imdatta0/qwen3-4b-swegym-moto-kl02-multitrain-grpo-b04-lr1e6-s12-adapter

Qwen3-4B SWE-Gym Moto KL02 Multi-Train GRPO Probe Adapter

LoRA adapter checkpoint from the short multi-file-train GRPO probe.

Local run: 20260604_231110_swegym_q4b-kl02-multitrain-grpo-b04-lr1e6-s12_389b336

Source checkpoint: checkpoints/best_holdout

This run started from the Qwen3-4B KL02 adapter and trained for 12 GRPO steps on the generated multi-file train split with beta 0.04 and learning rate 1e-6. The held-out baseline at step 0 and periodic step 12 eval were both greedy 8/35 with mean reward 0.4234, so best_holdout is the step-0 best checkpoint selected by the run.

This checkpoint is preserved for reproducibility, not as a new best model.