Update README
Browse files
README.md
CHANGED
@@ -1,3 +1,39 @@
|
|
1 |
-
|
2 |
-
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# JustLogic
|
2 |
+
|
3 |
+
[[Paper]](https://arxiv.org/abs/2501.14851) [[Github]](https://github.com/michaelchen-lab/JustLogic)
|
4 |
+
|
5 |
+
JustLogic is a deductive reasoning datataset that is
|
6 |
+
|
7 |
+
1. highly complex, capable of generating a diverse range of linguistic patterns, vocabulary, and argument structures;
|
8 |
+
2. prior knowledge independent, eliminating the advantage of models possessing prior knowledge and ensuring that only deductive reasoning is used to answer questions; and
|
9 |
+
3. capable of in-depth error analysis on the heterogeneous effects of reasoning depth and argument form on model accuracy.
|
10 |
+
|
11 |
+
## Dataset Format
|
12 |
+
|
13 |
+
- `premises`: List of premises in the question, in the form of a Python list.
|
14 |
+
- `paragraph`: A paragraph consisting of the above `premises`. This is given as input to models.
|
15 |
+
- `conclusion`: The expected conclusion of the given premises.
|
16 |
+
- `question`: The statement in which models must determine its truth-value.
|
17 |
+
- `label`: True | False | Uncertain
|
18 |
+
- `arg`: The argument structure
|
19 |
+
- `statements`: Matching symbols in `arg` to their corresponding natural language statements.
|
20 |
+
- `depth`: The argument depth of the given question
|
21 |
+
|
22 |
+
## Dataset Construction
|
23 |
+
|
24 |
+
JustLogic is a synthetically generated dataset. The script to construct your own dataset can be found in the [Github repo](https://github.com/michaelchen-lab/JustLogic).
|
25 |
+
|
26 |
+
## Citation
|
27 |
+
|
28 |
+
```
|
29 |
+
@article{chen2025justlogic,
|
30 |
+
title={JustLogic: A Comprehensive Benchmark for Evaluating Deductive Reasoning in Large Language Models},
|
31 |
+
author={Chen, Michael K and Zhang, Xikun and Tao, Dacheng},
|
32 |
+
journal={arXiv preprint arXiv:2501.14851},
|
33 |
+
year={2025}
|
34 |
+
}
|
35 |
+
```
|
36 |
+
|
37 |
+
---
|
38 |
+
license: mit
|
39 |
+
---
|