paperbd/paper_instructions_300K-v1

---
task_categories:
- question-answering
- summarization
- text-to-speech
language:
- en
size_categories:
- 10K<n<100K
---

## Dataset Summary

This dataset contains synthetic supervised fine-tuning data generated from academic papers using `text-albumentations`.

- Rows: 300,000
- Source documents: 1,500 papers
- Format: Alpaca-style instruction tuning rows

Data was synthetically generated using Qwen3.5-4B with `text-albumentations` library (https://github.com/avbiswas/text-albumentations)

Each row is derived from source text through structured augmentations such as:

- bullet extraction
- question-answer generation
- rephrasing
- continuation-style supervision
- comparison and retrieval-style tasks
- knowledge graph triplet extraction

The goal is to turn long-form technical text into diverse, task-shaped supervision for distillation and SFT workflows.

## Supported Tasks

- supervised fine-tuning
- instruction tuning
- distillation

## Data Structure

Each example follows an Alpaca-style schema:

```json
{
"instruction": "string",
"input": "string",
"output": "string"
}
```

## Source Data

The dataset was generated from a collection of 1,500 ML/AI papers. The source material was transformed into synthetic instruction-response pairs through structured augmentation pipelines rather than copied as raw passages alone.

## Limitations

- This is synthetic data, not human-written gold supervision.
- Output quality depends on the underlying model and prompting pipeline used during generation.
- The dataset may contain factual omissions, formatting inconsistencies, or augmentation artifacts.
- Coverage and style are shaped by the source papers and the selected augmentation families.

## License

This dataset is for educational purposes. Please ensure your use is compatible with the licenses and terms of the original source papers.