first commit
Browse files- requirements.txt +9 -0
- src/__pycache__/env_options.cpython-311.pyc +0 -0
- src/__pycache__/lmsys_dataset_handler.cpython-311.pyc +0 -0
- src/__pycache__/lmsys_dataset_wrapper.cpython-311.pyc +0 -0
- src/__pycache__/text_classification_functions.cpython-311.pyc +0 -0
- src/env_options.py +57 -0
- src/lmsys_dataset_handler.py +274 -0
- src/lmsys_dataset_wrapper.py +328 -0
- src/text_classification_functions.py +401 -0
requirements.txt
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
streamlit
|
2 |
+
streamlit-aggrid
|
3 |
+
duckdb
|
4 |
+
pandas
|
5 |
+
requests
|
6 |
+
json
|
7 |
+
torch
|
8 |
+
transformers
|
9 |
+
textwrap
|
src/__pycache__/env_options.cpython-311.pyc
ADDED
Binary file (4.08 kB). View file
|
|
src/__pycache__/lmsys_dataset_handler.cpython-311.pyc
ADDED
Binary file (18.4 kB). View file
|
|
src/__pycache__/lmsys_dataset_wrapper.cpython-311.pyc
ADDED
Binary file (20.4 kB). View file
|
|
src/__pycache__/text_classification_functions.cpython-311.pyc
ADDED
Binary file (26.4 kB). View file
|
|
src/env_options.py
ADDED
@@ -0,0 +1,57 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
|
2 |
+
import sys
|
3 |
+
import os
|
4 |
+
import torch
|
5 |
+
import transformers
|
6 |
+
|
7 |
+
def check_env(colab:bool=False, use_dotenv:bool=True, dotenv_path:str=None, colab_secrets:dict=None) -> tuple:
|
8 |
+
# Checking versions and GPU availability:
|
9 |
+
print(f"Python version: {sys.version}")
|
10 |
+
print(f"PyTorch version: {torch.__version__}")
|
11 |
+
print(f"Transformers version: {transformers.__version__}")
|
12 |
+
if torch.cuda.is_available():
|
13 |
+
print(f"CUDA device: {torch.cuda.get_device_name(0)}")
|
14 |
+
print(f"CUDA Version: {torch.version.cuda}")
|
15 |
+
print(f"FlashAttention available: {torch.backends.cuda.flash_sdp_enabled()}")
|
16 |
+
else:
|
17 |
+
print("No CUDA device available")
|
18 |
+
|
19 |
+
if use_dotenv:
|
20 |
+
print("Retrieved token(s) from .env file")
|
21 |
+
from dotenv import load_dotenv
|
22 |
+
load_dotenv(dotenv_path) # path to your dotenv file
|
23 |
+
hf_token = os.getenv("HF_TOKEN")
|
24 |
+
hf_token_write = os.getenv("HF_TOKEN_WRITE") # Only used for updating the Reddgr dataset (privileges needed)
|
25 |
+
openai_api_key = openai_api_key = os.getenv("OPENAI_API_KEY")
|
26 |
+
elif colab:
|
27 |
+
hf_token = colab_secrets.get('HF_TOKEN')
|
28 |
+
hf_token_write = colab_secrets.get('HF_TOKEN_WRITE')
|
29 |
+
openai_api_key = colab_secrets.get("OPENAI_API_KEY")
|
30 |
+
else:
|
31 |
+
print("Retrieved HuggingFace token(s) from environment variables")
|
32 |
+
hf_token = os.environ.get("HF_TOKEN")
|
33 |
+
hf_token_write = os.environ.get("HF_TOKEN_WRITE")
|
34 |
+
openai_api_key = openai_api_key = os.getenv("OPENAI_API_KEY")
|
35 |
+
|
36 |
+
def mask_token(token, unmasked_chars=4):
|
37 |
+
return token[:unmasked_chars] + '*' * (len(token) - unmasked_chars*2) + token[-unmasked_chars:]
|
38 |
+
|
39 |
+
if hf_token is None:
|
40 |
+
print("HF_TOKEN not found in the provided .env file" if use_dotenv else "HF_TOKEN not found in the environment variables")
|
41 |
+
if hf_token_write is None:
|
42 |
+
print("HF_TOKEN_WRITE not found in the provided .env file" if use_dotenv else "HF_TOKEN_WRITE not found in the environment variables")
|
43 |
+
if openai_api_key is None:
|
44 |
+
print("OPENAI_API_KEY not found in the provided .env file" if use_dotenv else "OPENAI_API_KEY not found in the environment variables")
|
45 |
+
|
46 |
+
masked_hf_token = mask_token(hf_token) if hf_token else None
|
47 |
+
masked_hf_token_write = mask_token(hf_token_write) if hf_token_write else None
|
48 |
+
masked_openai_api_key = mask_token(openai_api_key) if openai_api_key else None
|
49 |
+
|
50 |
+
if masked_hf_token:
|
51 |
+
print(f"Using HuggingFace token: {masked_hf_token}")
|
52 |
+
if masked_hf_token_write:
|
53 |
+
print(f"Using HuggingFace write token: {masked_hf_token_write}")
|
54 |
+
if masked_openai_api_key:
|
55 |
+
print(f"Using OpenAI token: {masked_openai_api_key}")
|
56 |
+
|
57 |
+
return hf_token, hf_token_write, openai_api_key
|
src/lmsys_dataset_handler.py
ADDED
@@ -0,0 +1,274 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import pandas as pd
|
3 |
+
import textwrap
|
4 |
+
import random
|
5 |
+
from datasets import load_dataset
|
6 |
+
from IPython.display import display
|
7 |
+
|
8 |
+
class LMSYSChat1MHandler:
|
9 |
+
def __init__(self, hf_token, streaming=False, verbose=True):
|
10 |
+
self.hf_token = hf_token
|
11 |
+
self.streaming = streaming
|
12 |
+
self.lmsys_dataset = load_dataset(
|
13 |
+
'lmsys/lmsys-chat-1m',
|
14 |
+
revision="main",
|
15 |
+
token=self.hf_token,
|
16 |
+
streaming=self.streaming
|
17 |
+
)
|
18 |
+
self.verbose = verbose
|
19 |
+
if verbose:
|
20 |
+
print(self.lmsys_dataset)
|
21 |
+
self.df_sample = None
|
22 |
+
self.df_prompts = None
|
23 |
+
self.unwrapped_turns_df = None
|
24 |
+
|
25 |
+
if not self.streaming and verbose:
|
26 |
+
print('Data is cached at:\n')
|
27 |
+
for file_info in self.lmsys_dataset['train'].cache_files:
|
28 |
+
filename = file_info['filename']
|
29 |
+
file_size = os.path.getsize(filename)
|
30 |
+
i = int((len(filename) - 41) / 2)
|
31 |
+
print(f"Filename: {filename[:i]}*{filename[-41:]}\nSize: {file_size} bytes")
|
32 |
+
|
33 |
+
def extract_df_sample(self, n_samples=None, conversation_ids=None):
|
34 |
+
"""
|
35 |
+
Extracts a sample of conversations or specific conversations based on their conversation IDs.
