# dpo_w2c — RAI content-safety classifier (DPO precision-tuned)

**Nemotron-3-Nano-4B → v3d_tjpi3o6 (full-SFT) → DPO.** lowest v14, most recall cost.

DPO (β0.05, LoRA r32) from `v3d_tjpi3o6` on combined v1+wave2 (1250) — preference pairs that fix confident benign
false-blocks (JB own-agent-control §07, IP own-content §08, PII synthetic §09, feedback-reading §13)
with contrastive recall-preserve pairs. DPO's non-saturating gradient moved the confident-cluster
false-blocks that full-SFT (100+ arms) walled at 0.36%.

## Metrics (v3_1d policy; gate = v14≤0.35 · hello≥0.47 · v200-P≥0.95 · rvR≥0.80 · rvP≥0.90) — **5/5**
| v14 Block% | hello R | v200-P | v200-R | rvR | rvP |
|---|---|---|---|---|---|
| 0.2% | 0.707 | 0.996 | 0.865 | 0.86 | 0.933 |

vs the tjpi3o6 base (v14 0.36 / rvP 0.914 / v200-P 0.990): **v14 down, precision up (rvP, v200-P), recall held** —
DPO reduced benign false-blocks, which lifts every precision metric at once.

## Contents
- root: merged weights + tokenizer + modeling_nemotron_h.py (trust_remote_code)
- eval/: raw inferences on v14/hello/v200/robustness under v3_1d + metrics
- Serve: `vllm serve <repo> --trust-remote-code --mamba_ssm_cache_dtype float32`