Gates Foundation AI Fellows · India 2026

Sarvam-30B Responsible-AI Audit

A pre-registered audit of Sarvam-30B — India's sovereign-built reasoning LLM — across cross-lingual safety, maternal & child health factuality, agricultural advisory, demographic bias, and Indian PII handling. Built on top of the CeRAI AIEvaluationTool with seven patches for reasoning-model support.

Artefacts

Auto-generated test-suite report — five chart-driven category sections + nine data tables, rendered from results/findings.json + cerai_scores_*.jsonl.
HuggingFace dataset — 120-prompt manifest with full per-prompt provenance, viewer + API access.
Source repository — indic-eval CLI, seven CeRAI patches, bootstrap script, this site.

The hand-crafted findings paper (this homepage) is a placeholder until the Nerfies-styled write-up lands. The auto-generated report.html above is complete and reflects the current audit state.