Spaces:
Running
Running
Update README.md
Browse files
README.md
CHANGED
@@ -7,4 +7,17 @@ sdk: static
|
|
7 |
pinned: false
|
8 |
---
|
9 |
|
10 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
7 |
pinned: false
|
8 |
---
|
9 |
|
10 |
+
### Summary
|
11 |
+
|
12 |
+
News Media Bias is an open-source package to identify and mitigate bias in texts using large language models. News Media Bias is part of Vector Institute's efforts to improve fairness in ML. The package will be open sourced soon (tentatively end of August).
|
13 |
+
|
14 |
+
### Design principles
|
15 |
+
In order of importance, here are the qualities we want this package to have:
|
16 |
+
|
17 |
+
1. Accurate. It's very important to both discern bias accurately and rephrase sentences properly.
|
18 |
+
2. Free of charge. Paid LLMs such as GPT-3.5 or Clause produce good results, but can be expensive for individuals or small organizations, as well as large workloads. This package, on the other hand, will always remain free.
|
19 |
+
3. Performant. While not as important as the first two factors, we do want to make our package reasonably fast and not too resource-hungry.
|
20 |
+
|
21 |
+
### Current implementation
|
22 |
+
|
23 |
+
This package currently uses llama-2-7b for both bias classification and debiasing. The model is quantized to 4 bit so it's able to run on low-compute environments (such as free Google Colab).
|