laion/a3-rl-laion_nemotron-gym-agent-calendar-80-8B

a3-rl-laion_nemotron-gym-agent-calendar (step 80, 8B)

RL-trained (SkyRL) agent checkpoint, selected by 5-period EMA (alpha=1/3) of reward/avg_raw_reward across all 80 steps of the training chain.

• —Selected checkpoint: global_step_80 (EMA = 0.9785, raw reward at step = 0.9727)
• —Final-step metrics: reward/avg_pass_at_8 = 1.0, reached max_steps = 80/80 (clean exit-0 completion)
• —Base model: laion/GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink (Qwen3-8B architecture)
• —Training dataset: laion/nemotron-gym-agent-calendar
• —Training type: RL (SkyRL)

The launch RL config is included as rl_config.yaml. Parsed training metrics, plots, and raw logs are under training_logs/.

Training Traces

Training-time Daytona/Harbor rollouts for this run are uploaded as a companion dataset: [open-athena/a3-rl-laion_nemotron-gym-agent-calendar](https://huggingface.co/datasets/open-athena/a3-rl-laion_nemotron-gym-agent-calendar)

The dataset contains the last episode of each trial (per make_and_upload_trace_dataset --episodes last) — the same rollouts the policy was trained on after rollback / truncation.