|
36 |
+
|
37 |
+
Parameters:
|
38 |
+
- n_samples (int): Number of random samples to extract. Ignored if `conversation_ids` is provided.
|
39 |
+
- conversation_ids (list): List of conversation IDs to extract. If provided, this takes precedence over `n_samples`.
|
40 |
+
|
41 |
+
Returns:
|
42 |
+
- pd.DataFrame: A DataFrame containing the extracted conversations.
|
43 |
+
"""
|
44 |
+
if conversation_ids:
|
45 |
+
# Filter conversations based on the provided conversation IDs
|
46 |
+
df_sample = self.lmsys_dataset['train'].to_pandas()
|
47 |
+
df_sample = df_sample[df_sample['conversation_id'].isin(conversation_ids)]
|
48 |
+
print(f"Retrieved {len(df_sample)} conversations based on specified IDs")
|
49 |
+
else:
|
50 |
+
# Randomly sample conversations if no IDs are provided
|
51 |
+
if not self.streaming:
|
52 |
+
df_sample = self.lmsys_dataset['train'].to_pandas().sample(n_samples)
|
53 |
+
print(f"Retrieved {len(df_sample)} random conversations from lmsys/lmsys-chat-1m")
|
54 |
+
else:
|
55 |
+
# Take a sample from the streamed dataset
|
56 |
+
streamed_samples = []
|
57 |
+
for i, row in enumerate(self.lmsys_dataset['train']):
|
58 |
+
streamed_samples.append(row)
|
59 |
+
if i + 1 == n_samples: # Collect only the desired number of samples
|
60 |
+
break
|
61 |
+
# Shuffle and convert the collected samples to a Pandas DataFrame
|
62 |
+
random.shuffle(streamed_samples)
|
63 |
+
df_sample = pd.DataFrame(streamed_samples)
|
64 |
+
|
65 |
+
self.df_sample = df_sample
|
66 |
+
if self.verbose and len(df_sample) > 4:
|
67 |
+
display(df_sample.head(2))
|
68 |
+
print('...')
|
69 |
+
display(df_sample.tail(2))
|
70 |
+
return df_sample
|
71 |
+
|
72 |
+
def parquet_sampling(self, n_samples):
|
73 |
+
base_url = "https://huggingface.co/datasets/lmsys/lmsys-chat-1m/resolve/main/data/"
|
74 |
+
data_files = [
|
75 |
+
"train-00000-of-00006-4feeb3f83346a0e9.parquet",
|
76 |
+
"train-00001-of-00006-4030672591c2f478.parquet",
|
77 |
+
"train-00002-of-00006-1779b7cec9462180.parquet",
|
78 |
+
"train-00003-of-00006-2fa862bfed56af1f.parquet",
|
79 |
+
"train-00004-of-00006-18f4bdd50c103e71.parquet",
|
80 |
+
"train-00005-of-00006-fe1acc5d10a9f0e2.parquet"
|
81 |
+
]
|
82 |
+
sample_file = random.choice(data_files)
|
83 |
+
print(f"Sampling from {sample_file}")
|
84 |
+
data_files = {"train": base_url + sample_file}
|
85 |
+
parquet_sample = load_dataset("parquet", data_files=data_files, split="train")
|
86 |
+
df_sample = parquet_sample.to_pandas().sample(n_samples)
|
87 |
+
print(f"Retrieved {len(df_sample)} random conversations from lmsys/lmsys-chat-1m/{sample_file}")
|
88 |
+
self.df_sample = df_sample
|
89 |
+
if self.verbose and len(df_sample) > 4:
|
90 |
+
display(df_sample.head(2))
|
91 |
+
print('...')
|
92 |
+
display(df_sample.tail(2))
|
93 |
+
return df_sample
|
94 |
+
|
95 |
+
def add_turns_to_conversations(self):
|
96 |
+
"""
|
97 |
+
Adds 'turn' keys to each conversation in the 'conversation' column of the dataframe.
|
98 |
+
"""
|
99 |
+
self.df_sample['conversation'] = self.df_sample['conversation'].apply(
|
100 |
+
lambda conv: Conversation(conv).add_turns()
|
101 |
+
)
|
102 |
+
df_with_turns = self.df_sample
|
103 |
+
return df_with_turns
|
104 |
+
|
105 |
+
def unwrap_turns(self):
|
106 |
+
"""
|
107 |
+
Creates a dataframe where each row corresponds to a pair of user-assistant messages in a conversation and turn.
|
108 |
+
The 'prompt' column contains the user's message, and the 'response' column contains the assistant's message.
|
109 |
+
Each row includes a 'turn_id' column, which numbers the turns uniquely per conversation.
|
110 |
+
"""
|
111 |
+
paired_data = []
|
112 |
+
for _, row in self.df_sample.iterrows():
|
113 |
+
conversation_id = row['conversation_id']
|
114 |
+
row_data = row.to_dict()
|
115 |
+
row_data.pop('conversation') # Remove the 'conversation' field as it's being unwrapped
|
116 |
+
|
117 |
+
current_prompt = None
|
118 |
+
turn_id = None
|
119 |
+
|
120 |
+
for message in row['conversation']:
|
121 |
+
if message['role'] == 'user':
|
122 |
+
current_prompt = message['content']
|
123 |
+
turn_id = f"{conversation_id}{message['turn']:03}" # Create turn_id
|
124 |
+
elif message['role'] == 'assistant' and current_prompt is not None:
|
125 |
+
# Create a new row with the user-assistant pair
|
126 |
+
paired_row = {
|
127 |
+
**row_data,
|
128 |
+
'turn_n': message['turn'],
|
129 |
+
'prompt': current_prompt,
|
130 |
+
'response': message['content'],
|
131 |
+
}
|
132 |
+
paired_data.append(paired_row)
|
133 |
+
current_prompt = None # Reset after pairing
|
134 |
+
|
135 |
+
unwrapped_turns_df = pd.DataFrame(paired_data)
|
136 |
+
unwrapped_turns_df.rename(columns={"turn": "conversation_turns"}, inplace=True) # The naming in the original dataset is ambiguous
|
137 |
+
self.unwrapped_turns_df = unwrapped_turns_df
|
138 |
+
return unwrapped_turns_df
|
139 |
+
|
140 |
+
def extract_prompts(self, filter_language=None, min_char_length=20, max_char_length=500, exclusions=None):
|
141 |
+
"""
|
142 |
+
Extracts user prompts from the sample dataframe, optionally filtering by language and limiting the character length.
