NewSonnet/cua-s1-forge-rlcd

CUA-S1 Forge RLCD Smoke Checkpoint

Experimental verifier-driven fine-tuning of the small CUA-S1 form-action scorer. This model selects among supplied GUI form actions; it is a classifier, not a general-purpose computer-use agent or an autoregressive language model. The architecture is derived from the open CUA-S1 source and the checkpoint was trained from scratch on the included synthetic recipe.

Status

This is a reproducibility checkpoint from a Google Colab NVIDIA Tesla T4 smoke run. It demonstrates that the RLCD objective trains and produces a loadable safetensors checkpoint. It is not presented as a quality improvement over the supervised baseline.

On the untouched synthetic test split:

The RLCD run improved top-1 accuracy by only 0.27 percentage points while worsening NLL and calibration. More seeds, stronger reward design, and better held-out evaluation are required before drawing conclusions.

Training

• —600 deterministic synthetic episodes, split by form signature
• —12,079 train rows, 1,631 validation rows, 1,821 untouched test rows
• —2 RLCD epochs, batch size 128, learning rate 5e-4
• —Entropy bonus, KL anchor to the supervised baseline, and supervised stabilizer
• —Upstream source commit: 27a318c3a616f9ff19d24fe2acca7517c5f8fa7b

See training_config.json, t4_smoke_report.md, and code/ for provenance and the experimental implementation. The source project is cua-s1-forge.

Scope and limitations

This is an early research artifact. It has not been evaluated on real-world forms, consequential workflows, or a broad computer-use benchmark. Do not use it to submit forms, alter accounts, make financial or legal decisions, or take other consequential actions without an independent verifier and human review.

This release uses no Jev or TypeSafe outputs and does not claim to reproduce any private Jev training implementation.