language:
- en
- es
tags:
- home-assistant
- structured-output
- instruction-tuning
- ha-action-v3
pretty_name: Home Assistant Requests V3
Home Assistant Requests V3
V3 is a migrated and validated predecessor of Home Assistant Requests V4.
Provenance and attribution
V3 is not entirely synthetic. Accepted rows were derived from:
Rows were treated as raw material, then parsed, normalized, validated against the Home Assistant service schema, and filtered for entity/domain consistency. Project-curated cases were added for missing information, refusal, domain mismatch, and contract coverage.
Accepted split counts:
| Split | Train | Validation | Test |
|---|---|---|---|
| Upstream-derived | 3,236 | 410 | 411 |
| Project-curated | 325 | 4 | 5 |
| Total | 3,561 | 414 | 416 |
Rejected source rows are preserved in rejected.jsonl with their rejection reason instead of being silently repaired.
Relationship to V4
Use V4 for current training and cross-size comparisons. V3 remains public for historical reproducibility and dataset-evolution analysis. Do not mix V3 and V4 test results.
Contract
The task uses constrained JSON responses for Home Assistant actions, with emphasis on valid status decisions, service selection, entity grounding, and safe refusal when context is insufficient.
Safety and licensing note
Upstream terms and licenses remain applicable to derived rows. Review both upstream dataset cards before redistribution or commercial use. Project-curated rows are released for research and reproducibility.
Never execute model output directly. Apply a runtime schema, allowlist, and entity validator before any Home Assistant action.