Wild Feedback Dataset With 186k Ordinal Satisfaction Scores

WildFB contains 186,000 instances of human-LLM conversational interactions, each labeled with a 4-level ordinal satisfaction score. The dataset was created by THU-KEG to train reward models using implicit feedback signals extracted from user follow-up queries.

Use Cases

• Train a reward model to predict the 4-level ordinal satisfaction score from conversational text.
• Analyze patterns in user follow-up queries to identify implicit feedback signals for LLM alignment.
• Fine-tune a language model using the satisfaction score labels as a supervision signal for conversational quality.

Strengths

• Contains 186,000 high-quality instances filtered from a larger source.
• Provides a 4-level ordinal satisfaction score for each conversational instance.
• Derives implicit reward signals from real-world human-LLM interactions.

Limitations

• Specific column names and data structure are not provided in the input.
• The dataset's geographic and temporal coverage are unknown.
• The original source, WildChat-4.8M, may introduce biases from its in-the-wild collection.

Provenance

Source

THU-KEG

Collection Method

Filtered and refined from the WildChat-4.8M dataset, extracting implicit reward signals from user follow-up queries.

Time Range

null

Freshness

Last updated February 2026.

Geography

null