| --- |
| license: apache-2.0 |
| base_model: maxituc/aimi-base |
| tags: |
| - aimi |
| - lfm2-vl |
| - lora |
| - hot-swap |
| - chat |
| --- |
| |
| # aimini-chat — Aimi chat aimini (V0.1) |
|
|
| LoRA adapter (rank 64, INT8 ~44 MB) hot-swappable on the `maxituc/aimi-base` LFM2.5-VL graph. Trained 2026-05-22 → 23 on ~1200 conversational examples covering: |
|
|
| - `generate_file` / `edit_file` / `read_file` / `ask_question` / `list_files` tool calls |
| - French chat (vouvoiement, structured replies) |
| - Multi-turn conversational patterns |
| - Clarification when content is ambiguous (asks instead of inventing) |
|
|
| ## Usage |
|
|
| Hot-swap via `@aimi/provider-transformers`: |
|
|
| ```ts |
| { |
| id: 'aimi-chat', |
| hotSwap: { |
| base: 'maxituc/aimi-base', |
| baseData: 'onnx/model_q4.onnx_data', |
| adapterRepo: 'maxituc/aimini-chat', |
| adapterFile: 'chat.data', |
| }, |
| } |
| ``` |
|
|
| ## Training |
|
|
| - Base : LFM2.5-VL stripped 1.3B |
| - LoRA : r=64 α=64, attention + MLP (q/k/v/out_proj, in_proj, w1/w2/w3) |
| - QLoRA NF4, 4 epochs, lr 2e-4 cosine, packing bfd |
| - 4 checkpoints saved ; `ep1` selected as winner via behavior eval sweep (best balance: least tool hallucination, cohérent intro/tool, template-mode preserved) |
|
|