|
143 |
+
|
144 |
+
Parameters:
|
145 |
+
- filter_language (list of str or None): A list of specific languages to filter prompts by. If None, no language
|
146 |
+
filter is applied. Examples of valid values include ['English'], ['English', 'Portuguese'], or
|
147 |
+
['Spanish', 'French', 'German'].
|
148 |
+
- min_char_length (int): The minimum character length for user prompts to include. Defaults to 20.
|
149 |
+
- max_char_length (int): The maximum character length for user prompts to include. Defaults to 500.
|
150 |
+
- exclusions (str or None): Path to a text file containing phrases. Prompts containing any of these phrases
|
151 |
+
will be excluded from the results. If None, no exclusions are applied.
|
152 |
+
|
153 |
+
Returns:
|
154 |
+
- pd.DataFrame: A DataFrame containing extracted prompts with columns 'prompt' and 'language'.
|
155 |
+
"""
|
156 |
+
df_sample = self.df_sample
|
157 |
+
if filter_language:
|
158 |
+
extracted_data = df_sample[df_sample['language'].isin(filter_language)].apply(
|
159 |
+
lambda row: [
|
160 |
+
{'content': entry['content'], 'language': row['language']}
|
161 |
+
for entry in row['conversation']
|
162 |
+
if entry['role'] == 'user' and min_char_length <= len(entry['content']) <= max_char_length
|
163 |
+
], axis=1
|
164 |
+
).explode().dropna()
|
165 |
+
else:
|
166 |
+
extracted_data = df_sample.apply(
|
167 |
+
lambda row: [
|
168 |
+
{'content': entry['content'], 'language': row['language']}
|
169 |
+
for entry in row['conversation']
|
170 |
+
if entry['role'] == 'user' and min_char_length <= len(entry['content']) <= max_char_length
|
171 |
+
], axis=1
|
172 |
+
).explode().dropna()
|
173 |
+
|
174 |
+
df_prompts = pd.DataFrame(extracted_data.tolist())
|
175 |
+
df_prompts.rename(columns={'content': 'prompt'}, inplace=True)
|
176 |
+
|
177 |
+
orig_length = len(df_prompts)
|
178 |
+
if exclusions:
|
179 |
+
# Excluding prompts with phrases that are repeated often in this dataset
|
180 |
+
with open(exclusions, 'r') as f:
|
181 |
+
exclusions = [line.strip() for line in f.readlines()]
|
182 |
+
df_prompts = df_prompts[~df_prompts['prompt'].apply(lambda x: any(exclusion in x for exclusion in exclusions))]
|
183 |
+
print(f"Excluded {orig_length - len(df_prompts)} prompts.")
|
184 |
+
|
185 |
+
self.df_prompts = df_prompts
|
186 |
+
if self.verbose and len(df_sample) > 4:
|
187 |
+
display(df_prompts.head(2))
|
188 |
+
print('...')
|
189 |
+
display(df_prompts.tail(2))
|
190 |
+
return df_prompts
|
191 |
+
|
192 |
+
def extract_prompt_sample(self):
|
193 |
+
prompt_sample = self.df_prompts.sample(1)['prompt'].values[0]
|
194 |
+
if self.verbose:
|
195 |
+
wrapped_message = textwrap.fill(prompt_sample, width=120)
|
196 |
+
print(wrapped_message)
|
197 |
+
return prompt_sample
|
198 |
+
|
199 |
+
def search_conversations(self, search_term):
|
200 |
+
"""
|
201 |
+
Searches the dataset for a given string and returns a DataFrame with matching records.
|
202 |
+
|
203 |
+
Parameters:
|
204 |
+
- search_term (str): The string to search for in the dataset.
|
205 |
+
|
206 |
+
Returns:
|
207 |
+
- pd.DataFrame: A DataFrame containing conversations where the search term is found.
|
208 |
+
"""
|
209 |
+
if self.streaming:
|
210 |
+
raise ValueError("Search is not supported in streaming mode.")
|
211 |
+
df = self.lmsys_dataset['train'].to_pandas()
|
212 |
+
# Filter rows where the search term appears in the 'conversation' column
|
213 |
+
matching_records = df[df['conversation'].apply(
|
214 |
+
lambda conv: any(search_term.lower() in message['content'].lower() for message in conv)
|
215 |
+
)]
|
216 |
+
if self.verbose:
|
217 |
+
print(f"Found {len(matching_records)} matching conversations for search term: '{search_term}'")
|
218 |
+
return matching_records
|
219 |
+
|
220 |
+
def print_language_counts(self, df):
|
221 |
+
language_counts = df['language'].value_counts()
|
222 |
+
print("Language Record Counts:")
|
223 |
+
print(language_counts.to_frame('Count').reset_index().rename(columns={'index': 'Language'}))
|
224 |
+
|
225 |
+
|
226 |
+
class Conversation:
|
227 |
+
def __init__(self, conversation_data):
|
228 |
+
"""
|
229 |
+
Initializes the Conversation object with the conversation data.
|
230 |
+
|
231 |
+
Parameters:
|
232 |
+
- conversation_data (list): A list of dictionaries representing a conversation.
|
233 |
+
"""
|
234 |
+
self.conversation_data = conversation_data
|
235 |
+
|
236 |
+
def add_turns(self):
|
237 |
+
"""
|
238 |
+
Adds a 'turn' key to each dictionary in the conversation,
|
239 |
+
identifying the turn (pair of user and assistant messages).
|
240 |
+
|
241 |
+
Returns:
|
242 |
+
- list: The updated conversation with 'turn' keys added.
|
243 |
+
"""
|
244 |
+
turn_counter = 0
|
245 |
+
for message in self.conversation_data:
|
246 |
+
if message['role'] == 'user':
|
247 |
+
turn_counter += 1
|
248 |
+
message['turn'] = turn_counter
|
249 |
+
return self.conversation_data
|
250 |
+
|
251 |
+
def pretty_print(self, user_prefix, assistant_prefix, width=80):
|
252 |
+
"""
|
253 |
+
Prints the conversation with specified prefixes and wrapped text.
|
254 |
+
|
255 |
+
Parameters:
|
256 |
+
- user_prefix (str): Prefix to prepend to user messages.
