AnmolNimmala0/general-language-v1

General Language Dataset: English, Hindi and Hinglish

Overview

general-language-v1 is a multilingual text corpus containing English, Hindi written in Devanagari, and Hinglish written in the Latin script.

The dataset is being collected for research and experimentation involving multilingual language models, small language models (SLMs), tokenizer development, and language-model pretraining.

The collection pipeline combines multiple upstream datasets into a unified JSONL format while retaining source attribution and metadata.

Collection target: 300,000 records.

The collection targets have been reached. Deduplication, language validation, data-quality checks, and licensing review are still required before treating the corpus as training-ready.

Dataset Composition

Language category	Collected records	Script
English	100,000	Latin
Hindi	100,000	Devanagari
Hinglish	100,000	Latin
Total	300,000	Multiple

These are collected-record counts, not verified counts of unique or high-quality examples.

Data Sources

English

• Dataset: HuggingFaceFW/fineweb
• Configuration: sample-10BT
• Split: train
• Text field: text
• Collected records: 100,000

Hindi

• Dataset: HuggingFaceFW/fineweb-2
• Configuration: hin_Deva
• Split: train
• Text field: text
• Collected records: 100,000

Hinglish

Source 1

• Dataset: debajyotimaz/trilingual-hinglish-corpus
• Split: train
• Text field: hinglish
• Collected records: 13,520

Source 2

• Dataset: IB-Emper/hinglish-translit-v2
• Split: train
• Text field: roman
• Collected records: 86,480

Combined Hinglish-category total: 100,000 records.

Romanized Hindi does not necessarily contain English-Hindi code-switching in every sentence. The Hinglish records require language-composition validation.

Data Format

The collection pipeline stores records in JSON Lines (.jsonl) files under data/v2/.

Each line contains one JSON object with the following fields:

Field	Description
id	SHA-256-based record identifier
text	Normalized text
lang	Configured language label: en, hi, or hi-en
script	Expected script: Latn or Deva
is_code_mixed	Source-level code-mixing flag
source_name	Upstream dataset identifier
source_config	Upstream configuration, if applicable
source_id	Upstream identifier or generated row identifier
source_url	Source URL when available
license_status	unverified
content_type	Configured content category
quality_status	pending_review
dedup_hash	SHA-256 hash of normalized text
token_count	Currently null
collected_at_utc	UTC collection timestamp

Collection Methodology

The collector uses the Hugging Face datasets library in streaming mode and uploads batches of JSONL shards to the repository.

The pipeline currently:

1. Streams records from configured datasets.
2. Extracts text from the configured text field.
3. Normalizes Unicode to NFC.
4. Removes null bytes and normalizes whitespace.
5. Filters text by character length.
6. Generates record identifiers and text hashes.
7. Writes accepted records to JSONL shards.
8. Uploads multiple shards per commit.
9. Saves a remote checkpoint to support resuming interrupted collections.

Collection settings used for this version:

• Shard size: 500 records.
• Batch size: up to 10 shards per commit.
• Minimum text length: 40 characters.
• Maximum text length: 100,000 characters.
• Target: 100,000 collected records per language category.

These are ingestion filters and do not constitute comprehensive quality assurance.

Intended Uses

Potential research applications include:

• Multilingual language-model pretraining experiments.
• English and Hindi tokenizer evaluation.
• Romanized Hindi and code-mixed text research.
• Language identification and script classification.
• Multilingual and code-mixed language modeling.
• Dataset provenance, deduplication, and quality evaluation.

The dataset has not yet been benchmarked for a specific downstream task.

Limitations

• Duplicates: Global deduplication has not been verified.
• Language labels: Labels are assigned from source configuration and are not independently verified.
• Hinglish composition: Some Roman-script records may be predominantly Hindi or English rather than genuinely code-mixed.
• Quality: Comprehensive automated and manual quality reviews are pending.
• Privacy: Personally identifiable information has not been comprehensively audited.
• Tokenization: Token counts have not been computed.
• Source distribution: Upstream sources may differ in domain, style, and text quality.
• Licensing: Source-specific license compatibility requires review.

The 300,000-record count should not be interpreted as 300,000 unique, validated, or training-ready examples.

Licensing and Attribution

The combined corpus does not currently have a verified unified license.

Each upstream dataset may have different licensing, attribution, and redistribution requirements. The trilingual Hinglish source is listed under CC BY-NC-SA 4.0, which includes noncommercial and share-alike conditions. Review its terms carefully before commercial use or redistribution.

Users must independently review upstream terms before using the records for training or distribution.

The license_status field remains unverified pending review.

Planned Improvements

• Validate JSONL syntax and shard integrity.
• Measure exact and near-duplicate rates.
• Detect malformed, empty, or corrupted text.
• Validate language and script labels.
• Measure actual Hindi-English code-switching.
• Review source licenses and privacy risks.
• Compute token counts using the intended tokenizer.
• Create reproducible training and validation splits.
• Publish a data-quality report.

Version Information

• Repository: AnmolNimmala0/general-language-v1
• Version: Initial collection, v1
• Format: JSONL shards
• Collection status: 300,000-record target reached
• Data-quality audit: Pending
• Licensing review: Pending

This dataset is a work in progress. Future versions may change as source validation, quality control, deduplication, and licensing reviews are completed.