| # ablation-pymethods2test-seqnorm-15-8B | |
| RLOO length-bias ablation — **arm1 seqnorm (base)** arm. RL-trained from the | |
| a3 pre-RL base on the `exp_rpt_pymethods2test-large` task set with SkyRL | |
| (RLOO-n advantage estimator, `seq_mean_token_sum_norm_global` loss reduction). | |
| - **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` | |
| - **Checkpoint:** `global_step_15` (selected by trailing-5 EMA of `reward/avg_raw_reward`, alpha=1/3, over the full chain; max_steps=80, hf_save_interval=5) | |
| - **Training framework:** SkyRL (FSDP2, vLLM), Jupiter GH200 | |
| The launch config is included as `rl_config.yaml`. | |
| ## Training Traces | |
| Training-time Daytona/Harbor rollouts for this run are uploaded as | |
| a companion dataset: | |
| **[penfever/ablation-pymethods2test-seqnorm](https://huggingface.co/datasets/penfever/ablation-pymethods2test-seqnorm)** | |
| 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. | |