nohaah/laya-home-fine-tune-en

Gantry decision model (English)

A fine-tuned decision model that turns a spoken or typed home command ("turn off the kitchen lights") into a device action. Instead of generating text, it answers a multiple-choice question (which action, which room, which door) with a calibrated probability for each option, in a single forward pass. So it can't produce a malformed command, and the hub can ask you to repeat when it isn't sure.

421M parameters (ModernBERT-large encoder, 395M, plus the Laya decision head). In Gantry it answers English requests; Korean requests go to the bilingual model, nohaah/laya-home-fine-tune-kr.

Part of Gantry, a fully offline voice assistant for the home that runs on a Raspberry Pi (source: github.com/noahhps/gantry).

Files

File	What it is
laya-home.int8.onnx	int8 ONNX export: what Gantry runs (598 MB)
model.safetensors	PyTorch weights, fp16 (843 MB)
rl_agent_config.json, encoder/config.json, tokenizer/	Model configuration and tokenizer
questions.json	The question format it was trained on
train_report.json	Training settings and held-out results before and after fine-tuning
SHA256SUMS	Checksums for every file

Training

• Base model: convaiinnovations/laya, encoder answerdotai/ModernBERT-large (the "typed-decisions" checkpoint).
• Data: 4,000 synthetic English commands, generated by training/generate_data.py: 5,751 training questions after splitting each command into its action, room and door questions. The data covers varied home layouts, follow-ups, and requests that aren't device commands. Published as nohaah/laya-home-fine-tune-en-training-set.
• Settings: 1 epoch, effective batch 16 (8 × 2 gradient accumulation), learning rate 2.5e-5 (encoder) and 1e-4 (head), label smoothing 0.1, seed 42, Laya's rlcd loss, on an Apple M1 Pro (MPS).
• Calibration: temperature fitted on 400 held-out training items, so the probabilities can be used as confidence.

How to use

The model answers one multiple-choice question at a time and returns a calibrated probability for every option. It is loaded with the laya library from convaiinnovations/laya.

from huggingface_hub import snapshot_download
from laya.onnx_agent import ONNXAgent

snapshot_download("nohaah/laya-home-fine-tune-en", local_dir="gantry-laya-en")
agent = ONNXAgent("gantry-laya-en", onnx_path="gantry-laya-en/laya-home.int8.onnx")
answer = agent.predict(
    'Latest request: "turn off the kitchen lights"',
    {"action": {"type": "choice", "instructions": "Which action?",
                 "criteria": ["turn_on_lights", "turn_off_lights", "set_temperature", "raise_temperature",
                              "lower_temperature", "open_garage", "close_garage", "lock_door", "unlock_door", "none"]}},
)["answers"]["action"]
print(answer["choice"], answer["probabilities"])

The PyTorch weights (model.safetensors) load with laya.load("gantry-laya-en").

Input format

The input describes the request, and for follow-ups the earlier commands, exactly as Gantry's hub writes it:

Earlier requests, oldest first:
- "turn on the kitchen lights" -> turn_on_lights {"room":"kitchen"}
Latest request: "turn them off"

It was trained on three short questions with bare option ids:

Question	Options
Which action?	turn_on_lights, turn_off_lights, set_temperature, raise_temperature, lower_temperature, open_garage, close_garage, lock_door, unlock_door, none
Which room?	all plus the home's rooms (English ids such as kitchen, living room)
Which door?	the home's doors (English ids such as front, back)

In Gantry, services/laya_server.py serves this folder to the hub: python services/laya_server.py --checkpoint gantry-laya-en --onnx gantry-laya-en/laya-home.int8.onnx.

Evaluation

Measure (held-out phrasings never seen in training)	Result
All questions (action, room, door), fp16, before → after fine-tuning	87.8% → 97.9% correct (189 questions)
Calibration error (ECE), before → after	0.184 → 0.081
Action question, int8 ONNX as deployed	95.2% correct (126 requests)
Speed, int8, Raspberry Pi 4, 4 threads	~1.1 s per question
Speed, int8, Apple M1 Pro, 4 threads	~68 ms per question

Longer input

Single held-out requests with 0–3 unrelated earlier commands added in front, int8 ONNX, Apple M1 Pro, 4 threads:

Language	Earlier commands in input	Tokens	Correct action	Time per question
English	0	14	96.9%	68 ms
English	1	43	86.5%	93 ms
English	2	60	88.5%	108 ms
English	3	82	83.3%	129 ms

Limitations

• Synthetic data. Training and test requests are generated from templates (training/generate_data.py), not recorded from people. Accuracy on real speech, through a speech-to-text model, will be lower.
• A narrow job. It chooses among a smart home's actions, rooms and doors. It isn't a general language model and won't answer questions.
• Unrelated context hurts. Accuracy drops when unrelated earlier commands are included (see the table above), so Gantry only sends history for follow-ups.
• Not a safety mechanism on its own. Gantry requires 80% confidence for unlocking doors and opening the garage and asks again below that, and the panel asks for confirmation. Keep a check like that around any action with consequences.

License

Apache-2.0, the same license as the base model convaiinnovations/laya and its encoder answerdotai/ModernBERT-large. See LICENSE and NOTICE.