hisaac617/hearth-lights

Hearth light-control dataset

Prepared for the Hearth on-device tool-calling case study. Derived from acon96/Home-Assistant-Requests-V2.

Generated by prepare_hearth_dataset.py v3.1.1 (schema v3) on 2026-08-27T05:09:08+00:00, seed 42.

Everything is pre-rendered. The consuming notebook does no filtering, sampling, auditing or prompt construction - it loads this dataset and starts modelling.

Splits

Split	Rows	Contents
train	700	600 light requests (stratified by tool) + 100 out-of-scope anchors
validation	100	same mix, smaller
test	300	light requests only, stratified by tool (see below)
out_of_scope	100	frozen non-light requests, for the decline metric
demonstrations	3	one worked example per tool, for the few-shot condition

Columns

Column	Type	Meaning
example_id	str	stable id from the source corpus, or "" if it has none
in_scope	bool	True when the gold answer is a light-control call
user_request	str	the user's utterance
system_prompt	str	fully rendered system prompt (registry + rules)
target	str	supervised target: tool-call block, or the decline sentence
tools	str	JSON string - list of light tool schemas
registry	list[str]	household lights as light.x 'Friendly Name'
registry_ids	list[str]	entity ids only, lowercased, for scoring
gold_tool	str	gold tool name
gold_entity	str	gold entity id, lowercased; "" on out-of-scope rows
gold_arguments	str	JSON string - gold arguments object
prompt_tokens	int	rendered prompt length
near_duplicate	bool	close match to a training utterance (measured on test; False elsewhere)
max_similarity	float	closest training-utterance similarity (measured on test; -1.0 = not measured on other splits)
source	str	provenance field from the source corpus, or "" if it has none

tools and gold_arguments are JSON strings so this loads on any datasets version. Arbitrary-shaped JSON otherwise needs the Json() feature type, added in datasets 4.7.0.

Every split carries every column with the same dtype, which DatasetDict.push_to_hub requires. Columns that are only measured on one split are present on the others with an explicit sentinel rather than being absent.

Usage

import json
from datasets import load_dataset
from transformers import AutoTokenizer

ds = load_dataset("hisaac617/hearth-lights")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct")

row = ds["test"][0]
prompt = tokenizer.apply_chat_template(
    [{"role": "system", "content": row["system_prompt"]},
     {"role": "user", "content": row["user_request"]}],
    tools=json.loads(row["tools"]),
    tokenize=False,
    add_generation_prompt=True,
)

Length gate

Prompts were rendered with Qwen/Qwen2.5-0.5B-Instruct and rows exceeding the budget were dropped at prep time, so no row in this dataset exceeds 2048 tokens including its target.

Split	prompt tokens (min / median / p95 / max)	dropped
train	430 / 713 / 1047 / 1366	0
validation	436 / 701 / 1031 / 1232	0
test	433 / 737 / 1112 / 1370	0
out_of_scope	430 / 709 / 1101 / 1218	0
demonstrations	896 / 976 / 1307 / 1307	0

Prompt length is driven by household size, not by the request. The largest homes produce the longest prompts and therefore the slowest responses.

Design decisions

Tool menu restricted to HassTurnOff, HassTurnOn, HassToggle. A light-control hub should not advertise lock or climate tools to the model at all; the allow-list, not the model's judgement, is what prevents an unsafe action.

Registry trimmed to id + friendly name. Current state (on/off, brightness, colour) is dropped: none of the three tools reads it, since Home Assistant performs the toggle inversion itself. This also makes the prompt static per household, enabling KV-cache reuse. Regenerate if a colour-setting or state-querying tool is added.

Out-of-scope anchors are in train and validation. Without them every gradient step says "emit a light call" and the model's ability to decline collapses. A lights-only validation set cannot detect that failure, which is why anchors appear in both.

test IS stratified by tool, and the natural proportions are recorded separately. The in-scope test pool is heavily skewed toward one tool. Sampling it naturally would produce a test set where a constant predictor scores the majority share on tool_acc, and where the weakest class has too few rows for per-class or confusion analysis to mean anything.

Stratifying costs nothing and buys measurement power on the weak class. The pool's natural proportions are stored in prep_manifest.json under natural_test_distribution, so a traffic-weighted figure can be recovered from the per-class table:

# natural_share values are percentages and sum to 100
traffic_weighted_acc = sum(
    natural_share[tool] / 100 * per_class_acc[tool] for tool in tools
)

The notebook reports this alongside the macro figure. The two diverge when the model is uneven across tools, and that divergence is itself the finding.

Pass --natural-test to sample naturally instead.

Targets use the <tool_call> convention, shared by Qwen and Hermes-style models. A different format can be rebuilt from gold_tool and gold_arguments.

Tool schemas are defined by the preparation script, not taken from the corpus. The hub's API surface is a design-time artefact. This also means the dataset works regardless of whether the source corpus ships a tools column.

Entity arguments are resolved against the household registry. Corpora differ on whether the tool argument is an entity id (light.porch_light) or a friendly name (Porch Light); both are accepted and normalised to the entity id, so gold_entity, gold_arguments and target always agree.

Rows needing arguments beyond name are dropped. The three schemas declare exactly one argument. A request like "set the pantry light to yellowgreen" whose gold answer carries only the entity would train the model to silently discard the colour, and exact_call_acc would reward it for doing so.

English only (unless --keep-all-languages). A sub-1B model split across several languages on a few hundred rows does poorly at all of them.

Filter funnel

How every source row was classified. A very small in_scope share means the filter and the corpus disagree, and nothing downstream should be trusted.

Split	in_scope	out_of_scope	extra_args	unresolved	not_english	no_call
train	50970	38852	0	21014	77856	19194
test	5868	4126	0	3812	11616	5462

Label distribution (in-scope)

test rows are stratified; natural share is the tool's proportion in the underlying test pool, for recovering a traffic-weighted figure.

Tool	train	test	natural share of pool
HassTurnOff	200	100	13.84%
HassTurnOn	200	100	82.96%
HassToggle	200	100	3.2%

Contamination audit

Two independent checks, because they catch different things.

Text overlap

Measured between the sampled train in-scope rows and the frozen test split. Every test row carries its own near_duplicate and max_similarity.

Check	Value
example_id overlap	n/a (source has no example ids)
exact text duplicates	48 (16.0%)
near duplicates (>= 0.90)	60 (20.0%)
median max-similarity	0.772

Household overlap

Text similarity compares utterances. It cannot see a different utterance aimed at the same household — and if a registry appears in both splits, the model can learn its entity names during training and recall them at test time. Every entity metric would be inflated without a single duplicate utterance.

Check	Value
distinct households in train	577
distinct households in test	268
test rows sharing a train household	0 (0.0%)
shared households excluded	False

Both are reported, not corrected for (unless --exclude-shared-households was passed). A high rate on either means accuracy measured on this test set is an upper bound on real-world performance.

Limitations

• Synthetic/template-generated; phrasing diversity is lower than real user speech.
• Entity ids are household-specific and supplied in the prompt, so generalisation to new naming conventions is untested.
• Out-of-scope anchors cover unsupported domains only - not ambiguity, multiple matching devices, missing devices, or multi-action requests.
• Train/test independence is enforced only where the audits above report it. Households shared between splits are measured, and removed only with --exclude-shared-households.
• Inherits the source corpus's licence and biases.