Skip to content

Commit 89df72d

Browse files
committed
Add Needle
1 parent edd1d5a commit 89df72d

8 files changed

+1581
-3
lines changed

README.md

+5-1
Original file line numberDiff line numberDiff line change
@@ -12,4 +12,8 @@ Implementation of ["Fast Adversarial Attacks on Language Models In One GPU Minut
1212

1313
**[Llama PGD](scripts/pgd.py)**
1414

15-
Implementation of ["Attacking Large Language Models with Projected Gradient Descent"](https://arxiv.org/abs/2402.09154) for Llama model variants with LitGPT. At teh time of release the authors have not posted any reference code, so be careful.
15+
Implementation of ["Attacking Large Language Models with Projected Gradient Descent"](https://arxiv.org/abs/2402.09154) for Llama model variants with LitGPT. At teh time of release the authors have not posted any reference code, so be careful.
16+
17+
**Needle [Triage](notebooks/Needle%20-%20Triage.ipynb)/[Fix](notebooks/Needle%20-%20Fix.ipynb)**
18+
19+
Research in partnership with [OpenSSF](https://openssf.org) for the [AIxCC Event](https://aicyberchallenge.com/).

notebooks/Mistral - Adversarial Suffix.ipynb

+1-1
Original file line numberDiff line numberDiff line change
@@ -109,7 +109,7 @@
109109
"\n",
110110
"prompts: list[Prompt] = []\n",
111111
"\n",
112-
"with open(\"prompts.json\", \"r\") as f:\n",
112+
"with open(\"data/prompts.json\", \"r\") as f:\n",
113113
" for name, content in json.load(f).items():\n",
114114
" if (\n",
115115
" len(content) < PROMPT_LEN_RANGE[0]\n",

notebooks/Mistral - BEAST Beam Attack.ipynb

+1-1
Original file line numberDiff line numberDiff line change
@@ -111,7 +111,7 @@
111111
"\n",
112112
"prompts: list[Prompt] = []\n",
113113
"\n",
114-
"with open(\"prompts.json\", \"r\") as f:\n",
114+
"with open(\"data/prompts.json\", \"r\") as f:\n",
115115
" for name, content in json.load(f).items():\n",
116116
" if (\n",
117117
" len(content) < PROMPT_LENGTHS[0]\n",

0 commit comments

Comments
 (0)