Datasets:
Update README.md
Browse files
README.md
CHANGED
@@ -8,4 +8,52 @@ tags:
|
|
8 |
pretty_name: CausalGym
|
9 |
size_categories:
|
10 |
- 10K<n<100K
|
11 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
8 |
pretty_name: CausalGym
|
9 |
size_categories:
|
10 |
- 10K<n<100K
|
11 |
+
---
|
12 |
+
|
13 |
+
**CausalGym** is a benchmark for comparing the performance of causal interpretability methods
|
14 |
+
on a variety of simple linguistic tasks taken from the SyntaxGym evaluation set
|
15 |
+
([Gauthier et al., 2020](https://aclanthology.org
|
16 |
+
/2020.acl-demos.10/), [Hu et al., 2020](https://aclanthology.org/2020.acl-main.158/))
|
17 |
+
and converted into a format suitable for interventional interpretability.
|
18 |
+
|
19 |
+
The dataset includes train/dev/test splits (exactly as used in the experiments in the paper).
|
20 |
+
The `base`/`src` columns are the prompts on which intervention is done. Each of these is a list of strings,
|
21 |
+
with each string being a span in the template which is aligned by index and may have an unequal number
|
22 |
+
of tokens. The `base_label` and `src_label` columns are the ground truth next-token predictions that we
|
23 |
+
train/evaluate on, and the `base_type` and `src_type` columns indicate the class (always binary) of the prompts.
|
24 |
+
Finally, the `task` column indicates which task this row is from. You should train separately on each task since
|
25 |
+
each one studies a different linguistic feature.
|
26 |
+
|
27 |
+
## Citation
|
28 |
+
|
29 |
+
Paper TBA. Please cite the SyntaxGym papers too:
|
30 |
+
|
31 |
+
```bibtex
|
32 |
+
@inproceedings{gauthier-etal-2020-syntaxgym,
|
33 |
+
title = "{S}yntax{G}ym: An Online Platform for Targeted Evaluation of Language Models",
|
34 |
+
author = "Gauthier, Jon and Hu, Jennifer and Wilcox, Ethan and Qian, Peng and Levy, Roger",
|
35 |
+
editor = "Celikyilmaz, Asli and Wen, Tsung-Hsien",
|
36 |
+
booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations",
|
37 |
+
month = jul,
|
38 |
+
year = "2020",
|
39 |
+
address = "Online",
|
40 |
+
publisher = "Association for Computational Linguistics",
|
41 |
+
url = "https://aclanthology.org/2020.acl-demos.10",
|
42 |
+
doi = "10.18653/v1/2020.acl-demos.10",
|
43 |
+
pages = "70--76",
|
44 |
+
}
|
45 |
+
|
46 |
+
@inproceedings{hu-etal-2020-systematic,
|
47 |
+
title = "A Systematic Assessment of Syntactic Generalization in Neural Language Models",
|
48 |
+
author = "Hu, Jennifer and Gauthier, Jon and Qian, Peng and Wilcox, Ethan and Levy, Roger",
|
49 |
+
editor = "Jurafsky, Dan and Chai, Joyce and Schluter, Natalie and Tetreault, Joel",
|
50 |
+
booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
|
51 |
+
month = jul,
|
52 |
+
year = "2020",
|
53 |
+
address = "Online",
|
54 |
+
publisher = "Association for Computational Linguistics",
|
55 |
+
url = "https://aclanthology.org/2020.acl-main.158",
|
56 |
+
doi = "10.18653/v1/2020.acl-main.158",
|
57 |
+
pages = "1725--1744",
|
58 |
+
}
|
59 |
+
```
|