Ultrachat Conversations With Filtered Bot Responses

Ultrachat-Uncensored modifies the original Ultrachat dataset by removing bot messages containing specific terms listed in a filters.txt file. Author branles14 created this variant in July 2023 to promote more neutral and constructive chatbot interactions. The dataset is hosted on Hugging Face with an estimated size between 100K and 1M text examples.

Use Cases

• Training language models to avoid generating responses containing the filtered terms from filters.txt.
• Analyzing the impact of term removal on conversation flow and bot neutrality.
• Comparing dialogue quality between the original Ultrachat and this filtered variant.
• Benchmarking safety filters for chatbot outputs using the specified exclusion list.

Strengths

• Derived from the established Ultrachat dataset, providing a known base.
• Explicit filtering process documented via a filters.txt file.
• Hosted on a major platform with library support for Polars, Dask, and Datasets.

Limitations

• Exact row count, column names, and sample data are unavailable for assessment.
• Filtering criteria are opaque without reviewing the specific terms in filters.txt.
• Dataset may inherit any biases or limitations present in the original Ultrachat source.

Provenance

Source

Variant of the Ultrachat dataset, hosted by user branles14 on Hugging Face.

Collection Method

Created by programmatically filtering the original dataset based on a list of terms.

Freshness

Last updated July 2023.