Training Language Models on the Knowledge Graph: Insights on Hallucinations and Their Detectability
Fuente:
arXiv
Saved in:
| Main Authors: | , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2024
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866929460088930304 |
|---|---|
| author | Hron, Jiri Culp, Laura Elsayed, Gamaleldin Liu, Rosanne Adlam, Ben Bileschi, Maxwell Bohnet, Bernd Co-Reyes, JD Fiedel, Noah Freeman, C. Daniel Gur, Izzeddin Kenealy, Kathleen Lee, Jaehoon Liu, Peter J. Mishra, Gaurav Mordatch, Igor Nova, Azade Novak, Roman Parisi, Aaron Pennington, Jeffrey Rizkowsky, Alex Simpson, Isabelle Sedghi, Hanie Sohl-dickstein, Jascha Swersky, Kevin Vikram, Sharad Warkentin, Tris Xiao, Lechao Xu, Kelvin Snoek, Jasper Kornblith, Simon |
| author_facet | Hron, Jiri Culp, Laura Elsayed, Gamaleldin Liu, Rosanne Adlam, Ben Bileschi, Maxwell Bohnet, Bernd Co-Reyes, JD Fiedel, Noah Freeman, C. Daniel Gur, Izzeddin Kenealy, Kathleen Lee, Jaehoon Liu, Peter J. Mishra, Gaurav Mordatch, Igor Nova, Azade Novak, Roman Parisi, Aaron Pennington, Jeffrey Rizkowsky, Alex Simpson, Isabelle Sedghi, Hanie Sohl-dickstein, Jascha Swersky, Kevin Vikram, Sharad Warkentin, Tris Xiao, Lechao Xu, Kelvin Snoek, Jasper Kornblith, Simon |
| contents | While many capabilities of language models (LMs) improve with increased training budget, the influence of scale on hallucinations is not yet fully understood. Hallucinations come in many forms, and there is no universally accepted definition. We thus focus on studying only those hallucinations where a correct answer appears verbatim in the training set. To fully control the training data content, we construct a knowledge graph (KG)-based dataset, and use it to train a set of increasingly large LMs. We find that for a fixed dataset, larger and longer-trained LMs hallucinate less. However, hallucinating on $\leq5$% of the training data requires an order of magnitude larger model, and thus an order of magnitude more compute, than Hoffmann et al. (2022) reported was optimal. Given this costliness, we study how hallucination detectors depend on scale. While we see detector size improves performance on fixed LM's outputs, we find an inverse relationship between the scale of the LM and the detectability of its hallucinations. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2408_07852 |
| institution | arXiv |
| publishDate | 2024 |
| record_format | arxiv |
| spellingShingle | Training Language Models on the Knowledge Graph: Insights on Hallucinations and Their Detectability Hron, Jiri Culp, Laura Elsayed, Gamaleldin Liu, Rosanne Adlam, Ben Bileschi, Maxwell Bohnet, Bernd Co-Reyes, JD Fiedel, Noah Freeman, C. Daniel Gur, Izzeddin Kenealy, Kathleen Lee, Jaehoon Liu, Peter J. Mishra, Gaurav Mordatch, Igor Nova, Azade Novak, Roman Parisi, Aaron Pennington, Jeffrey Rizkowsky, Alex Simpson, Isabelle Sedghi, Hanie Sohl-dickstein, Jascha Swersky, Kevin Vikram, Sharad Warkentin, Tris Xiao, Lechao Xu, Kelvin Snoek, Jasper Kornblith, Simon Computation and Language Artificial Intelligence Machine Learning While many capabilities of language models (LMs) improve with increased training budget, the influence of scale on hallucinations is not yet fully understood. Hallucinations come in many forms, and there is no universally accepted definition. We thus focus on studying only those hallucinations where a correct answer appears verbatim in the training set. To fully control the training data content, we construct a knowledge graph (KG)-based dataset, and use it to train a set of increasingly large LMs. We find that for a fixed dataset, larger and longer-trained LMs hallucinate less. However, hallucinating on $\leq5$% of the training data requires an order of magnitude larger model, and thus an order of magnitude more compute, than Hoffmann et al. (2022) reported was optimal. Given this costliness, we study how hallucination detectors depend on scale. While we see detector size improves performance on fixed LM's outputs, we find an inverse relationship between the scale of the LM and the detectability of its hallucinations. |
| title | Training Language Models on the Knowledge Graph: Insights on Hallucinations and Their Detectability |
| topic | Computation and Language Artificial Intelligence Machine Learning |
| url | https://arxiv.org/abs/2408.07852 |