mbotta/trippet-lfm2-1.2b-edit-ONNX

trippet-lfm2-1.2b — unified on-device model for the trippet trip-planning app

A LoRA fine-tune of LFM2.5-1.2B-Instruct, exported to q4 ONNX for in-browser inference (transformers.js + WebGPU). It is the single model that powers the whole trippet app — trained, exported, and validated end-to-end on a MacBook Pro (M2 Max), no cloud.

What it does (one model, routed by context)

It uses LFM2's native tool-calling as a mode switch. The app declares the extract_intent tool when parsing a trip brief; otherwise the model just talks.

job	how it's invoked	output
Parse a trip brief → structured Intent	declare extract_intent tool	native tool-call with the Intent (region, bases+roles+nights, intermediates, certainty, when/duration, party, interests, constraints, asks)
Select day stops from a candidate pool	the selection prompt (no tool)	{interestId:[indices]} — the strongest few per interest
Converse (narrate the plan, confirm an edit, answer a question)	a chat/draft prompt (no tool fires)	grounded plain prose

The native format is what makes one model possible: the tool declaration is a learned switch, so a model fine-tuned to tool-call on briefs still chats on everything else — no second model, no catastrophic forgetting.

Training (all local, Apple Silicon)

• PEFT LoRA (r=16, all-linear, prompt-masked), 3 epochs, MPS, merged to fp16.
• Data: ~1000 valid-by-construction brief→Intent examples (native tool-call targets) + a boundary set (chitchat/draft → prose, teaching when not to call) + selection examples distilled from a strong teacher over real candidate pools (the harness's own gatherCandidates over real towns).
• Export: LiquidAI's onnx-export → q4 ONNX (transformers.js layout).

Validation (fp16 merged, same scorer as the baseline)

• Routing: 7/7 — tool-calls on briefs, chats on chitchat ("Rome, Florence, Venice"; "Paris").
• Intent — in-distribution: slot-F1 1.000, schema-valid 40/40 (perfect structure).
• Intent — real briefs: slot-F1 0.636, schema-valid 6/10 (real/long/multi-constraint is the hard case).
• Selection vs teacher gold: pick-F1 0.915, coverage 1.0, set-Jaccard 0.92, 10/10 in-bounds.
• vs base LFM2.5-1.2B: the base can't produce the Intent at all (recall 0.00) and over-triggers/garbles tool-calls; this model is grounded and routes.

Use (transformers.js)

import { pipeline, env } from '@huggingface/transformers';
const gen = await pipeline('text-generation', '<this-repo>', { dtype: 'q4', device: 'webgpu' });
// parse a brief — declare the tool:
const out = await gen([{ role:'user', content: brief }], { tools: [EXTRACT_INTENT_TOOL], max_new_tokens: 400 });
// converse — no tool declared:
const reply = await gen([{ role:'user', content: draftPrompt }], { max_new_tokens: 300 });

Limitations

• Real, long, multi-constraint briefs still degrade (the Puglia-class stress test) — the gap is training data coverage (more long/messy/real briefs), not the method.
• Selection is trained on a modest teacher-labeled set; grow it for production.
• Romanization of non-Latin place names is better handled by a deterministic library than the model.
• The harness must declare the tool when parsing briefs and parse the native call (small change).