README.md

---
language:
- en
pretty_name: PACE
task_categories:
- question-answering
tags:
- benchmark
- retrieval
- reasoning
- personalization
- conflict-aware
size_categories:
- 100K<n<1M
configs:
  - config_name: queries
    data_files:
      - split: test
        path: data/queries.parquet

  - config_name: corpus
    data_files:
      - split: test
        path: data/corpus.parquet

  - config_name: profiles
    data_files:
      - split: test
        path: data/profiles.parquet
---

# PACE 🏃‍♂️

## Dataset Description
- **Repository:** [Code](https://github.com/p2chp2t/pacemaker)
- **Paper:** [PACE: Towards Surfacing Hidden Conflicts in User Requests](https://arxiv.org/abs/2609.03293)
- **Point of Contact:** [Yoojin Kim](mailto:kimyujin1224@postech.ac.kr)

## Dataset Summary
PACE (**P**ersonalized **A**ssistants for **C**onflict **E**valuation) is a benchmark for evaluating whether personalized assistants can identify situational conflicts between user requests and facts stored in a user-centric knowledge base.

PACE contains 3,249 queries across 185 user-specific cases, together with 376,448 atomic facts. Each query is labeled as either `conflict` or `non_conflict` and is associated with gold evidence from its corresponding knowledge base.

## Dataset Structure
PACE is organized into three configurations that together represent each user-specific case:
- **Queries** contains the user requests to be evaluated, along with their feasibility labels, reference judgments, and gold evidence document IDs.
- **Corpus** contains the atomic facts that form each user-specific knowledge base. It includes both gold evidence and distractor documents used for retrieval.
- **Profiles** contains the structured user profiles used to construct each case, including user attributes and information about relevant related individuals.


### Queries
| Field | Description |
| --- | --- |
| `case_id` | Identifier of the user-specific case / KB |
| `window_start` | Start of the temporal context window |
| `ref_date` | Reference date for the case |
| `window_end` | End of the temporal context window |
| `query_id` | Query identifier within the case |
| `situation_type` | Situation category (`temporal`, `personal`, or `state`) |
| `label` | Feasibility status (`conflict` or `non_conflict`) |
| `query` | User request |
| `context` | Context used during dataset construction |
| `judgment` | Reference explanation for the feasibility decision |
| `gold_doc_ids` | IDs of the gold corpus documents needed to resolve the query |

### Corpus
| Field | Description |
| --- | --- |
| `case_id` | Identifier of the user-specific case / KB |
| `window_start` | Start of the temporal context window |
| `ref_date` | Reference date for the case |
| `window_end` | End of the temporal context window |
| `doc_id` | Document identifier |
| `text` | Atomic fact text |
| `situation_type` | Situation category associated with the source query |
| `label` | Feasibility status associated with the source query |
| `kind` | `gold` or `distractor` |

### Profiles
| Field | Description |
| --- | --- |
| `case_id` | Identifier shared with the `queries` and `corpus` subsets |
| `persona_id` | Persona identifier from the original profile |
| `seed` | Short profile summary |
| `spec` | Nested structured profile information |
| `relationships` | Nested relationship information for relevant alters |

## Citation
Please cite our work if you find the resources in this repository useful:
```bibtex
@misc{kim2026pacesurfacinghiddenconflicts,
      title={PACE: Towards Surfacing Hidden Conflicts in User Requests}, 
      author={Yoojin Kim and Jihyoung Jang and Hyounghun Kim},
      year={2026},
      eprint={2609.03293},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2609.03293}, 
}
```