Mass-Editing Memory in a Transformer
About
Recent work has shown exciting promise in updating large language models with new memories, so as to replace obsolete information or add specialized knowledge. However, this line of work is predominantly limited to updating single associations. We develop MEMIT, a method for directly updating a language model with many memories, demonstrating experimentally that it can scale up to thousands of associations for GPT-J (6B) and GPT-NeoX (20B), exceeding prior work by orders of magnitude. Our code and data are at https://memit.baulab.info.
Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, David Bau• 2022
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Knowledge Editing | CounterFact | Efficacy9.38e+3 | 362 | |
| Multitask Language Understanding | MMLU (test) | Accuracy21.83 | 312 | |
| Knowledge Editing | zsRE | Generality96.4 | 268 | |
| Lifelong Free-text Knowledge Editing | MRLF-Bench | BLEU36.36 | 140 | |
| Commonsense Question Answering | CommonsenseQA | Accuracy20.23 | 92 | |
| Lifelong Model Editing | zsRE | Efficacy10.54 | 89 | |
| Sequential Model Editing | CounterFact | Efficacy98.55 | 81 | |
| Model Editing | UltraEditBench | Efficacy0.82 | 78 | |
| Model Editing | zsRE | Efficacy94.91 | 71 | |
| Lifelong Model Editing | WikiBigEdit | Efficacy1.59 | 63 |
Showing 10 of 185 rows
...