Open-source commitment

• Code (training, data processing, evaluation): Apache-2.0, on GitHub.
• Weights: LoRA adapters and merged 4-bit models on Hugging Face, under the base model's Apache-2.0 licence.
• NO-FC benchmark: CC BY 4.0, with the translation and validation scripts.
• Results: every config, seed, log and score, published even when the result is negative.
• Data hygiene: only openly licensed sources. Each source's licence is listed in the model card. No personal data is collected.

Limitations and safety

Small models can still make up tool calls and arguments. NO-FC deliberately measures declining and clarifying, and the model card will document the error rates we measure. The models are for research and local assistants, not for decisions without a human in the loop.