|
257 |
+
- assistant_prefix (str): Prefix to prepend to assistant messages.
|
258 |
+
- width (int): Maximum characters per line for wrapping.
|
259 |
+
"""
|
260 |
+
wrapper = textwrap.TextWrapper(width=width)
|
261 |
+
|
262 |
+
for message in self.conversation_data:
|
263 |
+
if message['role'] == 'user':
|
264 |
+
prefix = user_prefix
|
265 |
+
elif message['role'] == 'assistant':
|
266 |
+
prefix = assistant_prefix
|
267 |
+
else:
|
268 |
+
continue # Ignore roles other than 'user' and 'assistant'
|
269 |
+
|
270 |
+
# Split on existing newlines, wrap each line, and join back with newlines
|
271 |
+
wrapped_content = "\n".join(
|
272 |
+
wrapper.fill(line) for line in message['content'].splitlines()
|
273 |
+
)
|
274 |
+
print(f"{prefix} {wrapped_content}\n")
|
src/lmsys_dataset_wrapper.py
ADDED
@@ -0,0 +1,328 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import pandas as pd
|
3 |
+
import textwrap
|
4 |
+
import random
|
5 |
+
import duckdb
|
6 |
+
import requests
|
7 |
+
import json
|
8 |
+
import tempfile
|
9 |
+
from datasets import load_dataset
|
10 |
+
from IPython.display import display
|
11 |
+
from collections import defaultdict
|
12 |
+
|
13 |
+
class DatasetWrapper:
|
14 |
+
def __init__(self, hf_token, dataset_name="lmsys/lmsys-chat-1m", verbose=True,
|
15 |
+
conversations_index="json/conversations_index.json", cache_size=50, request_timeout=20):
|
16 |
+
self.hf_token = hf_token
|
17 |
+
self.dataset_name = dataset_name
|
18 |
+
self.headers = {"Authorization": f"Bearer {self.hf_token}"}
|
19 |
+
self.timeout = request_timeout
|
20 |
+
self.cache_size = cache_size
|
21 |
+
self.verbose = verbose
|
22 |
+
parquet_list_url = f"https://datasets-server.huggingface.co/parquet?dataset={self.dataset_name}"
|
23 |
+
response = self._safe_get(parquet_list_url)
|
24 |
+
# Extract URLs from the response JSON
|
25 |
+
if response is not None:
|
26 |
+
self.parquet_urls = [file['url'] for file in response.json()['parquet_files']]
|
27 |
+
if self.verbose:
|
28 |
+
print("\nParquet URLs:")
|
29 |
+
for url in self.parquet_urls:
|
30 |
+
print(url)
|
31 |
+
head_response = self._safe_head(url)
|
32 |
+
file_size = int(head_response.headers['Content-Length'])
|
33 |
+
print(f"{url.split('/')[-1]}: {file_size} bytes")
|
34 |
+
|
35 |
+
# Loading the index
|
36 |
+
try:
|
37 |
+
with open(conversations_index, "r", encoding="utf-8") as f:
|
38 |
+
self.conversations_index = json.load(f)
|
39 |
+
except (FileNotFoundError, json.JSONDecodeError):
|
40 |
+
print(f"Conversations index file not found or invalid. Creating a new one at {conversations_index}.")
|
41 |
+
# Ensure directory exists
|
42 |
+
os.makedirs(os.path.dirname(conversations_index), exist_ok=True)
|
43 |
+
self.create_conversations_index(output_index_file=conversations_index)
|
44 |
+
with open(conversations_index, "r", encoding="utf-8") as f:
|
45 |
+
self.conversations_index = json.load(f)
|
46 |
+
|
47 |
+
# Initialize active conversation and DataFrame
|
48 |
+
# Read from "pkl/cached_chats.pkl" if available:
|
49 |
+
try:
|
50 |
+
self.active_df = pd.read_pickle("pkl/cached_chats.pkl")
|
51 |
+
print(f"Loaded {len(self.active_df)} cached chats")
|
52 |
+
self.active_df = self.active_df.sample(self.cache_size).reset_index(drop=True)
|
53 |
+
except (FileNotFoundError, ValueError):
|
54 |
+
self.active_df = pd.DataFrame()
|
55 |
+
print("No cached chats found")
|
56 |
+
if not self.active_df.empty:
|
57 |
+
try:
|
58 |
+
self.active_conversation = Conversation(self.active_df.iloc[0])
|
59 |
+
except Exception as e:
|
60 |
+
print(f"No conversations available: {e}")
|
61 |
+
else:
|
62 |
+
self.active_conversation = None
|
63 |
+
|
64 |
+
def _safe_get(self, url):
|
65 |
+
if self.timeout == 0:
|
66 |
+
print("Timeout is set to 0. Skipping GET request.")
|
67 |
+
return None
|
68 |
+
else:
|
69 |
+
try:
|
70 |
+
response = requests.get(url, headers=self.headers, timeout=self.timeout)
|
71 |
+
if response.status_code != 200:
|
72 |
+
raise ValueError(f"Failed to retrieve {url}. Status code: {response.status_code}")
|
73 |
+
return response
|
74 |
+
except requests.exceptions.Timeout:
|
75 |
+
print(f"Timeout occurred for GET {url}. Skipping.")
|
76 |
+
return None
|
77 |
+
|
78 |
+
def _safe_head(self, url):
|
79 |
+
if self.timeout == 0:
|
80 |
+
print("Timeout is set to 0. Skipping HEAD request.")
|
81 |
+
return None
|
82 |
+
try:
|
83 |
+
response = requests.head(url, allow_redirects=True, headers=self.headers, timeout=self.timeout)
|
84 |
+
return response
|
85 |
+
except requests.exceptions.Timeout:
|
86 |
+
print(f"Timeout occurred for GET {url}. Skipping.")
|
87 |
+
return None
|
88 |
+
|
89 |
+
def extract_sample_conversations(self, n_samples):
|
90 |
+
url = random.choice(self.parquet_urls)
|
91 |
+
print(f"Sampling conversations from {url}")
|
92 |
+
# Download file with auth headers using requests
|
93 |
+
r = self._safe_get(url)
|
94 |
+
if r is None:
|
95 |
+
print(f"Timeout occurred for GET {url}. Skipping sample extraction.")
