From fe08be14794e438b2bc10d45894655b94e2ede4b Mon Sep 17 00:00:00 2001
From: Frank Liu <62361017+franciscoliu@users.noreply.github.com>
Date: Sat, 18 May 2024 07:03:29 +0800
Subject: [PATCH] Update README.md

---
 README.md | 23 ++++++++++++++++-------
 1 file changed, 16 insertions(+), 7 deletions(-)

diff --git a/README.md b/README.md
index 5ffee09..eaf2ebe 100644
--- a/README.md
+++ b/README.md
@@ -1,25 +1,34 @@
-## Selective Knowledge negation Unlearning (SKU)
+# Selective Knowledge negation Unlearning (SKU)
 
-### Instructions
+
+## Environment Setup
+Create an environment from the yml file using:
+
+```bash
+conda env create -f llm_unlearn.yml
+```
+
+
+## Instructions
 We conduct our experiments on OPT2.7b, llama2-7b and llama2-13b model. 
 
-### GPU requirements
+## GPU requirements
 Since task vector would require state_dict deduction and addition between two models, we conduct
 our OPT model experiments on 3 A100 GPUs (80GiB), and llama model experiments on 4 A100 GPUs (80GiB).
 
 
-### Dataset
+## Dataset
 We use the TruthfulQA dataset (avilable on github, one can simply use ```wget``` command to pull the data) as the normal data. 
 And we use PKU-SafeRLHF (avilable on Huggaingface) as  the harmful data.
 
-### Unlearn
+## Unlearn
 We can perform the unlearning by run the following command. Here we show an example of how to unlearn harmfulness learned from the pretrained OPT2.7B. 
 
-#### Before Unlearn
+### Before Unlearn
 After downloading the data, create a directory called ```dataloader``` to store the data. Then, create a directory named ```logs``` in
 harmful_unlearn directory to store the log files.
 
-#### Unlearn process
+### Unlearn process
 ```bash
 python unlearn_harm_new.py --bad_weight 2.5 --random_weight 2.5 --normal_weight 1 
 --lr 2e-4 --max_unlearn_steps 1000 --model_name=facebook/opt-2.7b