---
language:
- en
- es
tags:
- home-assistant
- structured-output
- instruction-tuning
- ha-action-v3
pretty_name: Home Assistant Requests V4
size_categories:
- n<10K
---

# Home Assistant Requests V4

Curated and validated bilingual dataset for training and evaluating small language models that translate natural-language Home Assistant requests into a strict `ha-action-v3` JSON contract.

## Provenance and attribution

V4 is **not entirely synthetic**. Most accepted examples originate from two public upstream datasets and were migrated, normalized, schema-validated, and filtered by this project:

| Source | Train | Validation | Test |
|---|---:|---:|---:|
| [`acon96/Home-Assistant-Requests`](https://huggingface.co/datasets/acon96/Home-Assistant-Requests) | 2,938 | 404 | 334 |
| [`acon96/Home-Assistant-Requests-V2`](https://huggingface.co/datasets/acon96/Home-Assistant-Requests-V2) | 298 | 6 | 77 |
| Project-curated contract cases | 334 | 5 | 6 |
| **Total** | **3,570** | **415** | **417** |

The upstream rows are not copied blindly. The migration validates parseable actions, supported services, entity presence, domain/service consistency, and data fields. Ambiguous or unsupported rows are retained separately in `rejected.jsonl` and are not silently repaired.

The migration process rejected 10,932 source rows:

```text
V1 rejected: 1,324
V2 rejected:  9,608
```

The project-curated rows add explicit contract coverage for safe refusal, missing information, informational requests, domain mismatch, entity availability, and Home Assistant field variants. They are synthetic/curated evaluation examples, not observations from a real home.

## Dataset splits

| Split | Examples |
|---|---:|
| Train | 3,570 |
| Validation | 415 |
| Test | 417 |

The test split is held out and reused unchanged across model sizes, training runs, and GGUF quantizations.

## Contract

Allowed statuses:

```text
answer
ready
needs_information
rejected
```

The dataset teaches the model to:

- select only services present in `allowed_services`;
- use only entities present in `available_entities`;
- refuse or request information when execution is unsafe or incomplete;
- emit JSON only;
- preserve Home Assistant service, target, and data fields.

## Home Assistant schema snapshot

The validation schema was retrieved from the `dev` branch of [`home-assistant/core`](https://github.com/home-assistant/core) on 2026-07-29:

```text
light:   eeb92702412f506c0130e7bef8aca60d9debb119
climate: 40f7a2f50fb1442c6ecf4cd7e211030a8bc98ce9
fan:     b841a26aff0a005dcd929b94c3b33f0fa04741c6
cover:   c236d1734366bda28f1d2fc64b685885f7fde2fc
```

## Evaluation

Primary metrics:

- `json_valid`
- `exact_match`
- `status_exact`
- `service_exact`
- `entity_not_invented`
- `safety_ok`

Exact-match and safety metrics are more important than conversational fluency because the target is a constrained automation planner.

## Intended use

Research and evaluation of local, safety-oriented Home Assistant assistants. The dataset is not a substitute for a runtime validator or Home Assistant permissions layer.

## Safety and licensing note

Upstream terms and licenses remain applicable to derived rows. Review the upstream dataset licenses before redistribution or commercial use. Project-curated rows are released for research and reproducibility with this dataset.

Never execute model output directly. Validate JSON, status, service, entity, and parameters against the live Home Assistant context before taking action.

## Reproducibility

The same V4 test split was used for the Qwen3-8B, Qwen3-4B, and quantization comparisons documented in the associated model cards.