|
96 |
+
return self.active_df
|
97 |
+
# Write the downloaded content into a temporary file
|
98 |
+
with tempfile.NamedTemporaryFile(suffix=".parquet", delete=False) as tmp:
|
99 |
+
tmp.write(r.content)
|
100 |
+
# tmp.flush()
|
101 |
+
tmp_path = tmp.name
|
102 |
+
try:
|
103 |
+
query_result = duckdb.query(f"SELECT * FROM read_parquet('{tmp_path}') USING SAMPLE {n_samples}").df()
|
104 |
+
self.active_df = query_result
|
105 |
+
try:
|
106 |
+
self.active_conversation = Conversation(query_result.iloc[0])
|
107 |
+
except Exception as e:
|
108 |
+
print(f"No conversations available: {e}")
|
109 |
+
finally:
|
110 |
+
# Clean up the temporary file
|
111 |
+
if os.path.exists(tmp_path):
|
112 |
+
os.unlink(tmp_path)
|
113 |
+
|
114 |
+
return query_result
|
115 |
+
|
116 |
+
def extract_conversations(self, conversation_ids):
|
117 |
+
|
118 |
+
# Create a lookup table for file names -> URLs
|
119 |
+
file_url_map = {url.split("/")[-1]: url for url in self.parquet_urls}
|
120 |
+
|
121 |
+
# Group conversation IDs by file
|
122 |
+
file_to_conversations = defaultdict(list)
|
123 |
+
for convid in conversation_ids:
|
124 |
+
if convid in self.conversations_index:
|
125 |
+
file_to_conversations[self.conversations_index[convid]].append(convid)
|
126 |
+
|
127 |
+
result_df = pd.DataFrame()
|
128 |
+
|
129 |
+
for file_name, conv_ids in file_to_conversations.items():
|
130 |
+
if file_name not in file_url_map:
|
131 |
+
print(f"File {file_name} not found in URL list, skipping.")
|
132 |
+
continue
|
133 |
+
|
134 |
+
file_url = file_url_map[file_name]
|
135 |
+
print(f"Querying file: {file_name} for {len(conv_ids)} conversations")
|
136 |
+
|
137 |
+
try:
|
138 |
+
r = self._safe_get(file_url)
|
139 |
+
if r == None:
|
140 |
+
print(f"Timeout occurred for GET {file_url}. Skipping file {file_name}.")
|
141 |
+
continue
|
142 |
+
|
143 |
+
with tempfile.NamedTemporaryFile(suffix=".parquet", delete=False) as tmp:
|
144 |
+
tmp.write(r.content)
|
145 |
+
tmp_path = tmp.name
|
146 |
+
try:
|
147 |
+
conv_id_list = "', '".join(conv_ids)
|
148 |
+
query_str = f"""
|
149 |
+
SELECT * FROM read_parquet('{tmp_path}')
|
150 |
+
WHERE conversation_id IN ('{conv_id_list}')
|
151 |
+
"""
|
152 |
+
df = duckdb.query(query_str).df()
|
153 |
+
finally:
|
154 |
+
if os.path.exists(tmp_path):
|
155 |
+
os.unlink(tmp_path)
|
156 |
+
|
157 |
+
if not df.empty:
|
158 |
+
print(f"Found {len(df)} conversations in {file_name}")
|
159 |
+
result_df = pd.concat([result_df, df], ignore_index=True)
|
160 |
+
|
161 |
+
except Exception as e:
|
162 |
+
print(f"Error processing {file_name}: {e}")
|
163 |
+
|
164 |
+
self.active_df = result_df
|
165 |
+
try:
|
166 |
+
self.active_conversation = Conversation(self.active_df.iloc[0])
|
167 |
+
except Exception as e:
|
168 |
+
print(f"No conversations available: {e}")
|
169 |
+
|
170 |
+
return result_df
|
171 |
+
|
172 |
+
def literal_text_search(self, filter_str, min_results=1):
|
173 |
+
# If filter_str is empty, sample random conversations
|
174 |
+
if filter_str == "":
|
175 |
+
result_df = self.extract_sample_conversations(50)
|
176 |
+
urls = self.parquet_urls.copy()
|
177 |
+
random.shuffle(urls)
|
178 |
+
|
179 |
+
result_df = pd.DataFrame()
|
180 |
+
|
181 |
+
for url in urls:
|
182 |
+
print(f"Querying file: {url}")
|
183 |
+
r = self._safe_get(url)
|
184 |
+
if r == None:
|
185 |
+
print(f"Timeout occurred for GET {url}. Skipping file {url}.")
|
186 |
+
continue
|
187 |
+
with tempfile.NamedTemporaryFile(suffix=".parquet", delete=False) as tmp:
|
188 |
+
tmp.write(r.content)
|
189 |
+
tmp_path = tmp.name
|
190 |
+
|
191 |
+
try:
|
192 |
+
query_str = f"""
|
193 |
+
SELECT * FROM read_parquet('{tmp_path}')
|
194 |
+
WHERE contains(lower(cast(conversation as VARCHAR)), lower('{filter_str}'))
|
195 |
+
"""
|
196 |
+
df = duckdb.query(query_str).df()
|
197 |
+
finally:
|
198 |
+
if os.path.exists(tmp_path):
|
199 |
+
os.unlink(tmp_path)
|
200 |
+
|
201 |
+
print(f"Found {len(df)} result(s) in {url.split('/')[-1]}")
|
202 |
+
|
203 |
+
if len(df) > 0:
|
204 |
+
result_df = pd.concat([result_df, df], ignore_index=True)
|
205 |
+
|
206 |
+
if len(result_df) >= min_results:
|
207 |
+
break
|
208 |
+
if len(result_df) == 0:
|
209 |
+
print("No results found. Returning empty DataFrame.")
|
210 |
+
placeholder_row = {'conversation_id': "No result found",
|
211 |
+
'model': "-",
|
212 |
+
'conversation': [
|
213 |
+
{'content': '-', 'role': 'user'},
|
214 |
+
{'content': '-', 'role': 'assistant'}
|
215 |
+
],
|
216 |
+
'turn': "-",
|
217 |
+
'language': "-",
|
218 |
+
'openai_moderation': "[{'-': '-', '-': '-'}]",
|
219 |
+
'redacted': "-",}
|
220 |
+
result_df = pd.DataFrame([placeholder_row])
|
221 |
+
print(result_df)
|
222 |
+
self.active_df = result_df
|
223 |
+
try:
|
224 |
+
self.active_conversation = Conversation(self.active_df.iloc[0])
|
225 |
+
except Exception as e:
|
226 |
+
print(f"No conversations available: {e}")
|
227 |
+
return result_df
|
228 |
+
|
229 |
+
def create_conversations_index(self, output_index_file="json/conversations_index.json"):
|
230 |
+
"""
|
231 |
+
Builds an index of conversation IDs from a list of Parquet file URLs.
|
232 |
+
Stores the index as a JSON mapping conversation IDs to their respective file names.
