it as a research PoC. The flagship production chat model remains `navyam-ai/navya-1c-sft`.

Research model — not investment advice. Small model; verify anything consequential.

Intended use

India personal-finance Q&A (savings, SIPs, loans, credit, insurance, tax) in English and Hindi/Hinglish, with visible step-by-step reasoning.

Details

• —Params: 1.31B · Author: Navyam AI (Bachatt)
• —Architecture: Llama-recipe — dim 2048, 26 layers, GQA 16/4 heads, SwiGLU, RMSNorm, RoPE, tied embeddings, 64k vocab, 4096 context
• —Tokenizer: custom tokenizer-v0.3-64k
• —Chat template: navya-chat-v1 (role tokens <|reserved_0..4|>)
• —Lineage: navya-1c (base) → sft → dpo → grpo → reasoning distillation (deepseek-r1:14b teacher)
• —Source: https://github.com/bachatt-app/navyam-gpt
• —License: Apache-2.0