Update README.md
Browse files
README.md
CHANGED
@@ -1,27 +1,52 @@
|
|
1 |
-
---
|
2 |
-
license: mit
|
3 |
-
dataset_info:
|
4 |
-
features:
|
5 |
-
- name: instruction
|
6 |
-
dtype: string
|
7 |
-
- name: output
|
8 |
-
dtype: string
|
9 |
-
- name: task
|
10 |
-
dtype: string
|
11 |
-
splits:
|
12 |
-
- name: train
|
13 |
-
num_bytes: 8972956600
|
14 |
-
num_examples: 503698
|
15 |
-
- name: validation
|
16 |
-
num_bytes: 1259708059
|
17 |
-
num_examples: 71638
|
18 |
-
download_size: 4925396868
|
19 |
-
dataset_size: 10232664659
|
20 |
-
configs:
|
21 |
-
- config_name: default
|
22 |
-
data_files:
|
23 |
-
- split: train
|
24 |
-
path: data/train-*
|
25 |
-
- split: validation
|
26 |
-
path: data/validation-*
|
27 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: mit
|
3 |
+
dataset_info:
|
4 |
+
features:
|
5 |
+
- name: instruction
|
6 |
+
dtype: string
|
7 |
+
- name: output
|
8 |
+
dtype: string
|
9 |
+
- name: task
|
10 |
+
dtype: string
|
11 |
+
splits:
|
12 |
+
- name: train
|
13 |
+
num_bytes: 8972956600
|
14 |
+
num_examples: 503698
|
15 |
+
- name: validation
|
16 |
+
num_bytes: 1259708059
|
17 |
+
num_examples: 71638
|
18 |
+
download_size: 4925396868
|
19 |
+
dataset_size: 10232664659
|
20 |
+
configs:
|
21 |
+
- config_name: default
|
22 |
+
data_files:
|
23 |
+
- split: train
|
24 |
+
path: data/train-*
|
25 |
+
- split: validation
|
26 |
+
path: data/validation-*
|
27 |
+
---
|
28 |
+
# Lawma fine-tuning dataset
|
29 |
+
|
30 |
+
This fine-tuning dataset contains 260 legal classification tasks derived from the [Supreme Court](http://scdb.wustl.edu/data.php) and [Songer Court of Appeals](www.songerproject.org/us-courts-of-appeals-databases.html) databases, totalling over 500k training examples and 2B tokens. This dataset was used to train [Lawma 8B](https://huggingface.co/ricdomolm/lawma-8b) and [Lawma 70B](https://huggingface.co/ricdomolm/lawma-70b). The Lawma models outperform GPT-4 on 95\% of these legal tasks, on average by over 17 accuracy points. See our [arXiv preprint](https://arxiv.org/abs/2407.16615) and [GitHub repository](https://github.com/socialfoundations/lawma) for more details.
|
31 |
+
|
32 |
+
Our reasons to study these legal classification tasks are both technical and substantive. From a technical machine learning perspective, these tasks provide highly non-trivial classification problems where
|
33 |
+
even the best models leave much room for improvement. From a substantive legal perspective, efficient
|
34 |
+
solutions to such classification problems have rich and important applications in legal research.
|
35 |
+
|
36 |
+
This dataset was created for the project
|
37 |
+
|
38 |
+
*Lawma: The Power of Specizalization for Legal Tasks. Ricardo Dominguez-Olmedo and Vedant Nanda and Rediet Abebe and Stefan Bechtold and Christoph Engel and Jens Frankenreiter and Krishna Gummadi and Moritz Hardt and Michael Livermore. 2024*
|
39 |
+
|
40 |
+
Please cite as:
|
41 |
+
|
42 |
+
```
|
43 |
+
@misc{dominguezolmedo2024lawmapowerspecializationlegal,
|
44 |
+
title={Lawma: The Power of Specialization for Legal Tasks},
|
45 |
+
author={Ricardo Dominguez-Olmedo and Vedant Nanda and Rediet Abebe and Stefan Bechtold and Christoph Engel and Jens Frankenreiter and Krishna Gummadi and Moritz Hardt and Michael Livermore},
|
46 |
+
year={2024},
|
47 |
+
eprint={2407.16615},
|
48 |
+
archivePrefix={arXiv},
|
49 |
+
primaryClass={cs.CL},
|
50 |
+
url={https://arxiv.org/abs/2407.16615},
|
51 |
+
}
|
52 |
+
```
|