kami/Maven

Files

T

kami 0c65387a5f feat(router): array route contract + shorter RU router prompt

Route contract is now a JSON array of action objects (one per ask) so
compound utterances route all their intents, not just the first. Grammar
root emits `[{intent...},...]`; parseActions tolerates a bare object.
Cascade still returns one Decision — full N-action dispatch lands with the
engine turn-on (marked in-code).

Router prompt rewritten shorter + decision-ordered (prompt-guy feedback),
fact redefined as "implicit update" not "trackable state", kept in Russian
to match the CPT base + phraser. "интент" → "намерение".

CLAUDE.md: routing-architecture section + refreshed open items.
docs/plans: route-data generation plan.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017GMrVfuYN3nE4L1vEiFYC9

2026-07-11 23:27:44 +04:00

2.2 KiB

Raw Blame History

Route-training data plan (LLM-as-router)

Goal: training data that teaches the CPT'd Qwen3-1.7B to emit the route contract — {"intent":<enum>, key?, value?, text?, verb?}, GBNF-constrained — matching internal/router/llmrouter.go (routeSystem + routeGrammar) verbatim. Train=deploy parity: label with the EXACT prompt the daemon sends.

7 intents: fact, reminder, note, query, act, chat, system. Hard cases (grammar can't enforce): note vs reminder («запомни» vs «напомни»), fact vs note (trackable state vs static memo).

Steps

Source utterances — real RU turns, not synthetic. Main: function_calling.jsonl (473); plus user-*.jsonl fragments for chat/fact/system coverage. gen_route_data.py dedups across all.
Relabel, don't convert — old taxonomy (time/weather/timer) ≠ 7 intents. Feed each utterance through routeSystem to a strong router model → take its {intent,...}.
Validate — intent ∈ enum, keys ⊆ {intent,key,value,text,verb}. Drop invalid.
Balance check — after a run, count intents. act/system/fact likely thin (function_calling skews query/act). Author extra examples for the holes; re-run.
Better prompt first — improve routeSystem for sub-1B disambiguation before a big generation run (awaiting prompt-guy input). Re-labeling is cheap; regenerate.
Train — route-LoRA on top of CPT base, OR fold into the persona SFT as a second contract (decide once volume known). Eval = route accuracy on a held-out REAL set.

Blocking

Router at inference.kvmx.ru / localhost:6446 must be up (currently down).
CPT must finish before the route-LoRA trains on it.

Files

esp32-whisper-fine-tune/llm/gen_route_data.py — the relabeler (done, self-checks). ROUTE_SYSTEM const = verbatim copy of Go routeSystem; keep in sync.
Output: llm/data/route_train.jsonl (resumable append).

Prompt-guy question (sent 2026-07-11)

How to structure the router system prompt for a sub-1B model doing 7-intent classification + slot extraction, GBNF-constrained — example ordering/count, contrastive near-miss pairs (note vs reminder) vs more singles, rule placement.

2.2 KiB Raw Blame History