a3-rl-DCAgent_exp_rpt_e2egit-v2 (step 10, 8B)
RL checkpoint from SkyRL (RLOO_n, FSDP2) trained on the exp_rpt_e2egit-v2
terminal-bench task set. This is the EMA-best checkpoint (global_step_10)
selected from a legitimately-completed run (2 epochs x 7 steps/epoch = 14
training steps; lead SLURM job 515469, exit 0).
- Base model:
laion/GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink(Qwen3-8B architecture) - Dataset:
DCAgent/exp_rpt_e2egit-v2 - Algorithm: RLOO_n, no KL loss, eps_clip 0.2/0.05, lr 8e-6, train_batch_size 64
- Selection: EMA (alpha=1/3, 5-period) of
reward/avg_raw_reward, step 10 EMA=0.6787
Training Traces
Training-time Daytona/Harbor rollouts for this run are uploaded as a companion dataset: