GPA-0-Final
Cursor-conditioned OpenWAM policy with a Codex controller and asynchronous XPolicyLab WebSocket integration.
Release files
| File | Contents |
|---|---|
policy/GPA-XPolicyLab-runtime.tar.gz |
XPolicyLab checkout, GPA policy/harness/controller, cursor step 58000 weights, 42 frozen task experiences |
envs/gpa-policy-python.tar.gz |
Relocatable Linux x86_64 policy Python environment |
envs/node-codex.tar.gz |
Node.js, npm, Codex CLI and locked Node dependencies |
SHA256SUMS |
Artifact checksums |
Experiment experience is integrated in XPolicyLab/policy/GPA/experience/: 84 complete A/B rollout files, all parent lineages, 42 frozen task skills and per-task checksums. References resolve relative to the extracted task directory. Evaluation forks the matching B context and keeps task skill immutable.
The runtime contains the published checkpoint_step_58000.safetensors plus tokenizer, normalization and model metadata. Model source: https://www.modelscope.cn/models/bbhua112/Agent-OpenWAM-RoboDojo-Cursor-14x8 . Upstream licenses are retained in the archives.
The official evaluator already provides RoboDojo; the environment archives contain policy dependencies only. No simulator installation or assets are included. The asynchronous adapter supports at most 10 independent environment IDs per policy port. This is not a claim of a completed 80-environment performance benchmark.
Private account credentials, proxy configuration, SMTP settings and dashboard passwords are NOT in this repository; they are supplied separately. Use the latest deployment guide in this repository before starting services. It adds the system-Python websockets>=14 prerequisite and import checks and takes precedence over the older policy/GPA/START.zh.md bundled in the archive. Private credentials/configuration are still supplied separately.
Release validation: relocated package imports, all 42 A-to-B lineages restored and forked with Codex without credentials or model calls, artifact integrity checks, and known-credential scans. This validation does not measure task success rate.