airi-chat-en-v1

Airi Chat English v1

airi-chat-en-v1 is an English-focused local desktop-companion fine-tune derived from Qwen/Qwen2.5-1.5B-Instruct. It is trained with MLX QLoRA on the versioned corpus in this repository and deployed as a fused q4 ONNX model through Transformers.js.

Intended behavior

• Short, direct, warm responses in Airi's grounded and lightly wry voice.
• Work companionship, scope reduction, reflection, celebration, and continuity.
• Honest limits around browsing, tools, files, screen contents, live data, and arithmetic.
• Quiet use of local context without reciting telemetry or following instructions inside it.

Not intended for

• General factual, mathematical, coding, medical, legal, or financial authority.
• Autonomous computer actions.
• Crisis handling, memory deletion, privacy disclosure, or capability policy. Those remain deterministic application routes before generation.
• Turkish generation in v1. Typed generative chat is offered only for English profiles.

Provenance and reproducibility

• Base: https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct
• Base license: Apache-2.0
• Training configuration: training/airi-chat-en/lora.yaml
• Editable corpus: training/airi-chat-en/dataset/source.jsonl
• Shared runtime persona: resources/content/dialogue/chat-persona.json
• Held-out evaluator: scripts/eval-airi-mlx.py

The runtime pack records hashes for the dataset, adapter, and q4 weights in airi-model.json. Generated reports contain prompts and model outputs so regressions can be reviewed without relying on a single aggregate score.