tuxevil's picture
Update README.md
3f33565 verified
|
Raw
History Blame Contribute Delete
4.62 kB
---
license: other
license_name: mixed-mit-apache-2.0-cc-by-4.0
task_categories:
- text-generation
language:
- multilingual
tags:
- home-assistant
- tool-calling
- function-calling
- smart-home
- qwen3
pretty_name: Home Assistant Requests V5.1 Native Strict
size_categories:
- 1K<n<10K
configs:
- config_name: default
data_files:
- split: train
path: train.jsonl
- split: validation
path: validation.jsonl
- split: test
path: test.jsonl
---
# Home Assistant Requests V5.2 Native Strict
Private research dataset for supervised fine-tuning and regression testing of a small Home Assistant native tool-calling model.
Contract: `ha-native-tool-calling-v2`.
## Frozen snapshot
| Split | Rows | Direct speech | Multi-call | Maximum rendered tokens |
|---|---:|---:|---:|---:|
| train | 3,806 | 340 | 78 | 3,098 |
| validation | 530 | 52 | 4 | 2,874 |
| test | 633 | 102 | 22 | 2,925 |
Tokenizer audit:
- model: `unsloth/Qwen3-4B-Instruct-2507`
- revision: `992063681dc2f7de4ee976110199552935cad284`
- `MAX_SEQ_LENGTH=4096`
- overlength rows: `0`
- render failures: `0`
All 20 Home Assistant native tools are represented. Test contains 52 `GetLiveContext` cases, at least 20 positive cases per tool, 102 direct/no-tool cases, and 22 multi-call cases.
## Files
- `train.jsonl`, `validation.jsonl`, `test.jsonl`: frozen training/evaluation splits.
- `metadata.json`: sources, coverage, tool counts, and readiness report.
- `audit.json`: contract, duplicate, leakage, entity-cap, secret, and hash audit.
- `token_audit.json`: full tokenizer/template length and supervised-token audit.
- `provenance.json`: immutable revisions, input/output hashes, code hashes, and exact build command.
## Sources and licensing
This is a mixed-license derivative dataset. Preserve source attribution and applicable notices when redistributing derivatives.
1. `tuxevil/Home-Assistant-Requests-V5-Native`, revision `fbf847bb63cfb2e49a36435412d610ac8993e8fc`, derived from `acon96/Home-Assistant-Requests-V2` — MIT.
2. `enfuse/joint-intent-slot-smarthome`, revision `39056ca057ae92d71fb7a272f7960567e55dca70` — Apache-2.0. Used in train only through a deterministic strict mapper.
3. `home-assistant/intents`, revision `35466626accc5699203fc20a1af470ff173a7893` — CC-BY-4.0. Attribution: Home Assistant intents contributors.
4. Deterministic internal deficit-only fixtures generated for missing readiness classes. No teacher LLM or paid API generated these fixtures.
The repository-level metadata uses `license: other` because no single SPDX license describes the combined corpus.
## Construction policy
- Maximum eight exposed entities per row.
- Tool names and schemas are frozen from Home Assistant Core branch `2026.7.4`.
- Tool arguments must satisfy their supplied schema and refer only to exposed entities/areas.
- Invalid, hidden-target, conflicting, unsupported, or non-native examples are rejected.
- Enfuse data is train-only.
- Official Home Assistant conformance rows are allocated test first, then validation, then train, without row reuse.
- Final global isolation removes normalized user-text and structured semantic-signature collisions in priority order: test → validation → train.
- Coverage fixtures only fill measured deficits and receive unique split-specific utterances.
## Verified gates
`audit.json` reports:
- contract errors: 0
- duplicate example IDs: 0
- cross-split semantic overlap: 0
- cross-split normalized user overlap: 0
- secret-pattern hits: 0
- maximum exposed entities: 8
- training readiness: true
`token_audit.json` reports zero overlength rows and zero render failures over all 4,969 rows.
## Limitations
- Split isolation prevents leakage created by this build. It cannot prove that a pretrained base model never saw public Home Assistant or Enfuse text during pretraining.
- Official conformance sentences are useful deterministic holdouts, not a guarantee of out-of-distribution generalization.
- Deficit fixtures are synthetic deterministic coverage data and may be stylistically narrower than real voice traffic.
- Passing dataset gates does not promote a model to production. Adapter, merged FP16, Q5, Ollama, and physical Home Assistant smoke gates remain mandatory.
- Q5 artifacts produced from this dataset remain `research-only-pending-ollama-gate` until tested against an authorized physical entity.
## Reproducibility
See `provenance.json` for exact revisions, SHA-256 hashes, code hashes, tokenizer versions, and complete build command. Consumers should pin the Hugging Face dataset repository to an immutable commit SHA rather than `main`.