| --- |
| license: apache-2.0 |
| base_model: Qwen/Qwen3.5-35B-A3B |
| base_model_relation: adapter |
| library_name: mlx |
| pipeline_tag: text-generation |
| language: |
| - fr |
| - en |
| tags: |
| - lora |
| - mlx |
| - qwen3.5 |
| - moe |
| - chat-fr |
| - micro-kiki |
| - hypneum-lab |
| - historical |
| --- |
| |
| # micro-kiki stack-chat-fr (MLX adapter, French chat) |
|
|
| Single MLX adapter file (`adapters.safetensors`, 6.76 GB) plus its iteration-1300 |
| checkpoint (`0001300_adapters.safetensors`, same size). Published 2026-05-05 as a |
| "multi-adapter stack for French chat" on top of `Qwen/Qwen3.5-35B-A3B`. |
|
|
| > **Provenance warning.** No `adapter_config.json`, no training log and no eval of this |
| > exact file are shipped. Base model, rank, target layers, learning rate and data are |
| > therefore **not recorded** for this artefact; the base model in the frontmatter is the |
| > one declared by the original card and could not be verified from files. Prefer |
| > `micro-kiki-v4-sota/chat-fr` (fully documented) for any use. |
| |
| ## Recipe |
| |
| | Parameter | Value | |
| |---|---| |
| | Base model (declared) | `Qwen/Qwen3.5-35B-A3B` | |
| | Framework | MLX (`mlx_lm`), BF16 | |
| | Iterations | ≥ 1300 (from the checkpoint file name) | |
| | Rank / scale / layers / lr / seq length | not recorded | |
| | Training data | micro-kiki `chat-fr` domain (French instruction/chat corpus, incl. Opus-distilled samples); counts not recorded | |
|
|
| ## Training host & cost |
|
|
| Mac Studio M3 Ultra 512 GB, MLX BF16 (project default). Peak memory, wall-clock, energy: |
| not recorded. Host powered off since 2026-10. |
|
|
| ## Evaluation |
|
|
| The file `evals/chat-fr-164-cross.json` kept in this repository reports HumanEval+ |
| (n=164, max_tokens 384) **pass@1 = 0.543**, total generation 729.7 s — but its own |
| metadata says it was run with `base_model = Qwen3.6-35B-A3B` and |
| `adapter = lora-qwen36-35b-v4-sota/chat-fr`, i.e. on the **V4-SOTA chat-fr adapter, not |
| on this file**. No evaluation of this 6.76 GB adapter is recorded. |
|
|
| June 2026 iact-bench caveat: on Qwen3.6-35B-A3B, the sibling ailiance LoRA family showed |
| no gain over the base except `html-css`; no comparable measurement exists for this adapter. |
|
|
| ## Limitations |
|
|
| Undocumented recipe; research artefact; no safety tuning; not for medical, legal or |
| financial advice. |
|
|
| ## Citation |
|
|
| micro-kiki concept DOI **10.5281/zenodo.19666759**. |
|
|
| ## Provenance |
|
|
| <https://github.com/hypneum-lab/micro-kiki> · <https://github.com/ailiance/ailiance-mac-tuner> |
|
|
| ## EU AI Act transparency |
|
|
| LoRA adapter weights only, base unchanged; provider Clément Saillant / L'Electron Rare for |
| Hypneum Lab; limited risk, Article 50; contact clement@saillant.cc. Outputs may be |
| inaccurate, biased or fabricated. |
|
|
