Flipping the Dialogue: Training and Evaluating User Language Models
Paper • 2510.06552 • Published • 2
wildchat-intents-qwen3-32b
WildChat-1M conversations with user-intent summaries extracted by
qwen3-32b (temperature 0.7, top-p 0.9, max_tokens 16384), following the
intent-extraction pipeline of "Quantifying the Utility of User Simulators
for Building Collaborative LLM Assistants"
(https://github.com/schang-lab/utility-of-user-simulators), with the
intent prompt of the UserLM paper (Naous et al., arXiv:2510.06552).
Each row is the original allenai/WildChat-1M
record plus an intent field holding the extractor's raw reply (including
the model's <think>...</think> block; the intent summary follows it).
Conversations whose extraction prompt exceeded the serving context window
were dropped.
| split | rows |
|---|---|
| train | 270,067 |
| val | 14,737 |
| test | 22,124 |
Source data is subject to the AI2 ImpACT License (Low Risk Artifacts) of WildChat-1M.
- Downloads last month
- 7