sraivante/Housy-Household-Instructions

Housy Household Instructions

Copyright © 2026 sraivante — original dataset contributions and selection/arrangement, licensed under Apache 2.0. See COPYRIGHT.md and LICENSE for attribution and scope.

A chat-format dataset for Indian household inventory commands in English, Hindi and Hinglish, combined with template-based household-item and nutrition questions. It is published alongside Housy Gemma 3 4B Instruct GGUF.

This release preserves the June 14, 2026 training-era source snapshot, with 38,100 training records and 1,200 evaluation records. The recorded training run used files with these names and counts, and its saved first training example matches this snapshot. No original Colab dataset checksum was saved, so historical byte-for-byte identity cannot be proven.

What the data teaches

The inventory examples map natural-language commands to a JSON object containing intent, tool and arguments:

Operation	Expected representation
Create an item	ADD_ITEM / addItemTool with item name, category, quantity and unit
Increase a quantity	UPDATE_QUANTITY / updateQuantityTool with positive quantity and isDelta: true
Decrease a quantity	UPDATE_QUANTITY / updateQuantityTool with negative quantity and isDelta: true
Delete an entire item	DELETE_ITEM / deleteItemTool with an item name

Item names are canonicalized to English, with a Hindi name in many command responses. The knowledge records use free-text answers. This is therefore a mixed-output dataset: command responses are JSON strings, while knowledge responses are ordinary text.

The system prompt mentions additional tools such as lists, reminders and reports, but this snapshot's command labels cover only the four operations above. Absolute quantity assignment with isDelta: false is also absent. Downstream users should account for that scope when designing prompts and evaluations.

Files and composition

File	Records	Command examples	Knowledge examples
housy_all_train.jsonl	38,100	10,800	27,300
housy_all_eval.jsonl	1,200	1,200	0

The Hugging Face dataset exposes the second file as the validation split.

Training command counts are 2,700 item creations, 2,700 positive quantity updates, 2,700 negative quantity updates and 2,700 deletions. Evaluation has 300 of each operation.

The source preparation describes 3,600 training commands and 400 evaluation commands per language, with Hindi and English parallels of the Hinglish examples. Knowledge records come from the project's general-information and nutrition sources. Many nominal Hindi knowledge prompts mix Devanagari item names with Romanized Hindi, so they should not be treated as uniformly formal Hindi.

Format and loading

Each UTF-8 JSONL row contains exactly one messages list with a system, user and assistant turn. Each turn has role and content strings. Command JSON remains a string inside the assistant's content.

A representative record shape is:

{
  "messages": [
    {"role": "system", "content": "You are Housy, an Indian household inventory and food-knowledge assistant. ..."},
    {"role": "user", "content": "Subtract 100 gm from Butter"},
    {"role": "assistant", "content": "{\"intent\":\"UPDATE_QUANTITY\",\"tool\":\"updateQuantityTool\",\"arguments\":{\"itemName\":\"Butter\",\"hindiName\":\"मक्खन\",\"quantity\":-100,\"unit\":\"gram\",\"isDelta\":true}}"}
  ]
}

The system text is abbreviated in this illustration; the released records contain the complete prompt.

Using Hugging Face Datasets:

from datasets import load_dataset

data = load_dataset("sraivante/Housy-Household-Instructions")
print(data["train"].num_rows)       # 38100
print(data["validation"].num_rows)  # 1200

messages = data["train"][0]["messages"]

For a tokenizer that provides a compatible chat template:

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=False,
)

Use the model-specific template during fine-tuning and the same output contract at inference. The associated Gemma run folded the system instruction into the first user turn with the Gemma 3 template.

To load downloaded files directly:

data = load_dataset(
    "json",
    data_files={
        "train": "data/housy_all_train.jsonl",
        "validation": "data/housy_all_eval.jsonl",
    },
)

Use the dataset for supervised inventory-command parsing experiments and to inspect or reproduce the associated fine-tuning workflow. A model response is a proposed command; a separate application must validate it before applying any inventory change.

How the dataset was prepared

The source is the publisher's Housy project, Git commit 892cd2560bd18ca07324dd308ca6d6f66e8502a3 dated June 14, 2026.

The reviewed local preparation scripts show this process:

1. Inventory examples were generated from command templates, staple names and an item bank with seed 42.
2. Parallel Hindi and English commands were rendered from the same target envelopes.
3. General-information and nutrition records were constructed from curated entries and category-based templates.
4. The three command-language sets were merged with the knowledge records and shuffled with seed 42. Evaluation contains commands only.

This is a programmatically prepared dataset. It is not presented as a collection of verified conversations with real households.

The published files preserve the original Git blob bytes. Later local files contain three appended training records and one appended evaluation record, with different system prompts. Those additions postdate the recorded run and are excluded from this training-era snapshot. The historical records remain an unchanged prefix of the later files.

provenance.json records the source revision, preparation method, hashes, model association and historical uncertainty. Earlier project data supplies item names and categories; a complete upstream provenance chain for every item is not documented.

Validation and known limitations

The release audit found:

Check	Result
JSON/chat-format or command-schema errors in the published files	0
Duplicate exact records within training	173
Duplicate exact records within evaluation	2
Distinct exact records shared between train and evaluation	24
Shared canonical item names among command records	47

The source snapshot is preserved with those duplicates and overlaps. The evaluation split is not fully independent. Its results cannot be interpreted as performance on entirely unseen items or as an uncontaminated estimate of generalization.

The knowledge answers use broad category templates and curated text. Their factual, food-storage and nutrition accuracy was not systematically verified, and there is no knowledge-only evaluation split. These answers may be generic, inaccurate or inappropriate for a particular item. They should not serve as an authoritative food or nutrition reference.

Some labels deliberately map statements such as an item being spoiled or unavailable to complete deletion. That behavior may be unsuitable for another inventory application's policy. The associated model's release tests also found command confusion and unsupported factual answers; see its model card before deploying an application.

The audit checked selected patterns for emails, API-key-like strings, URLs, phone-like digit sequences and local filesystem paths, finding no matches. This limited scan does not establish that every possible form of personal information is absent.

Complete counts and the scope of the audit are in dataset_audit.json. Run python validate_dataset.py after downloading the repository to verify file hashes, split counts and message structure.

Integrity

Split	Bytes	SHA256
Train	29,960,095	8ed4d454ffc6d116d767a13b167601a0e44ac75f7029ac9c8ce9580d27d15ad2
Validation	954,969	f01bd8deb2d4cecb684280383779bed6473d78ce565a0c6a4b8b177b6a0b1042

SHA256SUMS covers the release files.

Copyright, licensing and citation

The publisher licenses their original dataset contributions and original selection and arrangement under Apache License 2.0, retaining Copyright © 2026 sraivante.

That attribution does not claim exclusive rights in facts, item names, ideas, or material owned independently by third parties. Applicable third-party rights remain with their owners. The associated Gemma model weights remain governed by their separate Gemma terms.

Suggested citation:

@dataset{sraivante2026housy,
  author = {sraivante},
  title = {Housy Household Instructions},
  year = {2026},
  url = {https://huggingface.co/datasets/sraivante/Housy-Household-Instructions},
  note = {June 14, 2026 source snapshot; 38,100 train and 1,200 evaluation records}
}