Ailiance-fr/router-v6-minilm

ailiance-router-v6 (MiniLM L6 v2 + 2-layer MLP, niche-augmented)

Domain classifier head used by the ailiance gateway to route incoming chat requests to the most adapted backend worker. Successor to v5; adds 150 manually-curated prompts across 13 niche technical domains.

Performance

• Top-1 accuracy: 87.7 %
• Top-3 accuracy: ~98 %
• Validation set: ~2 000 prompts across 32 domains.

Improvements over v5 (65.5 % top-1 on the same valid set):

• Niche domains (kicad-pcb, stm32, embedded, dsp, iot, music-audio, platformio, ml-training, llm-orch, web-backend, web-frontend, yaml-json) now route correctly thanks to the curated prompts in scripts/augment_niche_domains.py.
• Threshold calibrated to 0.50 (vs 0.12 in v5) — ambiguous prompts now fall back to Gemma instead of being force-routed.

Architecture

• Encoder: sentence-transformers/all-MiniLM-L6-v2 (frozen, 384d)
• Head: 2-layer MLP Linear(384→256) -> GELU -> Dropout(0.1) -> Linear(256→32)
• Loss: BCE-with-logits (multi-label sigmoid)
• Optimizer: AdamW lr=1e-3, batch=128, 30 epochs, seed=42

🇪🇺 EU AI Act transparency

Field	Value
Provider	Ailiance (clemsail)
Role under AI Act	GPAI provider (this head). Encoder is upstream Microsoft / sentence-transformers under Apache-2.0.
Adapter type	Custom 2-layer MLP head (~85 k parameters)
Base model	sentence-transformers/all-MiniLM-L6-v2
License	Apache-2.0
Intended use	Domain routing only — classify a user prompt over 32 technical domains, decide which downstream worker handles the request. Not for standalone classification of safety-critical text.
Out of scope	Healthcare triage, legal classification, autonomous safety decisions, anything requiring calibrated probabilities.
Risk classification	Limited risk — Article 50 obligations apply.
Training data origin	All open permissive-licensed datasets + Ailiance in-house curation. Full per-source provenance: https://ailiance.fr/transparency/router-provenance
Reproducibility	https://ailiance.fr/transparency/router-rebuild-recipe + scripts/augment_niche_domains.py
Contact	postmaster@saillant.cc

Files

• router.safetensors — MLP head weights
• meta.json — encoder name, dim, head config, domain list
• label_map.json — domain ↔ index

Validated in ailiance/ailiance-bench v0.2

This model is referenced in the Ailiance benchmark suite (Phase 6 scoreboard, 7-task hardware-design evaluation).

See the full scoreboard: ailiance-bench README#scoreboard-lora-phase-6.