Sovereign-Bharat-V1 — Multilingual Indic + Temporal RAG

Pipeline A: Bharat-Context (Indic Reasoning)

Data Source	Verified	Content
ai4bharat/indic-instruct-data-v0.1	✅	15 configs × Hi/En, ChatML messages
ai4bharat/IndicIFEval	✅	14 Indian languages, IF constraints

Indic Rewards: Language consistency (script detection) + reasoning quality + correctness

Pipeline B: Temporal RAG (Citation Tracking)

Citation Format: [CITE: source="...", date="YYYY-MM-DD", confidence=0.0-1.0]

Temporal Rewards: Citation completeness + quality + chronological consistency

Verified Results

Hindi→Hindi response:   +1.0 ✅ (correct script)
Hindi→English response: -0.5 ✅ (code-switch penalized)
Tamil→Tamil response:   +1.0 ✅ (correct script)
3 valid citations:      +1.0 ✅ (complete)
No citations + claims:  -0.5 ✅ (penalized)
Honest "can't verify":  0.0  ✅ (neutral)

Launch

MODEL_ID=ai4bharat/Airavata LANGUAGES=hi,mr,ta python bharat_temporal.py