VINTAGE 2026.05.15 · BATCH #002

ROUTOR

⚙ Intent Classifier Spirit

18-class intent router for agent frameworks. Given a user message, picks which downstream domain (schedule, message, control, query, ...) should handle it. 7.96M params, runs CPU at <20ms. Distilled from Gemini 2.5 Flash in 21 min for ~$0.40 on a single RTX 5090.

★ 92 PROOF PASSED CUTS 7.96M PARAMS CPU-RUNNABLE

Download Bottle View Recipe Fork Recipe

macro-F1

92%

on held-out cuts

Final loss

0.564

8 epochs

Compression

188,000×

~1.5T → 7.96M

Tasting Notes

✓ Strengths

Macro-F1 0.924 across 18 well-separated intents on 200 held-out cuts (stratified, 88/12 train/eval)
7 of 18 classes hit perfect F1 = 1.000 (cancel_action, control_device, generate_creative_content, greet_or_smalltalk, manage_todo_list, send_message, translate_language)
Trains in ~15 min on a single RTX 5090; data gen ~6 min in Gemini Flash (~$0.40 total)
Handles tricky cancel/undo phrasings ("Wait, scratch that, don't do it") at F1 1.000
Holds JSON output structure consistently across the held-out set

✕ Weaknesses

compute_math vs query_factual_info: "How many feet are in a mile?" routed to query_factual_info (gold = compute_math). The boundary between facts and conversions is genuinely ambiguous
summarize_text recall only 0.60 — a few summarize utterances get pulled to search_web or capture_note (precision still 1.0)
query_factual_info precision 0.733 — it over-attracts on borderline examples (the natural attractor for ambiguity)
Eval set is teacher-derived. v0.2 will add a human-curated 200-utterance hold-out so the proof is not just "agrees with Gemini Flash"

SAMPLE PREDICTIONS

Utterance	Gold	Predicted	Verdict
What's 15% of 250?	`compute_math`	`compute_math`	✓ exact
What time is it in Tokyo right now?	`query_factual_info`	`query_factual_info`	✓ exact
Put on some classical music, something by Beethoven.	`play_media`	`play_media`	✓ exact
Hey there, long time no see! How's your day been?	`greet_or_smalltalk`	`greet_or_smalltalk`	✓ exact
Wait, scratch that, don't do it.	`cancel_action`	`cancel_action`	✓ exact
Tell me a quick, lighthearted joke.	`generate_creative_content`	`generate_creative_content`	✓ exact
How many feet are in a mile?	`compute_math`	`query_factual_info`	✕ wrong

Distillation Run

8 epochs over distilled examples. Final loss: 0.564.

Provenance

Source paper: Routor recipe — Intent Classifier v1

Teacher: gemini-2.5-flash (~1.5T params)

Recipe: routor.intent-classifier-v1.yaml

Distilled: 2026-05-15 by The Distillery v0.1 engine

Author: @house

Download Bottle

Self-contained — model weights, tokenizer, and recipe in a single artifact. No API key needed at inference.

PYTORCH

30.5 MB

For training continuation, HuggingFace, custom Python.

Download

ONNX

14 MB (coming)

CPU/GPU inference, edge, browser via onnxruntime-web.

GGUF

4 MB (coming)

Quantized for llama.cpp, mobile, embedded targets.