laion/stageC-pbs-80-8B

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 (maxsteps) export — the legitimate end of training (`trainer.maxsteps=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, open-athena/stageD-thinkbudget-80-8B.

Final training metrics (globalstep80)

Training Traces

Training-time Daytona/Harbor rollouts: [open-athena/stageC-pbs](https://huggingface.co/datasets/open-athena/stageC-pbs) (the last episode of each trial — the rollouts the policy trained on after rollback/truncation).

Contents

4-shard safetensors weights + config + tokenizer/chattemplate + `generationconfig + rlconfig.json` (the launch config) + `traininglogs/ (per-step metrics CSVs + vLLM metrics + raw .out` chain logs — the W&B-equivalent, since Jupiter runs WANDB offline).