Skip to content

Commit a1b2f08

Browse files
shauli-ravfogelshauli-ravfogel
committed
update
1 parent 61d48e6 commit a1b2f08

File tree

6 files changed

+1173
-0
lines changed

6 files changed

+1173
-0
lines changed

‎analyze.py

+135
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,135 @@
1+
import pickle
2+
import matplotlib.pyplot as plt
3+
import numpy as np
4+
5+
models = ["counterfactuals2/wiki_Meta-Llama-3-8B-Instruct->mimic_gender_llama3_instruct_prompt:first_k_sents:500_prompt_first_k:5_max_new_tokens:25.pkl",
6+
"counterfactuals2/wiki_Meta-Llama-3-8B-Instruct->honest_steering_llama3_instruct_prompt:first_k_sents:500_prompt_first_k:5_max_new_tokens:25.pkl",
7+
#"counterfactuals2/wiki_gpt2-xl->rome_louvre_gpt2_xl_prompt:first_k_sents:500_prompt_first_k:5_max_new_tokens:25.pkl",
8+
"counterfactuals2/wiki_Meta-Llama-3-8B->chat_llama3_prompt:first_k_sents:500_prompt_first_k:5_max_new_tokens:25.pkl",
9+
"counterfactuals2/wiki_gpt2-xl->mimic_gender_gpt2_instruct_prompt:first_k_sents:500_prompt_first_k:5_max_new_tokens:25.pkl",
10+
"counterfactuals2/wiki_gpt2-xl->GPT2-memit-louvre-rome_prompt:first_k_sents:500_prompt_first_k:5_max_new_tokens:25.pkl",
11+
"counterfactuals2/wiki_gpt2-xl->GPT2-memit-koalas-new_zealand_prompt:first_k_sents:500_prompt_first_k:5_max_new_tokens:25.pkl"]
12+
#]#,
13+
#"gpt2-xl_rome-edits-louvre-rome_prompt:True.pkl",
14+
#"Llama-2-7b-hf_Llama-2-7b-chat-hf_prompt:True.pkl"]
15+
names = ["LLaMA3-Steering-Gender", "LLaMA3-Steering-Honest", "LLaMA3-Instruct", "GPT2-XL-Steering-Gender", "GPT2-XL-MEMIT-Louvre", "GPT2-XL-MEMIT-Koalas"] #["Honest-LLama", "GPT-XL-ROME", "LLama2-Chat", "GPT2-XL-MEMIT"]
16+
17+
name2data = {}
18+
colors = ['#1f77b4', '#ff7f0e', '#2ca02c', '#d62728', '#9467bd', "cyan", "purple", "red"] #['blue', 'orange', 'green', "red", "cyan"] # Define colors for consistency
19+
20+
plt.rcParams["font.family"] = "serif"
21+
plt.rcParams.update({'font.size': 15})
22+
plt.figure(figsize=(8, 6))
23+
for idx, (name, model) in enumerate(zip(names, models)):
24+
print(name)
25+
with open(model, "rb") as f:
26+
data = pickle.load(f)
27+
orig, count = data["original"], data["counter"]
28+
#counter = [d["counter"] for d in counter]
29+
#orig = [o.split(" ") for o in original]
30+
#count = [c.split(" ") for c in counter]
31+
32+
orig = [d["tokens"] for d in orig]
33+
count = [d["tokens"][1:] for d in count]
34+
name2data[name] = (orig, count)
35+
36+
37+
diffs=[]
38+
#print(len(orig), len(count))
39+
for o,c in zip(orig, count):
40+
41+
#print(o,c)
42+
i=0
43+
for oo,cc in zip(o,c):
44+
#print("try", cc,oo)
45+
if cc != oo:
46+
#print(i, len(oo))
47+
diffs.append(i/len(o))
48+
break
49+
i+=1
50+
#print(diffs)
51+
52+
53+
plt.hist(
54+
diffs,
55+
density=False,
56+
bins=15,
57+
alpha=0.5,
58+
label=name,
59+
color=colors[idx]
60+
)
61+
62+
# Calculate and plot median
63+
median_diff = np.median(diffs)
64+
plt.axvline(
65+
median_diff,
66+
color=colors[idx],
67+
linestyle='dashed',
68+
linewidth=2
69+
)
70+
"""
71+
plt.text(
72+
median_diff,
73+
plt.ylim()[1]*0.9 - idx*plt.ylim()[1]*0.08, # Adjust y-position for each label
74+
f'Median {name}: {median_diff:.2f}',
75+
rotation=0,
76+
color=colors[idx],
77+
verticalalignment='top',
78+
horizontalalignment='center',
79+
fontsize=20, # Increase font size of median labels
80+
bbox=dict(facecolor='white', alpha=0.5, edgecolor='none')
81+
)
82+
"""
83+
plt.xlabel("Normalized Length of Longest Common Prefix", fontsize=14)
84+
plt.ylabel("Counts", fontsize=14)
85+
plt.grid()
86+
plt.legend(fontsize=13)
87+
plt.savefig("common_prefix.pdf", dpi=800)
88+
89+
90+
91+
#### cosine sim
92+
####
93+
94+
95+
import torch.nn.functional as F
96+
97+
from torch import Tensor
98+
from transformers import AutoTokenizer, AutoModel
99+
import torch
100+
import sklearn
101+
from sklearn.metrics.pairwise import cosine_similarity
102+
103+
def average_pool(last_hidden_states: Tensor,
104+
attention_mask: Tensor) -> Tensor:
105+
last_hidden = last_hidden_states.masked_fill(~attention_mask[..., None].bool(), 0.0)
106+
return last_hidden.sum(dim=1) / attention_mask.sum(dim=1)[..., None]
107+
108+
109+
# Each input text should start with "query: " or "passage: ".
110+
# For tasks other than retrieval, you can simply use the "query: " prefix.
111+
input_texts = ['query: how much protein should a female eat',
112+
'query: summit define',
113+
"passage: As a general guideline, the CDC's average requirement of protein for women ages 19 to 70 is 46 grams per day. But, as you can see from this chart, you'll need to increase that if you're expecting or training for a marathon. Check out the chart below to see how much protein you should be eating each day.",
114+
"passage: Definition of summit for English Language Learners. : 1 the highest point of a mountain : the top of a mountain. : 2 the highest level. : 3 a meeting or series of meetings between the leaders of two or more governments."]
115+
116+
tokenizer = AutoTokenizer.from_pretrained('intfloat/e5-base-v2')
117+
model = AutoModel.from_pretrained('intfloat/e5-base-v2')
118+
119+
for idx, (name, model_path) in enumerate(zip(names, models)):
120+
print(name)
121+
with open(model_path, "rb") as f:
122+
data = pickle.load(f)
123+
orig, count = data["original"], data["counter"]
124+
original = [d["text"] for d in orig]
125+
counter = [d["text"] for d in count]
126+
with torch.no_grad():
127+
batch_dict_original = tokenizer(original, max_length=512, padding=True, truncation=True, return_tensors='pt')
128+
outputs_original = model(**batch_dict_original)
129+
embeddings_original = average_pool(outputs_original.last_hidden_state, batch_dict_original['attention_mask'])
130+
131+
batch_dict_counter = tokenizer(counter, max_length=512, padding=True, truncation=True, return_tensors='pt')
132+
outputs_counter = model(**batch_dict_counter)
133+
embeddings_counter = average_pool(outputs_counter.last_hidden_state, batch_dict_counter['attention_mask'])
134+
135+
print(name, np.diag(cosine_similarity(embeddings_original,embeddings_counter)).mean())

0 commit comments

Comments
 (0)