Commit
•
ba3f6f3
1
Parent(s):
3139eb9
Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,39 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
language:
|
3 |
+
- en
|
4 |
+
tags:
|
5 |
+
- argumentation
|
6 |
+
license: apache-2.0
|
7 |
+
metrics:
|
8 |
+
- perplexity
|
9 |
+
---
|
10 |
+
|
11 |
+
# Generate the conclusion of an argument
|
12 |
+
|
13 |
+
This model is a version of [`gpt-neo-2.7B`](https://huggingface.co/EleutherAI/gpt-neo-2.7B), where all parameters (both weights and biases) have been finetuned on the task of generating the conclusion of an argument given its premises. It was trained as part of a University of Melbourne [research project](https://github.com/Hunt-Laboratory/language-model-optimization) evaluating how large language models can best be optimized to perform argumentative reasoning tasks.
|
14 |
+
|
15 |
+
Code used for optimization and evaluation can be found in the project [GitHub repository](https://github.com/Hunt-Laboratory/language-model-optimization). A paper reporting on model evaluation is currently under review.
|
16 |
+
|
17 |
+
# Prompt Template
|
18 |
+
|
19 |
+
```
|
20 |
+
Consider the facts:
|
21 |
+
* [premise 1]
|
22 |
+
* [premise 2]
|
23 |
+
...
|
24 |
+
* [premise n]
|
25 |
+
|
26 |
+
We must conclude that: [generated conclusion]
|
27 |
+
```
|
28 |
+
|
29 |
+
# Dataset
|
30 |
+
|
31 |
+
The parameters were finetuned using argument maps scraped from the crowdsourced argument-mapping platform [Kialo](https://kialo.com/).
|
32 |
+
|
33 |
+
# Limitations and Biases
|
34 |
+
|
35 |
+
The model is a finetuned version of [`gpt-neo-2.7B`](https://huggingface.co/EleutherAI/gpt-neo-2.7B), so likely has many of the same limitations and biases. Additionally, note that while the goal of the model is to produce coherent and valid reasoning, many generated model outputs will be illogical or nonsensical and should not be relied upon.
|
36 |
+
|
37 |
+
# Acknowledgements
|
38 |
+
|
39 |
+
This research was funded by the Australian Department of Defence and the Office of National Intelligence under the AI for Decision Making Program, delivered in partnership with the Defence Science Institute in Victoria, Australia.
|