| ---
|
| license: cc-by-4.0
|
| language:
|
| - en
|
| configs:
|
| - config_name: Temporal
|
| data_files:
|
| - split: default
|
| path: Temporal/*.json
|
| - config_name: Invariant
|
| data_files:
|
| - split: default
|
| path: Invariant/*.json
|
| task_categories:
|
| - question-answering
|
| - text-generation
|
| tags:
|
| - circuit
|
| - temporal
|
| - knowledge
|
| - triplet
|
| size_categories:
|
| - 10K<n<100K
|
| ---
|
| |
| # \[ACL 2025\] Does Time Have Its Place? Temporal Heads: Where Language Models Recall Time-specific Information |
|
|
| <center><img src = "https://cdn-uploads.huggingface.co/production/uploads/5efbdc4ac3896117eab961a9/yzXDiuGZUHCaVkERZFSIO.png" width="1000" height="1000"></center> |
|
|
| **This repository contains two separate subsets of data (configs):** |
| - **Temporal**: JSON files in `Temporal` that include temporal knowledge. |
| - **Invariant**: JSON files in `Invariant` that describe time-invariant knowledge based on [LRE](https://arxiv.org/abs/2308.09124). |
|
|
| Each subset has its own schema. By defining them as two configs in the YAML header above, Hugging Face’s Dataset Viewer will show **“Temporal”** and **“Invariant”** as separate options in the configuration dropdown, allowing you to explore each schema independently without a schema‐mismatch error. |
|
|
| --- |
| ## Dataset Overview |
|
|
| **Motivation:** |
| Large language models (LLMs) often struggle to answer questions whose answers change over time. We investigated whether there exist specialized attention heads—**Temporal Heads**—that are triggered by explicit dates (e.g., “In 2004, …”) or by implicit textual cues (e.g., “In the year …”) and that help the model recall or update time-specific facts. |
|
|
| - **Method:** Using [Knowledge Circuit](https://arxiv.org/abs/2405.17969) analysis, we identified attention heads in LLMs that strongly activate on temporal signals (timestamps, years, etc.). |
| - **Findings:** These Temporal Heads are crucial for time-sensitive recall. When you ablate (disable) them, the model’s performance on time-dependent questions degrades significantly, whereas its performance on static (time-invariant) knowledge remains almost unchanged. |
| - **Implications:** By manipulating the outputs of these specific heads, one can potentially edit or correct a model’s temporal knowledge directly (e.g., if its internal knowledge about “Who was president in 1999?” is outdated). |
|
|
| --- |
|
|
| ## Usage |
|
|
| ```python |
| from datasets import load_dataset |
| |
| # 1. Load the "Temporal" config |
| # - Each example in this split has fields like: |
| # { "name": ..., "prompt_templates": [...], "samples": [ { "subject": ..., "object": ..., "time": ... }, ... ], ... } |
| Temporal = load_dataset("dmis-lab/TemporalHead", "Temporal")["default"] |
| |
| # 2. Load the "Invariant" config |
| # - Each example here has fields like: |
| # { "name": ..., "prompt_templates": [...], "properties": { "relation_type": ..., ... }, "samples": [ { "subject": ..., "object": ... }, ... ], ... } |
| Invariant = load_dataset("dmis-lab/TemporalHead", "Invariant")["default"] |
| ``` |
|
|
| --- |
|
|
| ## Citation and Acknowledgements |
|
|
| If you find our work is useful in your research, please consider citing our [paper](https://arxiv.org/abs/2502.14258): |
| ``` |
| @article{park2025does, |
| title={Does Time Have Its Place? Temporal Heads: Where Language Models Recall Time-specific Information}, |
| author={Park, Yein and Yoon, Chanwoong and Park, Jungwoo and Jeong, Minbyul and Kang, Jaewoo}, |
| journal={arXiv preprint arXiv:2502.14258}, |
| year={2025} |
| } |
| ``` |
|
|
| We also gratefully acknowledge the following open-source repositories and kindly ask that you cite their accompanying papers as well. |
|
|
| [1] https://github.com/zjunlp/KnowledgeCircuits |
| [2] https://github.com/hannamw/eap-ig |
| [3] https://github.com/evandez/relations |
| |
| --- |
|
|
| ## Contact |
| For any questions or issues, feel free to reach out to [522yein (at) korea.ac.kr]. |
