sraivante/tiny-superfast-agentic-MLM-6.5m-v1-data

Union Command Agent dataset snapshot v1.0.0

Copyright (c) 2026 sraivante. Apache License 2.0 applies to the author's original material and original selection/arrangement. No claim is made over third-party trademarks or other independently owned material. The author confirmed that all bundled A/B and augmentation rows are their original/synthetic work. No external corpus is represented as an independent third-party benchmark.

This release accompanies Union Command Agent: A 6.44M-Parameter English-Hinglish Parser for On-Device Command Execution, Zenodo DOI 10.5281/zenodo.23000171, by Lalit Belwal, Independent Researcher, India. The paired model is sraivante/tiny-superfast-agentic-MLM-6.5m-v1; its weights have SHA-256 30ff9a19e8f007cabcaa2133b61d0a3389adcac16e354763babf874d8d0d74c8. The publication and dataset are versioned as 1.0.0.

Purpose and scope

The corpus maps short English and Roman-script Hindi-English (Hinglish) utterances to one canonical device command. It supports training and evaluating action selection and slot filling for a 371-action catalog. It is synthetic text, not recorded speech, human interaction logs, or an ASR robustness benchmark. The author-created A and B labels identify earlier source collections within this project; they do not identify externally validated datasets.

The snapshot contains 1,230,162 training rows. There are 3,232 rows containing Devanagari characters (approximately 0.26%); no separately validated Devanagari test set is supplied. An exact language percentage for English versus code-mixed Roman-script text has not been measured.

Files and schema

union-train-bundle.zip preserves the inspected archive byte-for-byte. It contains union_train/data/train.jsonl, val.jsonl, test.jsonl, curated.jsonl, and auxiliary files under data/eval/, together with training/runtime code. snapshot-manifest.json records row counts, hashes, action distributions, group prefixes, exact and normalized overlaps, and source-version limitations. The archive also contains a historical cpu_run5_summary.json; it is not the final model's evaluation. The paper's final results are in its separate evidence file and model archive.

Each JSONL record has these fields:

Field	Type	Meaning
text	string	Input command utterance
target	string	Canonical action plus typed slots, with normalized text values
action	string	Action label, normally the first token of target
group	string	Source/template or augmentation identifier; not a guarantee of semantic independence

Illustrative schema example (not asserted to be a specific stored row):

{"text":"volume 40 kar do","target":"set_volume value=40","action":"set_volume","group":"illustrative"}

The catalog, tokenizer, grammar, normalization and target serialization are part of the task definition. Case-sensitive free-text values are normalized in canonical targets and recovered from the original input by the runtime. The encoder limit is 128 normalized characters and the decoder limit is 112 tokens. This audit found 9,510 training rows whose normalized input exceeds 128 characters; historical behavior is preserved and this is an additional truncation limitation.

Splits and saved model results

test_strict and curated_strict are subsets, not additional independent datasets. They exclude group prefixes matching G[57]:para:. The counts below recalculate saved grammar-constrained predictions before runtime repairs; they are not a claim of independently collected test performance.

Split	Rows	Exact correct	Normalized overlaps with current training file
curated	371	369	0
curated_strict	253	251	subset
test	10,618	10,445	0
test_strict	6,611	6,448	subset
val	10,413	10,319	1
A_challenge	1,152	1,122	0
A_practical-dev	1,321	1,259	1
A_practical-holdout	428	394	1
B_golden_dev	383	364	0
B_golden_final	384	374	0
B_typo	3,000	2,975	0
gap10_holdout	1,401	1,401	4
gap11_holdout	696	691	0
wifi_holdout	461	461	1

The main test and curated sets have zero direct normalized-text overlap with the current training file. Eight evaluation rows across five splits overlap after normalization: validation 1, A_practical-dev 1, A_practical-holdout 1, gap10_holdout 4, and wifi_holdout 1. The latter five also overlap verbatim. Shared group identifiers occur for 9,489 validation rows, 2,800 B_typo rows and 159 wifi_holdout rows. Zero duplicated text or distinct group identifiers does not rule out paraphrase/template relatives.

Evaluation scores were observed during model development. These sets must not be described as a newly frozen blind test. Removing the duplicated Wi-Fi and gap10 rows leaves 460/460 and 1,397/1,397 saved exact matches, respectively. A_practical-holdout becomes 393/427 (92.04%). Original splits are retained for traceability rather than silently edited.

Provenance and historical identity

The current training file SHA-256 is 66e358fe595b9c0438aacad2d589c7b18fcff27d8af7eb5ef52058f4cfa49a82. The complete bundle SHA-256 is 71137cf43486761705f108ed08a1959df825ed2b760a987b60621c435494d112. All 12 evaluation files match the model archive's prediction inputs and labels in order. The installed model matches both the scratch-model archive and the public model weights.

The project handover associates this bundle with the final scratch run. However, the checkpoint contains weights and architecture configuration, not a training-file digest, and earlier notes report smaller data counts. Consequently this release identifies the current associated snapshot, while exact historical training-file identity remains unproven. A/B source revisions and all augmentation-generation scripts were not preserved in the available artifacts. File hashes provide a reproducible release snapshot; they do not recreate missing generation history.

This release contains the final scratch model's associated data. Intermediate fine-tune checkpoints and their precise historical data snapshots are unavailable and are not reconstructed or claimed to be included.

Loading and training

from huggingface_hub import hf_hub_download
import json, zipfile

archive = hf_hub_download(
    repo_id="sraivante/tiny-superfast-agentic-MLM-6.5m-v1-data",
    repo_type="dataset", revision="v1.0.0",
    filename="union-train-bundle.zip",
)
with zipfile.ZipFile(archive) as z:
    with z.open("union_train/data/train.jsonl") as f:
        first_row = json.loads(next(f))
        print(first_row["action"])

After extracting the archive into a working directory, the training code is in union_train/. The scratch notebook specifies width 256, 5 encoder layers, 3 decoder layers, 8 attention heads, FFN 768, dropout 0.1, 18 epochs and seed 0 by default. Its high-memory-GPU branch uses batch 1,024, learning rate 0.003 and bfloat16 autocast. AdamW uses weight decay 0.01 and betas (0.9, 0.98), with label smoothing 0.05 and a warm-up/cosine schedule. See the paper and included scripts for the distinction between recorded code settings and historically authenticated run settings. Do not initialize from an intermediate model when attempting the reported from-scratch procedure.

Limitations and intended use

The corpus emphasizes a finite catalog and synthetic phrasings. It does not establish multi-step planning, general conversation, calibrated confidence, safe autonomous execution, speech accuracy, or performance on languages outside the documented scope. Some targets have label ambiguity (for example scan_lan versus discover_devices and bare pin numbering). Apply an independent execution policy and evaluate unfamiliar wording and out-of-scope inputs separately. Example credentials and identifiers are part of the author's synthetic task data; no real account access is required to load or evaluate the corpus.

The dataset, model, paper and reproducibility package are linked through DOI 10.5281/zenodo.23000171. License text and attribution are included. No third-party dataset was identified for exclusion after the author's confirmation.