clemsail/micro-kiki-router-v4

micro-kiki Router v4 — 34-Domain Multi-Label Classifier

MiniLM-L6 embedding + MLP head. Classifies a user prompt into up to 34 domains (sigmoid, not softmax — domains are not mutually exclusive). Used at inference time by the micro-kiki serving layer to pick the top-K LoRA adapters to load alongside the Qwen3.6-35B-A3B base.

Numbers (2026-04-22)

Metric	Value
Train / val	21 510 / 5 377
top-1 accuracy	0.464
top-3 accuracy	0.696

Files

• router.safetensors — MLP head weights
• label_map.json — domain ↔ index map
• meta.json — training config snapshot

Source

Training & serving pipeline at hypneum-lab/micro-kiki. Trained by scripts/train_router_v4.py.

🇪🇺 EU AI Act transparency

This adapter is provided as a fine-tuned LoRA under the AI Act framework (Regulation EU 2024/1689). Compliance metadata:

Field	Value
Provider	L'Électron Rare (clemsail / electron-rare)
Role under AI Act	GPAI provider for this adapter
Base model	sentence-transformers/all-MiniLM-L6-v2 — see upstream provenance
Adapter type	LoRA / PEFT — adapter weights only; base unchanged
Training data origin	L'Électron Rare proprietary technical corpus + curated public docs
License	Apache-2.0 (adapter). Upstream base licence applies separately.
Intended use	Domain-routing classifier
Out of scope	Healthcare diagnosis, legal advice, autonomous safety-critical decisions, generation of malicious code
Risk classification	Limited risk — Article 50 transparency obligations apply
Copyright respect	Training data does not include scraped copyrighted material. Opt-out signals (robots.txt, ai.txt) are honoured for web-sourced data.
Full provenance	https://github.com/L-electron-Rare/eu-kiki/tree/main/docs/provenance
Contact	postmaster@saillant.cc — biased output reports, copyright concerns, etc.

⚠️ You are using an AI model. Outputs may be inaccurate, biased or fabricated. Do not act on them without independent verification, especially in regulated domains.