# a3-rl-laion_nemotron-gym-knowledge-web-search-mcqa (global_step 25)

RL (fully-async, fan-out K=4) checkpoint trained on
`laion/nemotron-gym-knowledge-web-search-mcqa`, selected by EMA-best
(5-period EMA, alpha=1/3) of `reward/avg_raw_reward` over the full
step-1->80 training curve.

- **Base model (SFT):** `laion/GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink` (Qwen3-8B architecture)
- **Selected step:** 25 (reward=0.5918, EMA=0.5281)
- **Training:** Jupiter, 56 GPUs / 14 nodes, SkyRL fully-async, max_steps=80 (completed)

## Training Traces

Training-time Daytona/Harbor rollouts for this run are uploaded as
a companion dataset:
**[penfever/a3-rl-laion_nemotron-gym-knowledge-web-search-mcqa](https://huggingface.co/datasets/penfever/a3-rl-laion_nemotron-gym-knowledge-web-search-mcqa)**

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.