---
base_model: laion/GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink
tags:
- rl
- skyrl
- agentic
- swe
library_name: transformers
---

# ablation-pymethods2test-seqmean-arm0-30-8B

RL (SkyRL GRPO) checkpoint from the **sequence-mean / RLOO-n (arm0)** ablation of
the a3-successor study. The policy loss uses `loss_reduction=sequence_mean` with
the `advantage_estimator=rloo_n` (RLOO-n) estimator, contrasting with the
token-mean reduction of the a3 series.

- **Base model:** [laion/GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink](https://huggingface.co/laion/GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink) (a Qwen3-8B SFT)
- **Training dataset:** [DCAgent/exp_rpt_pymethods2test-large](https://huggingface.co/datasets/DCAgent/exp_rpt_pymethods2test-large)
- **Checkpoint:** `global_step_30`.
- **Training:** SkyRL GRPO, `hf_save_interval=5`, 14x GH200 nodes on JSC Jupiter.

The `rl_config.yaml` in this repo is the exact launch config used for reproducibility.
This is the step-30 checkpoint of the same run that produced
[laion/ablation-pymethods2test-seqmean-arm0-15-8B](https://huggingface.co/laion/ablation-pymethods2test-seqmean-arm0-15-8B) (step 15).

## Training Traces

Training-time Daytona/Harbor rollouts for this run are uploaded as
a companion dataset:
**[penfever/ablation-pymethods2test-seqmean-arm0](https://huggingface.co/datasets/penfever/ablation-pymethods2test-seqmean-arm0)**

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.