monsoon-nlp
commited on
Commit
·
876cbfe
1
Parent(s):
4b0d6b5
model card
Browse files
README.md
ADDED
@@ -0,0 +1,48 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# GPT-NYC
|
2 |
+
|
3 |
+
## About
|
4 |
+
|
5 |
+
GPT2-Medium fine-tuned on questions and responses from https://reddit.com/r/asknyc
|
6 |
+
|
7 |
+
I filtered comments to ones with scores >= 3, and responding directly
|
8 |
+
to the original post ( = ignoring responses to other commenters).
|
9 |
+
|
10 |
+
I added tokens to match NYC neighborhoods, subway stations, foods, and other
|
11 |
+
common terms in the original batches of questions and comments.
|
12 |
+
You would be surprised what is missing from GPT tokens!
|
13 |
+
|
14 |
+
Try prompting with ```question? %% ``` or ```question? - more info %%```
|
15 |
+
|
16 |
+
## Status
|
17 |
+
|
18 |
+
I would like to continue by:
|
19 |
+
- fine-tuning GPT2-Large with a larger dataset of questions
|
20 |
+
- examining bias and toxicity
|
21 |
+
- examining memorization vs. original responses
|
22 |
+
- releasing a reusable benchmark
|
23 |
+
|
24 |
+
## Blog
|
25 |
+
|
26 |
+
https://mapmeld.medium.com/gpt-nyc-part-1-9cb698b2e3d
|
27 |
+
|
28 |
+
## Notebooks
|
29 |
+
|
30 |
+
### Data processing / new tokens
|
31 |
+
|
32 |
+
https://colab.research.google.com/drive/13BOw0uekoAYB4jjQtaXTn6J_VHatiRLu
|
33 |
+
|
34 |
+
### Fine-tuning GPT2 (small)
|
35 |
+
|
36 |
+
https://colab.research.google.com/drive/1FnXcAh4H-k8dAzixkV5ieygV96ePh3lR
|
37 |
+
|
38 |
+
### Fine-tuning GPT2-Medium
|
39 |
+
|
40 |
+
Same code as small, but on Google Cloud to use an A100 GPU
|
41 |
+
|
42 |
+
### Predictive text and probabilities
|
43 |
+
|
44 |
+
Scroll to end of
|
45 |
+
|
46 |
+
https://colab.research.google.com/drive/1FnXcAh4H-k8dAzixkV5ieygV96ePh3lR
|
47 |
+
|
48 |
+
to see how to install git-lfs and trick ecco into loading this.
|