|
233 |
+
"""
|
234 |
+
index = {}
|
235 |
+
|
236 |
+
for url in self.parquet_urls:
|
237 |
+
file_name = url.split('/')[-1] # Extract file name from URL
|
238 |
+
print(f"Indexing file: {file_name}")
|
239 |
+
|
240 |
+
try:
|
241 |
+
# Download the file temporarily
|
242 |
+
r = requests.get(url, headers=self.headers)
|
243 |
+
with tempfile.NamedTemporaryFile(suffix=".parquet", delete=False) as tmp:
|
244 |
+
tmp.write(r.content)
|
245 |
+
# tmp.flush()
|
246 |
+
tmp_path = tmp.name
|
247 |
+
try:
|
248 |
+
query = f"SELECT conversation_id FROM read_parquet('{tmp_path}')"
|
249 |
+
df = duckdb.query(query).to_df()
|
250 |
+
finally:
|
251 |
+
if os.path.exists(tmp_path):
|
252 |
+
os.unlink(tmp_path)
|
253 |
+
|
254 |
+
# Map conversation IDs to file name (not the full URL)
|
255 |
+
for _, row in df.iterrows():
|
256 |
+
index[row["conversation_id"]] = file_name
|
257 |
+
|
258 |
+
except Exception as e:
|
259 |
+
print(f"Error indexing {file_name}: {e}")
|
260 |
+
|
261 |
+
# Save index for fast lookup
|
262 |
+
with open(output_index_file, "w", encoding="utf-8") as f:
|
263 |
+
json.dump(index, f, indent=2)
|
264 |
+
|
265 |
+
return output_index_file
|
266 |
+
|
267 |
+
|
268 |
+
class Conversation:
|
269 |
+
def __init__(self, data):
|
270 |
+
"""
|
271 |
+
Initialize a conversation object either from conversation data directly or from a DataFrame row.
|
272 |
+
|
273 |
+
Parameters:
|
274 |
+
- data: Can be either a list of conversation messages or a pandas Series/dict containing conversation data
|
275 |
+
"""
|
276 |
+
# Handle both direct conversation data and DataFrame row
|
277 |
+
if isinstance(data, (pd.Series, dict)):
|
278 |
+
# Store all metadata separately
|
279 |
+
self.conversation_metadata = {}
|
280 |
+
for key, value in (data.items() if isinstance(data, pd.Series) else data.items()):
|
281 |
+
if key == 'conversation':
|
282 |
+
self.conversation_data = value
|
283 |
+
else:
|
284 |
+
self.conversation_metadata[key] = value
|
285 |
+
else:
|
286 |
+
# Direct initialization with conversation data
|
287 |
+
self.conversation_data = data
|
288 |
+
self.conversation_metadata = {}
|
289 |
+
|
290 |
+
def add_turns(self):
|
291 |
+
"""
|
292 |
+
Adds a 'turn' key to each dictionary in the conversation,
|
293 |
+
identifying the turn (pair of user and assistant messages).
|
294 |
+
|
295 |
+
Returns:
|
296 |
+
- list: The updated conversation with 'turn' keys added.
|
297 |
+
"""
|
298 |
+
turn_counter = 0
|
299 |
+
for message in self.conversation_data:
|
300 |
+
if message['role'] == 'user':
|
301 |
+
turn_counter += 1
|
302 |
+
message['turn'] = turn_counter
|
303 |
+
return self.conversation_data
|
304 |
+
|
305 |
+
def pretty_print(self, user_prefix, assistant_prefix, width=80):
|
306 |
+
"""
|
307 |
+
Prints the conversation with specified prefixes and wrapped text.
|
308 |
+
|
309 |
+
Parameters:
|
310 |
+
- user_prefix (str): Prefix to prepend to user messages.
|
311 |
+
- assistant_prefix (str): Prefix to prepend to assistant messages.
|
312 |
+
- width (int): Maximum characters per line for wrapping.
|
313 |
+
"""
|
314 |
+
wrapper = textwrap.TextWrapper(width=width)
|
315 |
+
|
316 |
+
for message in self.conversation_data:
|
317 |
+
if message['role'] == 'user':
|
318 |
+
prefix = user_prefix
|
319 |
+
elif message['role'] == 'assistant':
|
320 |
+
prefix = assistant_prefix
|
321 |
+
else:
|
322 |
+
continue # Ignore roles other than 'user' and 'assistant'
|
323 |
+
|
324 |
+
# Split on existing newlines, wrap each line, and join back with newlines
|
325 |
+
wrapped_content = "\n".join(
|
326 |
+
wrapper.fill(line) for line in message['content'].splitlines()
|
327 |
+
)
|
328 |
+
print(f"{prefix} {wrapped_content}\n")
|
src/text_classification_functions.py
ADDED
@@ -0,0 +1,401 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import pipeline, AutoModelForSequenceClassification, AutoTokenizer
|
2 |
+
from datasets import Dataset
|
3 |
+
from tqdm import tqdm
|
4 |
+
import torch
|
5 |
+
import numpy as np
|
6 |
+
import os
|
7 |
+
from langdetect import detect
|
8 |
+
from sklearn.metrics import accuracy_score, f1_score, log_loss, confusion_matrix, ConfusionMatrixDisplay
|
9 |
+
import matplotlib.pyplot as plt
|
10 |
+
|
11 |
+
class Classifier:
|
12 |
+
def __init__(self, model_path, label_map, verbose = False):
|
13 |
+
self.model_path = model_path
|
14 |
+
self.classifier = pipeline("text-classification", model=model_path, tokenizer=model_path, device=0 if torch.cuda.is_available() else -1)
|
15 |
+
self.label_map = label_map
|
16 |
+
if verbose:
|
17 |
+
self.print_device_information()
|
18 |
+
|
19 |
+
def print_device_information(self):
|
20 |
+
# Check device information
|
21 |
+
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
22 |
+
device_properties = torch.cuda.get_device_properties(0) if device.type == "cuda" else "CPU Device"
|
23 |
+
|
24 |
+
print(f"Using device: {device}")
|
25 |
+
if device.type == "cuda":
|
26 |
+
print(f"Device Name: {device_properties.name}")
|
27 |
+
# print(f"Compute Capability: {device_properties.major}.{device_properties.minor}")
|
28 |
+
print(f"Total Memory: {device_properties.total_memory / 1e9:.2f} GB")
|
29 |
+
|
30 |
+
def tokenize_and_trim(self, text):
|
31 |
+
max_length = self.classifier.tokenizer.model_max_length
|
32 |
+
inputs = self.classifier.tokenizer(text, truncation=True, max_length=max_length, return_tensors="tf")
|
33 |
+
return self.classifier.tokenizer.decode(inputs['input_ids'][0], skip_special_tokens=True)
|
34 |
+
|
35 |
+
|
36 |
+
def classify_dataframe_column(self, df, target_column, feature_suffix):
|
37 |
+
|
38 |
+
tqdm.pandas()
|
39 |
+
df[f'trimmed_{target_column}'] = df[target_column].progress_apply(self.tokenize_and_trim)
|
40 |
+
|
41 |
+
results = []
|
42 |
+
for text in tqdm(df[f'trimmed_{target_column}'].tolist(), desc="Classifying"):
|
43 |
+
result = self.classifier(text)
|
44 |
+
results.append(result[0])
|
45 |
+
|
46 |
+
df[f'pred_label_{feature_suffix}'] = [self.label_map[int(result['label'].split('_')[-1])] for result in results]
|
47 |
+
df[f'prob_{feature_suffix}'] = [result['score'] for result in results]
|
48 |
+
df.drop(columns=[f'trimmed_{target_column}'], inplace=True)
|
49 |
+
return df
|
50 |
+
|
51 |
+
def test_model_predictions(self, df, target_column):
|
52 |
+
"""
|
53 |
+
Tests model predictions on a given dataframe column and computes evaluation metrics.
