NQP-Bench
NQP-Bench is a benchmark for next-query prediction in multi-turn conversations. Given the previous user-assistant dialogue history, the task is to predict the user's next query.
This release contains two public subsets:
| Subset | Source | Train | Test | Note |
|---|---|---|---|---|
NQP-Wild | WildChat | 15,683 | 2,329 | public training and test subset |
NQP-Share | ShareChat | - | 1,947 | test-only cross-source subset |
The private subset NQP-Priv is not released.
Data Format
Each line is a JSON object:
{
"session_id": "...",
"source_dataset": "sharechat",
"history": [
{
"turn_idx": 1,
"query": "...",
"response": "..."
}
],
"target": "...",
"difficulty": "Hard",
"intent_transfer_type": "deepening",
"intent_primary": "Reasoning",
"intent_secondary": "Math & Logic",
"target_intention": "...",
"intent_clarity": 3,
"num_turns": 4
}
Fields
| Field | Description |
|---|---|
session_id | Conversation/session identifier. |
source_dataset | Source dataset name, e.g., wildchat or sharechat. |
history | Previous dialogue turns used as input context. |
history[].turn_idx | Turn index in the conversation. |
history[].query | User query in a previous turn. |
history[].response | Assistant response in a previous turn. |
target | Ground-truth next user query to be predicted. |
difficulty | Difficulty label. |
intent_transfer_type | Intent transition type. |
intent_primary | Primary intent category. |
intent_secondary | Secondary intent category. |
target_intention | Short natural-language description of the target intent. |
intent_clarity | Clarity score of the target intent. |
num_turns | Number of turns in the original example. |
License
This dataset is released under a composite license because its public subsets are derived from different source datasets.
NQP-Wildfollows the license terms of WildChat.NQP-Sharefollows the license terms of ShareChat.
Citation
@article{chen2026onepred,
title={OnePred: Next-Query Prediction via Recursive Intent Memory in Multi-Turn Conversations},
author={Chen, Jiangwang and Zhang, Bowen and Song, Zixin and Kang, Jiazheng and Yang, Xiao and Zhu, Da and Jiang, Guanjun},
journal={arXiv preprint arXiv:2605.23668},
year={2026}
}