| 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. | |