30 lines
1.1 KiB
Markdown
30 lines
1.1 KiB
Markdown
---
|
|
license: apache-2.0
|
|
base_model: laion/GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink
|
|
tags:
|
|
- rl
|
|
- rloo
|
|
- terminal-bench
|
|
- openthoughts-agent
|
|
---
|
|
|
|
# ablation-pymethods2test-seqmean-arm0-tis (step 15)
|
|
|
|
RLOO length-bias A/B ablation, **arm0 = sequence_mean loss reduction WITH TIS**
|
|
(truncated importance sampling). Trained with SkyRL (RLOO-n, FSDP2) on
|
|
`DCAgent/exp_rpt_pymethods2test-large` from the Qwen3-8B pre-RL base
|
|
`laion/GLM-4_7-swesmith-...-fixthink`.
|
|
|
|
Checkpoint selected at **global_step 15** by trailing-5 EMA (alpha=1/3) of
|
|
`reward/avg_raw_reward` over the full training chain (cap step <= 80).
|
|
|
|
## Training Traces
|
|
|
|
Training-time Daytona/Harbor rollouts for this run are uploaded as
|
|
a companion dataset:
|
|
**[penfever/ablation-pymethods2test-seqmean-arm0-tis](https://huggingface.co/datasets/penfever/ablation-pymethods2test-seqmean-arm0-tis)**
|
|
|
|
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.
|