Dataset Card for lenamerkli/distilled-web

This dataset consists of web-scraped data using a custom crawler purpose-built for each website.

Dataset Details

Dataset Sources

• Repository: https://github.com/lenamerkli/distilled-web

Uses

This dataset is useful for training large language models.

• The train split provides instruction-following and chat data for supervised fine-tuning (SFT) and instruction tuning.
• The pretrain split provides raw text for pretraining.

Use Recommendations

• Use the train split for instruction tuning or chat fine-tuning of a base model.
• Use the pretrain split for continued pretraining or domain adaptation on Minecraft / Create mod knowledge.
• The ai_enhanced field can be used to filter examples: set it to False to keep only original scraped content, or True to keep only AI-enhanced examples.
• The source field records the original URL of each example for traceability and attribution.

Dataset Structure

Data Fields

The two splits have different schemas.

train split:

Field	Type	Description
messages	string (JSON)	A chat conversation in OpenAI-style format, with system, user, assistant, and tool roles. Content can be multimodal (text and image).
source	string	The URL the example was collected from.
collected_at	timestamp	When the example was collected.
ai_enhanced	boolean	Whether the example was generated or enhanced by an AI model.

pretrain split:

Field	Type	Description
text	string	The raw text.
source	string	The URL the text was collected from.
collected_at	timestamp	When the text was collected.
ai_enhanced	boolean	Whether the example was generated or enhanced by an AI model.

Data Splits

Split	Description
train	Instruction-tuning chat data
pretrain	Raw web-scraped text

Bias, Risks, and Limitations

• Web-scraped content: Web-scraped content may contain inaccuracies, outdated information, or vandalism that was not fully cleaned.
• AI-enhanced data: The AI-enhanced data may introduce hallucinations, stylistic biases, or factual errors. Filter with the ai_enhanced field when needed.
• Language bias: The dataset is predominantly English and German, with limited coverage of other languages.

Citation

If you use this dataset, please cite it as:

@misc{merkli2026distilled-web,
  author = {Merkli, Lena},
  title = {LLMinstruct},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/lenamerkli/distilled-web}
}