stageC-pbs-80-8B

Agentic RL (SkyRL, FSDP2) checkpoint from the stageC cell of the a3 / pymethods2test agentic-RL family. This is the global_step_80 (max_steps) export — the legitimate end of training (trainer.max_steps=80).

• Base model: laion/GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink (a Qwen3-8B SFT)
• Dataset: DCAgent/exp_rpt_pymethods2test-large
• Algorithm: RLOO-N (advantage_estimator=rloo_n_pbs), per-batch-shaped reward channel (enable_token_reward_channel=true), loss_reduction=seq_mean_token_sum_norm_global (seqnorm), TIS on (use_tis=true, tis_imp_ratio_cap=2.0), eps_clip=0.2/0.2, no KL loss.
• Training: 14 nodes, train_batch_size=64, 2 epochs, max_steps=80, ckpt_interval=2, hf_save_interval=5. WANDB offline (Jupiter).
• Sibling cells: laion/stageB-channel-80-8B, penfever/stageD-thinkbudget-80-8B.

Final training metrics (global_step_80)

metric	value
reward (avg_raw)	0.5645
pass@8	0.672
entropy	0.290
raw_grad_norm	~3.6e-5 (seqnorm global-denom artifact)
tis_imp_ratio_mean	0.987
tis_imp_ratio_capped_fraction	~1e-5

Training Traces

Training-time Daytona/Harbor rollouts: penfever/stageC-pbs (the last episode of each trial — the rollouts the policy trained on after rollback/truncation).

Contents

4-shard safetensors weights + config + tokenizer/chat_template + generation_config + rl_config.json (the launch config) + training_logs/ (per-step metrics CSVs + vLLM metrics + raw .out chain logs — the W&B-equivalent, since Jupiter runs WANDB offline).