monsoon-nlp
commited on
Commit
•
aea5dff
1
Parent(s):
959cde2
test post
Browse files- README.md +45 -0
- config.json +36 -0
- merges.txt +0 -0
- pytorch_model.bin +3 -0
- tokenizer.json +0 -0
- vocab.json +0 -0
README.md
ADDED
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
language: en
|
3 |
+
tags:
|
4 |
+
- exbert
|
5 |
+
|
6 |
+
license: mit
|
7 |
+
---
|
8 |
+
|
9 |
+
# no-phone-gpt2
|
10 |
+
|
11 |
+
This is a test to remove memorized private information, such as phone numbers, from a small GPT-2 model. This should not generate valid phone numbers.
|
12 |
+
|
13 |
+
Inspired by BAIR privacy research:
|
14 |
+
- https://bair.berkeley.edu/blog/2019/08/13/memorization/
|
15 |
+
- https://bair.berkeley.edu/blog/2020/12/20/lmmem/
|
16 |
+
|
17 |
+
[Blog post](https://mapmeld.medium.com/scrambling-memorized-info-in-gpt-2-60753d7652d8)
|
18 |
+
|
19 |
+
## Process
|
20 |
+
|
21 |
+
- All +## and +### tokens were replaced with new, randomly-selected 2- and 3-digit numbers in the vocab.json and tokenizer.json. You can identify these in outputs because the new tokens start with ^^.
|
22 |
+
- Input and output embeddings for +## and +### tokens were moved to the +00 and +000 embeddings.
|
23 |
+
- Removed associations between numbers from merges.txt
|
24 |
+
|
25 |
+
Using a library such as [ecco](https://github.com/jalammar/ecco), probabilities for next number token look equally likely, with +000 preferred.
|
26 |
+
|
27 |
+
Code: https://colab.research.google.com/drive/1X31TIZjmxlXMXAzQrR3Fl1AnLzGBCpWf#scrollTo=0GVFwrAgY68J
|
28 |
+
|
29 |
+
### Future goals
|
30 |
+
|
31 |
+
- Add new +### tokens to rebuild number generation
|
32 |
+
- Fine-tune new tokens on counting numbers and ended phone numbers
|
33 |
+
- Use [gpt2-large](https://huggingface.co/gpt2-large)
|
34 |
+
|
35 |
+
### BibTeX entry and citation info
|
36 |
+
|
37 |
+
Original GPT-2:
|
38 |
+
|
39 |
+
```bibtex
|
40 |
+
@article{radford2019language,
|
41 |
+
title={Language Models are Unsupervised Multitask Learners},
|
42 |
+
author={Radford, Alec and Wu, Jeff and Child, Rewon and Luan, David and Amodei, Dario and Sutskever, Ilya},
|
43 |
+
year={2019}
|
44 |
+
}
|
45 |
+
```
|
config.json
ADDED
@@ -0,0 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"_name_or_path": "./gpt2",
|
3 |
+
"activation_function": "gelu_new",
|
4 |
+
"architectures": [
|
5 |
+
"GPT2LMHeadModel"
|
6 |
+
],
|
7 |
+
"attn_pdrop": 0.1,
|
8 |
+
"bos_token_id": 50256,
|
9 |
+
"embd_pdrop": 0.1,
|
10 |
+
"eos_token_id": 50256,
|
11 |
+
"gradient_checkpointing": false,
|
12 |
+
"initializer_range": 0.02,
|
13 |
+
"layer_norm_epsilon": 1e-05,
|
14 |
+
"model_type": "gpt2",
|
15 |
+
"n_ctx": 1024,
|
16 |
+
"n_embd": 768,
|
17 |
+
"n_head": 12,
|
18 |
+
"n_inner": null,
|
19 |
+
"n_layer": 12,
|
20 |
+
"n_positions": 1024,
|
21 |
+
"resid_pdrop": 0.1,
|
22 |
+
"summary_activation": null,
|
23 |
+
"summary_first_dropout": 0.1,
|
24 |
+
"summary_proj_to_labels": true,
|
25 |
+
"summary_type": "cls_index",
|
26 |
+
"summary_use_proj": true,
|
27 |
+
"task_specific_params": {
|
28 |
+
"text-generation": {
|
29 |
+
"do_sample": true,
|
30 |
+
"max_length": 50
|
31 |
+
}
|
32 |
+
},
|
33 |
+
"transformers_version": "4.3.2",
|
34 |
+
"use_cache": true,
|
35 |
+
"vocab_size": 50257
|
36 |
+
}
|
merges.txt
ADDED
The diff for this file is too large to render.
See raw diff
|
pytorch_model.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:ca721d6d982f9bab9330542cca4d1caaa94bbfa04087ab956ecbd03c3b1bf8b2
|
3 |
+
size 510404323
|
tokenizer.json
ADDED
The diff for this file is too large to render.
See raw diff
|
vocab.json
ADDED
The diff for this file is too large to render.
See raw diff
|