purvansh01/safetypulse-probe-bundle
Updated • 46
SafetyPulse - data
SafetyPulse (Lexsi hackathon, Track 2b): eval splits, fine-tune data, contrast pairs, probe bundle, monitoring and audit logs.
Note What safetypulse check needs: refusal direction, 20-component circuit, probe pairs, thresholds.
purvansh01/safetypulse-logs
Preview • Updated • 47Note CIS, spot-check and ASR for 150 checkpoints; audits (harmful answers withheld); repair and report tables.
purvansh01/safetypulse-eval
Preview • Updated • 34Note Fixed dev/test splits: HarmBench, XSTest, GSM8K, MMLU.
purvansh01/safetypulse-finetune-data
Updated • 37Note Alpaca, CuratorKIT-curated Alpaca, GSM8K and negative-control fine-tuning sets.
purvansh01/safetypulse-pairs
Preview • Updated • 34Note 288 matched harmful / harmless contrast pairs.