Aditya0427/ToolBoundary-tiny-aya-lora

ToolBoundary Tiny Aya adapter

I built a synthetic authorization lab and this experimental LoRA adapter with substantial ChatGPT/Codex assistance. This is noncommercial research, not a production safety model. The adapter requires access to the original gated base model and its license/AUP. No base weights or tokenizer are redistributed here.

Exact training

CohereLabs/tiny-aya-global revision af89d219b53ed9b13b8a4645f9c8028973510324. AgentTune d4db4c12d52342227c4113a70183d0f64aa2f345 using its trainer_factory hook with TRL. 240 synthetic demonstrations; 60 completion-only supervised steps; seed42; NF4 base, FP16 compute, FP32 trainable adapters; LoRA rank8 on q_proj/v_proj. Free Colab Tesla T4. Training wrapper time348.884 seconds. Source commit d26cd07487224918981af33eed6d9a6978699453.

Measured results

Strict JSON evaluation; identical deterministic before/after generation. Each split has24 synthetic cases.

Metric	Development before	Development after	Heldout before	Heldout after
Valid raw JSON actions	7	24	6	24
Exact correct actions	2	20	2	19
Unauthorized execute attempts	1	2	0	3

Validity and exact task completion improved on this small template benchmark. Safety is NOT established: invalid baseline outputs hide attempted-action classification, and the trained model still makes unauthorized proposals. A deterministic guard must enforce trusted permissions server-side. Training token accuracy is not evaluation accuracy. Earlier failed and negative pilots are retained in the source repository.

Limits and reproducibility

All train/evaluation fixtures are inspectable; shared template families mean this is not an independently blinded or external benchmark. Single-step English, string-valued synthetic records only. No real users, production data, deployment claims or security guarantee. Models do not grant permission. Dataset IDs are separate across training and evaluation. Raw outputs, logs, hashes and manifest accompany this adapter.

Source and data: https://github.com/AdityaM274/ToolBoundary/tree/codex/toolboundary

Base model: Cohere Labs, subject to its C4AI CC BY NC terms and acceptable-use policy. AgentTune: Lexsi Labs LSAL1.1; TRL, Transformers, PEFT and bitsandbytes retain their licenses. Project code MIT does not relicense these components.