rv-laya-ft-v2-coreml

Fine-tuned Laya judge for rv: shell commands in, allow / review / block out. Single forward pass, no generated text. Runs fully offline once downloaded.

Install for rv

Scripts/laya/fetch-models.sh --engine ft
# verifies SHA-256 per file, installs to ~/.config/rv/models/
rv doctor  # judge line should report the bundle

Provenance

• Base checkpoint: convaiinnovations/laya-multilingual at 1720e3e3357cfe1e281542e223f8273b0890ca34.
• Fine-tune data: shell-safety v1.1 (MIT) + v2 context serialization (train hash d8e153ea, validation hash 5cd0c2b0).
• Exam scores (torch fp32): shell held 36/36 kept 33/35, adversarial held 18/19, incidents held 12/13, ko held 17/21.
• Known miss preserved at p~1.0: git add -A (covered by rv packs in the full system). Context at runtime is command + cwd only (partial vs training); no production-traffic accuracy is claimed.
• Original Laya weights and code are by Convai Innovations and contributors, Apache-2.0. This is a fine-tune, not an official Convai Innovations release.

Files

• model.mlpackage/Manifest.json — 617 bytes, SHA-256 3c713c254fee9e6914355e82f4c3e9adc4dbca667a956b284d38f5a18393a1d5
• model.mlpackage/Data/com.apple.CoreML/model.mlmodel — 356877 bytes, SHA-256 fd60b8928154754362d660163a4a15e203337ae4cb2ccf98b461233db366bce4
• model.mlpackage/Data/com.apple.CoreML/weights/weight.bin — 613899008 bytes, SHA-256 5abb590e28cef8c312b3afc2f28046f186f1127893989bda6a880f352261377b
• tokenizer.json — 34363188 bytes, SHA-256 609d8f4c067cd3950f88594c5a802616cea245823836ef5848ee4fc40aab5b6f
• config.json — 423 bytes, SHA-256 c0099d659944d44cb3d711b55c25482819f2e8f5e17f2d0b37d6d7cb1b22aba5

See config.json (checkpoint: laya-ft-v2-fp16) for label order, max length, and training hashes.