spitfire4794/Bangla-SFT-50k

Bangla-SFT

Bangla-SFT is an instruction-following dataset containing 50,053 Bengali prompt-response pairs. It was scaled up from a 500-sample seed dataset (spitfire4794/bang_seed).

Dataset Summary

The dataset covers 6 task categories. The prompts are designed to be self-contained (hydrated with appropriate contextual inputs), and the responses are formatted to be direct, omitting conversational prefaces and filler.

• Seed Generation: Baseline instructions generated using Gemini 3 Pro Preview.
• Scaling & Response Generation: Instructions scaled and responses generated using DeepSeek v4 Flash (via Pollinations AI, with reasoning disabled).

Category Distribution

The dataset contains exactly 50,053 rows with the following split:

Category	Count	Percentage	Description
Structured Formatting	12,517	25.01%	Formatting tabular, list, or structured data in Markdown/JSON.
Conversational Chat and QA	10,006	19.99%	Standard open-ended questions, dialogue, and general reasoning.
Summarization and Extraction	10,002	19.98%	Summarizing or extracting entities from Bengali texts.
Editing, Tone and Grammar	7,517	15.02%	Text editing, style transfers, and formal grammar corrections.
Simple Translation / Code-mixing	5,007	10.00%	English-Bengali translation and Banglish code-mixing.
Simple Classification	5,004	10.00%	Categorization, intent tagging, and sentiment classification.
Total	50,053	100.0%

Dataset Properties

File Schema

• hash (string): MD5 hash of the instruction used for deduplication tracking.
• category (string): The classification category of the task.
• instruction (string): The Bengali prompt (includes context/input where needed).
• response (string): The direct Bengali output.

Response Length Statistics

Evaluated using the spitfire4794/beng_bpe tokenizer over a 1,000-row sample:

• Minimum Response Length: 1 token (e.g., direct labels in classification)
• Maximum Response Length: 768 tokens (capped by budget limits)
• Average Response Length: 110.3 tokens

Dataset Creation & Pipeline Details

1. Hydrated Prompts: Instructions requiring input text (such as translation, editing, or summarization) have realistic Bengali context paragraphs generated and appended directly inside the instruction string. This prevents "orphan prompts" like "Summarize the text below" that lack the actual text.
2. No Meta-Completions: System prompts strictly forbid introductory conversational filler (e.g., "নিশ্চয়ই, আমি এটি করে দিচ্ছি" / "Sure, I am doing this") or polite sign-offs. Responses start immediately with the markdown list, table, or code block.
3. Token Budget Enforcement: Responses exceeding 768 tokens were automatically discarded during generation, and the pipeline requested shorter rewrites from the API.
4. Deduplication: Monitored at the hash level. Exact duplicate prompts were discarded on-the-fly.
5. Dynamic Rate Limiting: Managed via an AIMD (Additive Increase, Multiplicative Decrease) controller. Concurrency limits were automatically reduced when CloudFront timeouts (504/502) occurred and slowly restored on successful requests.

Licensing & Attribution

This dataset is licensed under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. Under this license, you may share and adapt the material, but you must give appropriate credit and you cannot use the material for commercial purposes.

Citation

If you use this dataset in your research or non-commercial projects, please credit the creator:

@misc{bangla_sft_2026,
  author = {spitfire4794},
  title = {Bangla-SFT: A Scaled Bengali Instruction-Tuning Dataset},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/spitfire4794/Bangla-SFT-50k}}
}