metadata
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_step10 — DrGRPO RL on top of llamacle_v6_clean
Continuation of ceselder/llamacle_v6_clean_step1875 via online Dr. GRPO RL
on the 2,500 held-out FineWeb LoRAs from v6 pretrain. 32 prompts/cycle x K=16
rollouts, lr=7e-6, eps_low/high=0.2/0.28, NF4-quantized base + DDP across 6 B200s,
sub-batched K=4x4 decode, forward_ckpt_inject for backward.
This is step 10 of 80 (early checkpoint).
Score progression (judge 1-10, mean over kept rollouts)
cycle 1: 4.08 cycle 2: 3.85 cycle 3: 4.45 cycle 4: 4.95 cycle 5: 4.95 cycle 6: 5.17 cycle 7: 5.77 cycle 8: 4.21 cycle 9: 4.82 cycle 10: 4.73