a3-rl-DCAgent_exp_rpt_e2egit-large-15-8B

a3-rl-DCAgent_exp_rpt_e2egit-large (step 15, 8B)

RL (GRPO/rloo_n via SkyRL) checkpoint trained from laion/GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink on dataset DCAgent/exp_rpt_e2egit-large.

This is global_step_15, selected as the best checkpoint by 5-period EMA of reward/avg_raw_reward (EMA 0.864, raw 0.842). The run trained the full 80 steps but collapsed late (entropy 0.0961 -> 0.0215 from ~step 56), so an early-mid checkpoint is correctly the best.

Launch config: hpc/skyrl_yaml/jupiter/56GPU_base.yaml (included as rl_config.yaml).

Training Traces

Training-time Daytona/Harbor rollouts for this run are deferred — the make_and_upload_trace_dataset script stalled in GPFS I/O (uninterruptible disk-sleep, 33 min, zero output) and was killed past a 15 min time-box. The companion dataset will live at open-athena/a3-rl-DCAgent_exp_rpt_e2egit-large once re-attempted (no partial upload was created).

Training Metrics

training_logs/ contains the per-chain-link metrics_job_*.csv, the merged metrics_table.csv / vllm_metrics_table.csv, the markdown report (metrics_report.md), and the reward/turn-count plots. Raw .out logs from all 10 chain links are included too.