---
base_model: laion/GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink
tags:
- rl
- skyrl
- agentic
- terminal-bench
- tis
---

# explore-tis-untrunc (global_step 45, 8B)

Agentic RL checkpoint (SkyRL) finetuned from
`laion/GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink` (Qwen3-8B SFT base).

## Run config
- **Algorithm:** RLOO-n advantage, `seq_mean_token_sum_norm_global` loss reduction, no KL loss.
- **TIS:** enabled (`tis_imp_ratio_cap=2.0`) — truncated importance sampling correcting the vLLM↔FSDP bf16 logprob gap.
- **Sampling (untruncated exploration variant):** the distinguishing feature of this ablation is untruncated sampling during exploration.
- **Dataset:** `DCAgent/exp_rpt_pymethods2test-large`.
- **Harness:** terminus-2 / Harbor agentic terminal-bench, n_samples_per_prompt=8.

## Checkpoint selection
This is **global_step 45**, selected by trailing-5 EMA (alpha=1/3) of `reward/avg_raw_reward`,
**restricted to the genuine training region (steps <= 78)**. Steps 79+ exhibited an
artifact reward jump (single-block greedy / eval-checkpoint passes, not a learning event)
and an eternal-retry tail, so they were excluded from candidate selection.
At step 45: raw reward ~0.518, trailing-5 EMA ~0.495 (highest among saved exports in-region).

## Training Traces
Rollout traces for this run: https://huggingface.co/datasets/penfever/explore-tis-untrunc

## Training logs
Parsed SkyRL metrics, vLLM metrics, and raw trainer logs are in the `training_logs/` folder of this repo.
The serialized launch config is `rl_config.yaml`.