explore-tis-minp-40-8B

RL (SkyRL, agentic terminal-bench / Harbor + Daytona) checkpoint from the explore-tis sampling-parameter ablation. This is the min-p sampling arm (explore-tis-minp).

• Base model: laion/GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink (an 8B model)
• Dataset: DCAgent/exp_rpt_pymethods2test-large
• Algorithm: RLOO-n, no KL loss, loss_reduction=seq_mean_token_sum_norm_global, TIS on (tis_imp_ratio_cap=2.0)
• Sampling: temperature=1.0, min_p=0.05, top_k=-1, top_p=1.0
• Selected checkpoint: global_step_40 (best trailing-5 EMA, α=1/3, of reward/avg_raw_reward over saved exports with step ≤ 78; EMA ≈ 0.4722). The 78-step cutoff was applied to exclude a step-79+ greedy/eval-pass reward artifact.
• Max training steps: 80

Training Traces

Rollout traces for this run: penfever/explore-tis-minp

Training Logs

Parsed metrics, reward-vs-steps plots, and raw console logs are under training_logs/ in this repo.