sraivante/home-commands-json-v1

Home Commands JSON v1

English and Romanized Hindi/Hinglish home commands mapped to structured JSON. This release contains the exact source and processed snapshot associated with Superfast Tiny Home Robotics JSON 1M v1. It also includes a separately authored evaluation challenge set. Native-script Hindi appears only in a small later diagnostic, not as a supported training language claim.

Publisher: sraivante. Release date: 2026-09-25. Version: v1.0.2.

Configurations and counts

Configuration / split	Rows	Role
original / source	3,458,816	Original received JSONL, losslessly gzip-compressed
prepared / train	910,114	Actual training examples
prepared / validation	103,089	Validation split
prepared / test	104,723	Original held-out test split
robotics_edge_v1 / test	284	Later stress-test cases; never used to train this model

The source split is provenance data, not an additional training split. Do not combine it with the prepared splits: it contains their source examples and many duplicates. The prepared configuration is the default.

Load the data

from datasets import load_dataset

repo = "sraivante/home-commands-json-v1"
data = load_dataset(repo, "prepared", revision="v1.0.2")
print(data["train"][0])

# For a small initial download:
stream = load_dataset(repo, "prepared", split="train", streaming=True, revision="v1.0.2")
print(next(iter(stream)))

challenge = load_dataset(repo, "robotics_edge_v1", split="test", revision="v1.0.2")
source = load_dataset(repo, "original", split="source", streaming=True, revision="v1.0.2")

JSONL files are gzip-compressed without changing the decompressed bytes. Each line is one UTF-8 JSON object. Download and decompress directly if not using the Hugging Face Datasets library.

Schema

Original source:

{"instruction":"Please turn on the fan","output":{"activity":"air","subject":"fan","action":"ON"}}

This illustrates the schema rather than asserting the example's source line. The prepared files retain instruction and the object-valued output, and add:

• group_id: canonical phrase-family string used for split assignment.
• source_line: one-based line number of the retained example in the original decompressed JSONL.

Actions are ON, OFF, STATUS. There are 30 observed combinations across fan, light, motor, speaker, geyser, cooler, washing_machine, tv, ac, and sprinkler. motor refers to the water pump. Full valid combinations and the field schema are in metadata/allowed_outputs.json and metadata/output_schema.json.

Challenge cases have id, category, text, nullable expected JSON, expected_behavior, scope, reason, and overlap-audit fields. expected=null means no single supported command is acceptable: the declared policy calls for abstention/clarification. It does not mean the trained model supports a reject class. The four input-guard cases include empty/overlong text intentionally.

Preparation and split construction

1. Audit JSONL records, output schema and conflicting normalized inputs.
2. Normalize input identity with Unicode NFKC, whitespace normalization and case-folding. Remove 2,340,890 duplicate normalized-input rows. Retain one source example per distinct normalized input, leaving 1,117,926 examples.
3. Group phrase families after removing punctuation, common greetings/politeness prefixes and suffixes, and the article the. Stratify family assignment by the 30 target commands, with approximately 80/10/10 family proportions and seed 20260925. This produces 2,292 / 288 / 288 train/validation/test families.
4. Fit the byte-level BPE tokenizer only on training examples; serialize input and JSON response with explicit role tokens. The training objective includes input and response next-token targets, excluding padding and initial BOS.

Training contains 22,655,302 prediction tokens for one pass. Only 4,120 training examples have STATUS targets, versus 905,994 ON/OFF examples; command imbalance and heavy templating should be considered when evaluating models.

The preparation code is in preparation/. The exact tokenizer, token-ID memmaps, indexes and normalized splits used by training are also included in artifacts/home_commands_training_bundle.zip for reproduction.

Evaluation scope and known limitations

The associated model reached 100% exact JSON accuracy on the original 104,723 test rows, but only 68/140 (48.6%) on later supported-command challenges and 48/120 (40%) on those without source-phrase overlap. This is evidence that the original split is not a sufficient real-world generalization test.

