30 lines
1.2 KiB
Markdown
30 lines
1.2 KiB
Markdown
|
---
|
||
|
|
base_model: laion/GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink
|
||
|
|
tags:
|
||
|
|
- rl
|
||
|
|
- skyrl
|
||
|
|
- qwen3
|
||
|
|
---
|
||
|
|
|
||
|
|
# a3-rl-laion_exp_rpt_stack-bash-v3 (global_step_70)
|
||
|
|
|
||
|
|
RL (RLOO_n / SkyRL FSDP2) fine-tune of
|
||
|
|
`laion/GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink`
|
||
|
|
(Qwen3-8B) on the `laion/exp_rpt_stack-bash-v3` agentic task set, agent `terminus-2`.
|
||
|
|
|
||
|
|
**PARTIAL / CANCELLED MID-RUN.** This run was part of the a3 RL series, which was
|
||
|
|
concluded as uninformative; SLURM job 589415 was cancelled at training step ~73
|
||
|
|
(2026-06-06). This checkpoint is the **latest available HF-ready export
|
||
|
|
(global_step_70)** — preserved as the run artifact, not a converged model. The
|
||
|
|
reward signal was low and flat throughout (avg_raw_reward ~0.08-0.21, EMA peak
|
||
|
|
~0.19 around step 60); grad-norm stayed healthy (<0.012) with no collapse.
|
||
|
|
|
||
|
|
## Training Traces
|
||
|
|
|
||
|
|
Training-time Daytona/Harbor rollouts for this run are uploaded as a companion
|
||
|
|
dataset:
|
||
|
|
**[DCAgent/a3-rl-laion_exp_rpt_stack-bash-v3](https://huggingface.co/datasets/DCAgent/a3-rl-laion_exp_rpt_stack-bash-v3)**
|
||
|
|
|
||
|
|
The dataset contains the last episode of each of the ~41,263 trials
|
||
|
|
(per make_and_upload_trace_dataset --episodes last).
|