README.md

---
language: en
tags:
  - loracle
  - lora-oracle
  - mechinterp
  - llama-3-3-70b
  - drgrpo
license: mit
base_model: meta-llama/Llama-3.3-70B-Instruct
---

# llamacle_drgrpo_v1_step20 - DrGRPO RL on top of llamacle_v6_clean (step 20)

Best loracle ckpt by AB Llama-70B any-match: 74.1% (vs 55.2% pretrain baseline).

## AB Llama-70B (3 prompts x 2 rollouts, judge=Sonnet 4.6)
- Overall any-match: **0.741** (baseline 0.552)
- synth_docs_high: 0.867 (baseline 0.667)
- synth_docs_kto:  0.857 (baseline 0.857)
- transcripts_high: 0.467 (baseline 0.200)
- transcripts_kto:  0.786 (baseline 0.500)
- rollout_mean: 0.517

Continuation of `ceselder/llamacle_v6_clean_step1875` via online Dr. GRPO RL
on the 2,500 held-out FineWeb LoRAs. 32 prompts/cycle x K=16 rollouts
(sub-batched 4xK=4), lr=7e-6, eps=0.2/0.28, NF4-DDP across 6 B200s.