The later 284-case suite was frozen before model inference. Twenty cases overlap an original source phrase (14 train, 3 validation, 3 test); the remaining 264 do not. It is a small assistant-authored diagnostic with correlated wording, not a probability sample, recorded-speech dataset or robotics safety benchmark.

The training output space has no reject/no-op class, negation semantics, device instances, speed/position parameters, schedules or multi-step commands. The associated model emitted recognized commands for all 130 requests requiring abstention, including 33 ON/OFF commands. Do not interpret valid JSON as safe actuation. Additional human-written data and independent testing are needed for practical command systems.

Snapshot identity

These are SHA-256 hashes of the decompressed original bytes:

File	SHA-256
Original source	10822f6617d2cc94dca66d609a1f41769cd5ef2ebba224d85fc7ae79cfbbaeba
Train JSONL	63313b7aafeb2a2a32e87c660015d745a84698a0db1c23e83c8da94105547430
Validation JSONL	df6821e7165af193a09b6790de530c83ef9466d25b4a7e01d25f5b392ec3b580
Test JSONL	a43dc4afaa386e5d4825e26b495159b0beace648488c4d843f153c650adfaa06
Later challenge JSONL	30cb9a6de18cc12cefcbb6e1aac8a29c7371b0f4ea537e516f38d02b3f417d6c

training_snapshot_manifest.json is the byte-identical manifest used by the run. release_manifest.json additionally hashes compressed/public artifacts and links the model version. The historical manifest's private/pre-authorization note describes its creation state; this release follows the publisher's later explicit authorization. No source JSONL or split content has been edited.

Provenance, license and attribution

The source file was supplied by sraivante for this project. Its upstream generation script, historical repository revision, and original creation date were not provided or independently verified. No external corpus was added in preparation. The exact file hash identifies the data actually used by the run; the release date is not a claim about the original source creation date.

Copyright (c) 2026 sraivante, limited to original contributions and original selection/arrangement, under Apache License 2.0. See LICENSE and NOTICE. Third-party ownership, attribution and terms remain unchanged. The material supplied to this project did not identify third-party source repositories; the original-contribution license does not purport to relicense any unidentified third-party material.

The later diagnostic was assistant-authored during evaluation for the publisher and is separately versioned and marked in the manifest. It was not mixed into the training snapshot. Models or revisions trained using this challenge must use a fresh independent evaluation set for new performance claims.

Raspberry Pi evaluation snapshot — 2026-09-26

Measurement/data revision: edge-rpi5-20260926.1. The original training/source files are unchanged, pinned at 8479d12161767c5aabd7dfc097e6f6c41790201a. No training or fine-tuning was performed for this update.

The edge_benchmarks/rpi5-2026-09-26 directory adds 284 quality cases, 10 latency inputs, and all quality predictions from the corresponding published model on a Raspberry Pi 5 (16 GB, aarch64). The complete model card reports the measurements.

Evaluation role: existing frozen stress/regression suite, constrained decoding only. Previously inspected synthetic suite; not a new blind holdout. Original per-case source-overlap metadata retained. The Pi result was 72/284, with the per-category breakdown in the linked model report. This data does not establish fresh blind or real-world accuracy.

quality_cases.jsonl and latency_cases.jsonl contain id, input text, serialized expected_json (null means reject/guard), scoring match (exact, subset, rejection), and category. quality_predictions.jsonl adds serialized output_json, correct, error, elapsed_ms, and post-inference cpu_temperature_c. Exact adapter inputs and original metadata are preserved in cases/*.json. provenance.json records counts, SHA-256, original source revisions and historical identity limitations. Keep this diagnostic separate from training and the original evaluation configurations.

Copyright (c) 2026 sraivante applies only to original benchmark/evaluation material and original selection/arrangement, under Apache License 2.0. Third-party ownership, attribution and licenses remain unchanged.