|
54 |
+
|
55 |
+
Args:
|
56 |
+
df (pd.DataFrame): Input dataframe containing the data.
|
57 |
+
target_column (str): The name of the column to classify.
|
58 |
+
|
59 |
+
Requirements:
|
60 |
+
- The dataframe must include a 'label' column for comparison with predictions.
|
61 |
+
|
62 |
+
Returns:
|
63 |
+
dict: A dictionary containing accuracy, F1 score, cross-entropy loss,
|
64 |
+
and the confusion matrix.
|
65 |
+
"""
|
66 |
+
# Convert pandas dataframe to Dataset
|
67 |
+
dataset = Dataset.from_pandas(df)
|
68 |
+
|
69 |
+
# Define a processing function for tokenization and classification
|
70 |
+
def process_data(batch):
|
71 |
+
trimmed_text = self.tokenize_and_trim(batch[target_column])
|
72 |
+
result = self.classifier(trimmed_text)
|
73 |
+
score = result[0]['score']
|
74 |
+
label = result[0]['label']
|
75 |
+
return {
|
76 |
+
'trimmed_text': trimmed_text,
|
77 |
+
'predicted_prob_0': score if label == 'LABEL_0' else 1 - score,
|
78 |
+
'predicted_prob_1': 1 - score if label == 'LABEL_0' else score,
|
79 |
+
}
|
80 |
+
|
81 |
+
# Apply processing with map
|
82 |
+
processed_dataset = dataset.map(process_data, batched=False)
|
83 |
+
|
84 |
+
# Convert back to pandas dataframe
|
85 |
+
processed_df = processed_dataset.to_pandas()
|
86 |
+
|
87 |
+
# Extract predicted probabilities and true labels
|
88 |
+
predicted_probs = processed_df[['predicted_prob_0', 'predicted_prob_1']].values
|
89 |
+
true_labels = df['label'].values
|
90 |
+
|
91 |
+
# Calculate metrics
|
92 |
+
accuracy = accuracy_score(true_labels, np.argmax(predicted_probs, axis=1))
|
93 |
+
f1 = f1_score(true_labels, np.argmax(predicted_probs, axis=1), average='weighted')
|
94 |
+
cross_entropy_loss = log_loss(true_labels, predicted_probs)
|
95 |
+
|
96 |
+
# Print metrics
|
97 |
+
print(f"Accuracy: {accuracy:.4f}")
|
98 |
+
print(f"F1 Score: {f1:.4f}")
|
99 |
+
print(f"Cross Entropy Loss: {cross_entropy_loss:.4f}")
|
100 |
+
|
101 |
+
# Confusion matrix
|
102 |
+
cm = confusion_matrix(true_labels, np.argmax(predicted_probs, axis=1))
|
103 |
+
cmap = plt.cm.Blues
|
104 |
+
disp = ConfusionMatrixDisplay(confusion_matrix=cm, display_labels=[0, 1])
|
105 |
+
disp.plot(cmap=cmap)
|
106 |
+
plt.show()
|
107 |
+
|
108 |
+
# Return metrics and probabilities for further inspection
|
109 |
+
return {
|
110 |
+
"accuracy": accuracy,
|
111 |
+
"f1_score": f1,
|
112 |
+
"cross_entropy_loss": cross_entropy_loss,
|
113 |
+
"confusion_matrix": cm,
|
114 |
+
"predicted_probs": predicted_probs # Include reconstructed probabilities
|
115 |
+
}
|
116 |
+
|
117 |
+
|
118 |
+
class LanguageDetector:
|
119 |
+
def __init__(self, dataframe):
|
120 |
+
"""
|
121 |
+
Initializes the LanguageDetector with the provided DataFrame.
|
122 |
+
"""
|
123 |
+
self.dataframe = dataframe
|
124 |
+
|
125 |
+
def detect_language_dataframe_column(self, target_column):
|
126 |
+
"""
|
127 |
+
Detects the language of text in the specified column using langdetect and adds
|
128 |
+
a 'detected_language' column to the DataFrame.
|
129 |
+
"""
|
130 |
+
def detect_language(text):
|
131 |
+
try:
|
132 |
+
return detect(text)
|
133 |
+
except Exception:
|
134 |
+
return None
|
135 |
+
|
136 |
+
tqdm.pandas()
|
137 |
+
self.dataframe['detected_language'] = self.dataframe[target_column].progress_apply(detect_language)
|
138 |
+
|
139 |
+
return self.dataframe
|
140 |
+
|
141 |
+
|
142 |
+
# Classifier with Tensorflow backend
|
143 |
+
class TensorflowClassifier(Classifier):
|
144 |
+
def __init__(self, model_path, label_map, verbose=False):
|
145 |
+
super().__init__(model_path, label_map, verbose=False)
|
146 |
+
self.is_tensorflow = False
|
147 |
+
|
148 |
+
if self._is_tensorflow_model(model_path):
|
149 |
+
self.model = tf.keras.models.load_model(model_path)
|
150 |
+
self.tokenizer = AutoTokenizer.from_pretrained("distilbert-base-uncased") # Adjust as per training tokenizer
|
151 |
+
self.is_tensorflow = True
|
152 |
+
if verbose:
|
153 |
+
print("Loaded TensorFlow model.")
|
154 |
+
else:
|
155 |
+
if verbose:
|
156 |
+
print("Fallback to HuggingFace pipeline.")
