Auto-router

Détails LoRA

Spécialistes LoRA routés

Sur les domaines hardware/code, l'auto-router délègue à un des 30 adaptateurs spécialistes qwen36-35B servis sur Mac Studio (:9360 / :9361).

Provenance · Annex IV §1(c)

{
  "_doc": "EU AI Act Annex IV §1(c) — model lineage / supply-chain record.",
  "deployment_id": "ailiance/auto",
  "deployed_at_utc": "2026-05-05T13:55:00Z",
  "deployed_on_host": "electron-server (gateway-side, CPU)",
  "served_via": "in-process MLP head invoked by the ailiance gateway (FastAPI :9300)",
  "source": {
    "provider": "Ailiance (head trained in-house) + Microsoft / Hugging Face (encoder)",
    "encoder": {
      "huggingface_repo": "sentence-transformers/all-MiniLM-L6-v2",
      "huggingface_url": "https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2",
      "license_spdx": "apache-2.0",
      "embedding_dim": 384,
      "parameters": 22700000
    },
    "head": {
      "checkpoint": "ailiance/output/router/router.safetensors",
      "previous_checkpoint": "ailiance/output/router-v1-backup-1777983938 (rollback)",
      "training_script": "scripts/train_router.py",
      "training_data_script": "scripts/build_router_data.py + scripts/augment_router_data.py"
    }
  },
  "weights_file": {
    "filename": "router.safetensors",
    "size_bytes": 432000,
    "sha256_expected": null,
    "comment": "Hash captured on next deploy. Reproducible from training data + scripts/train_router.py with fixed seed."
  },
  "architecture": {
    "family": "Encoder + classifier head",
    "encoder_kind": "BERT-style sentence transformer (MiniLM L6 v2, 384d)",
    "head_kind": "2-layer MLP (Linear → GELU → Dropout → Linear)",
    "head_hidden_dim": 256,
    "num_domains": 34,
    "total_parameters": 22700000,
    "active_parameters_per_token": 22700000
  },
  "training_data": {
    "source": "Ailiance proprietary corpus — ~46 500 user-style prompts classified across 34 technical domains",
    "datasets_combined": [
      "ailiance-mac-tuner/data/micro-kiki/classified/*.jsonl (32 domains, ~46 000 lines)",
      "scripts/augment_router_data.py — manually curated 528 prompts targeting docker/spice/calcul-normatif gaps + FR/EN code-switching"
    ],
    "split": "80/20 train/valid, deterministic seed 42",
    "language_mix": "primarily fr/en (mixed), small amounts of de/it/es",
    "copyright_status": "All training prompts are short user-style queries written or paraphrased internally; no scraped third-party copyrighted text."
  },
  "training_method": {
    "loss": "BCE-with-logits (multi-label sigmoid)",
    "optimizer": "AdamW lr=1e-3",
    "epochs": 30,
    "batch_size": 128,
    "hardware": "Mac Studio M3 Ultra (CPU encoder + CPU/Metal head)",
    "wall_clock_minutes_estimate": 6,
    "energy_estimate_wh": 50,
    "metrics_final": {
      "top1_accuracy": 0.652,
      "top3_accuracy": 0.857,
      "validation_set_size": 11644
    }
  },
  "modifications_post_download": [
    "Encoder used unchanged from sentence-transformers/all-MiniLM-L6-v2",
    "Classifier head trained from scratch on internal corpus"
  ],
  "intended_use": "Domain routing only — classifies an inbound user prompt into one of 34 technical domains and forwards the chat request to the most adapted backend worker (Apertus / Devstral / EuroLLM / Qwen / Gemma).",
  "out_of_scope": [
    "Standalone classification of safety-critical text (e.g. medical triage)",
    "Content moderation",
    "Anything requiring calibrated probabilities — head is not calibrated"
  ]
}