Qwen3.5-4B ALFWorld token-wise absolute-MI run
This repository is a complete archival export of:
qwen35_4b_absmi_w008_clip01_rawdiv01_within0_sameid_w120_8gpu_20260810_190116
The run used AReaL/GRPO with token-wise absolute mutual-information reward,
a final token reward clip of 0.1, cross-group diversity reward
0.1 * (1 - cosine), zero within-input similarity reward, and a
120-word skill output limit.
Checkpoint boundary
- Best completed evaluation checkpoint: global step 57, model version 58.
- Best all-unseen SR: 0.3339552239.
- Best ID SR: 0.4024390244.
- Best OOD SR: 0.2259615385.
- Latest saved model checkpoint: global step 58.
- Training recovery state: global step 58.
- Evaluation at global step 58 is partial/incomplete and is preserved as-is.
The best checkpoint was selected using alfworld_eval_all_sr.
Repository layout
checkpoints/: exact AReaL checkpoint root, containing the best step-57
model, latest step-58 model, distributed recovery checkpoint, and
recover_info.archives/outputs_and_trajectories.tar.gz: all generated skills,
prompts, training/evaluation trajectories, trajectory pool, metrics, and
the partial step-58 evaluation state.archives/areal_wandb_logs.tar.gz: AReaL worker logs, resolved config,
and all local W&B run files from the resume history.archives/launcher_monitor_logs.tar.gz: launcher/bootstrap logs and
system monitoring history.archives/quarantine_recovery_artifacts.tar.gz: artifacts quarantined
during two clean-resume operations, retained for auditability.metadata/: directly browsable best-checkpoint record, evaluation
history, resolved config, and selected skill/evaluation summaries.SHA256SUMS: integrity hashes for every uploaded archive, checkpoint,
and metadata file.
Loading a checkpoint
Download a repository snapshot and point Transformers at the desired local
checkpoint directory, for example:
checkpoints/default/epoch0epochstep57globalstep57
The repository root itself is an experiment archive, not a directly loadable
Transformers model directory.
Provenance and safety
The source run was no longer active when archived. A credential-pattern scan
covered 177,800 non-model files (11,223,506,161 bytes), plus checkpoint
configuration and recovery metadata. No candidate HF, OpenAI, W&B, AWS, or
DashScope credentials were detected. Original absolute runtime paths and
host-level diagnostics are retained in logs for reproducibility.
