micro-kiki router v0 — 3-domain toy classifier (deprecated)
Deprecated / historical. First routing experiment (pre-pivot, Qwen3.5-4B era). Superseded by
micro-kiki-router-v4(MiniLM, 34 domains) and the mpnet router in the GitHub repo.
Correction vs. the previous card: the encoder is not all-MiniLM-L6-v2.
router_metrics.json records embedding_method: causal_lm_inner_model_mean_pool with
embed_dim: 2560 — mean-pooled hidden states of a causal LM (hidden size 2560, i.e. the
Qwen3.5-4B base of the time).
Recipe (from router_metrics.json)
| Parameter | Value |
|---|---|
| Features | mean-pooled causal-LM hidden states, 2560d |
| Head | PyTorch classifier (router.pt, 5.3 MB, pickle format) |
| Domains | 3: chat-fr, python, reasoning |
| Train / eval | 531 / 100 examples |
| Epochs | 5 |
| Script | scripts/legacy/train_router_v0.py (hypneum-lab/micro-kiki) |
Evaluation
macro-F1 1.0, accuracy 1.0 on 100 eval examples — a saturated 3-class toy task; not indicative of real routing quality.
Limitations
router.pt is a pickle file: loading it executes Python code; load only from a trusted
copy (torch.load(..., weights_only=True) where possible). Three domains only.
Citation / provenance
micro-kiki concept DOI 10.5281/zenodo.19666759 · https://github.com/hypneum-lab/micro-kiki. Provider Clément Saillant / L'Electron Rare for Hypneum Lab; contact clement@saillant.cc.