Views
No views yet
global_step_80 (max_steps) export — the legitimate end of
training (trainer.max_steps=80).laion/GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink (a Qwen3-8B SFT)DCAgent/exp_rpt_pymethods2test-largeadvantage_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.train_batch_size=64, 2 epochs, max_steps=80, ckpt_interval=2,
hf_save_interval=5. WANDB offline (Jupiter).laion/stageB-channel-80-8B, open-athena/stageD-thinkbudget-80-8B.| 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 |
last episode of each trial — the rollouts the policy trained on after rollback/truncation).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).