|
157 |
+
|
158 |
+
def _is_tensorflow_model(self, model_path):
|
159 |
+
return os.path.isdir(model_path) and os.path.exists(os.path.join(model_path, "saved_model.pb"))
|
160 |
+
|
161 |
+
def classify(self, text):
|
162 |
+
if self.is_tensorflow:
|
163 |
+
inputs = self.tokenizer(text, truncation=True, max_length=self.tokenizer.model_max_length, return_tensors="np")
|
164 |
+
logits = self.model.predict([inputs["input_ids"], inputs["attention_mask"]])
|
165 |
+
probabilities = tf.nn.softmax(logits).numpy()
|
166 |
+
label_id = np.argmax(probabilities, axis=-1).item()
|
167 |
+
return {
|
168 |
+
"label": f"LABEL_{label_id}",
|
169 |
+
"score": probabilities.max()
|
170 |
+
}
|
171 |
+
else:
|
172 |
+
return self.classifier(text)[0]
|
173 |
+
|
174 |
+
def classify_dataframe_column(self, df, target_column, feature_suffix):
|
175 |
+
tqdm.pandas()
|
176 |
+
df[f'trimmed_{target_column}'] = df[target_column].progress_apply(
|
177 |
+
lambda text: self.tokenizer.decode(
|
178 |
+
self.tokenizer(text, truncation=True, max_length=self.tokenizer.model_max_length)["input_ids"],
|
179 |
+
skip_special_tokens=True
|
180 |
+
)
|
181 |
+
)
|
182 |
+
|
183 |
+
if self.is_tensorflow:
|
184 |
+
results = [self.classify(text) for text in df[f'trimmed_{target_column}']]
|
185 |
+
else:
|
186 |
+
results = [self.classifier(text)[0] for text in df[f'trimmed_{target_column}']]
|
187 |
+
|
188 |
+
df[f'pred_label_{feature_suffix}'] = [
|
189 |
+
self.label_map[int(result['label'].split('_')[-1])] for result in results
|
190 |
+
]
|
191 |
+
df[f'prob_{feature_suffix}'] = [result['score'] for result in results]
|
192 |
+
df.drop(columns=[f'trimmed_{target_column}'], inplace=True)
|
193 |
+
return df
|
194 |
+
|
195 |
+
|
196 |
+
class ZeroShotClassifier(Classifier):
|
197 |
+
|
198 |
+
def __init__(self, model_path, tokenizer_path, candidate_labels):
|
199 |
+
self.model_path = model_path
|
200 |
+
self.candidate_labels = candidate_labels
|
201 |
+
self.classifier = pipeline("zero-shot-classification", model=model_path, tokenizer=tokenizer_path, clean_up_tokenization_spaces=True, device=0 if torch.cuda.is_available() else -1)
|
202 |
+
|
203 |
+
def classify_text(self, text, top_n=None, multi_label=False):
|
204 |
+
"""
|
205 |
+
Classify a single text using zero-shot classification with truncated scores.
|
206 |
+
|
207 |
+
:param text: The text to classify
|
208 |
+
:param multi_label: Whether to allow multi-label classification
|
209 |
+
:return: Classification result as a dictionary with scores truncated to 3 decimals
|
210 |
+
"""
|
211 |
+
classification_output = self.classifier(text, self.candidate_labels, multi_label=multi_label, clean_up_tokenization_spaces=True)
|
212 |
+
classification_output['scores'] = [round(score, 3) for score in classification_output['scores']]
|
213 |
+
if top_n is not None:
|
214 |
+
classification_output = {
|
215 |
+
'sequence': classification_output['sequence'],
|
216 |
+
'labels': classification_output['labels'][:top_n],
|
217 |
+
'scores': classification_output['scores'][:top_n]
|
218 |
+
}
|
219 |
+
return classification_output
|
220 |
+
|
221 |
+
def classify_dataframe_column(self, df, target_column, feature_suffix, multi_label=False):
|
222 |
+
"""
|
223 |
+
Classify the contents of a dataframe column using zero-shot classification.
|
224 |
+
|
225 |
+
:param df: The dataframe to process
|
226 |
+
:param target_column: The column containing text to classify
|
227 |
+
:param feature_suffix: Suffix for the output columns
|
228 |
+
:param multi_label: Whether to allow multi-label classification
|
229 |
+
:return: The dataframe with classification results
|
230 |
+
"""
|
231 |
+
tqdm.pandas()
|
232 |
+
|
233 |
+
# Apply the classify_text method to each row
|
234 |
+
results = df[target_column].progress_apply(
|
235 |
+
lambda text: self.classify_text(text, multi_label=multi_label)
|
236 |
+
)
|
237 |
+
|
238 |
+
# Extract and store results
|
239 |
+
df[f'top_class_{feature_suffix}'] = results.apply(lambda res: res['labels'][0])
|
240 |
+
df[f'top_score_{feature_suffix}'] = results.apply(lambda res: res['scores'][0])
|
241 |
+
df[f'full_results_{feature_suffix}'] = results.apply(lambda res: list(zip(res['labels'], res['scores'])))
|
242 |
+
|
243 |
+
return df
|
244 |
+
|
245 |
+
def test_zs_predictions(self, df, target_column='text', true_classes_column='category', plot_conf_matrix=True):
|
246 |
+
"""
|
247 |
+
Tests model predictions on a given dataset column using the zero-shot classification pipeline.
|
248 |
+
|
249 |
+
Args:
|
250 |
+
df (pd.DataFrame): Input dataframe containing texts for zero-shot classification.
|
251 |
+
target_column (str): The name of the column containing text to classify.
|
252 |
+
true_classes_column (str): The column containing annotated classes.
|
253 |
+
|
254 |
+
Returns:
|
255 |
+
dict: A dictionary containing accuracy, F1 score, and confusion matrix.
|
256 |
+
"""
|
257 |
+
# Progress bar for classification
|
258 |
+
tqdm.pandas(desc=f"Zero-shot classification with {self.model_path}")
|
259 |
+
|
260 |
+
# Function to classify each row
|
261 |
+
def classify_row(row):
|
262 |
+
classification_output = self.classifier(
|
263 |
+
row[target_column],
|
264 |
+
self.candidate_labels,
|
265 |
+
multi_label=False,
|
266 |
+
clean_up_tokenization_spaces=True,
|
267 |
+
)
|
268 |
+
return classification_output["labels"][0]
|
269 |
+
|
270 |
+
# Apply classification with progress bar
|
271 |
+
df = df.copy()
|
272 |
+
df.loc[:, 'predicted_class'] = df.progress_apply(classify_row, axis=1)
|
273 |
+
|
274 |
+
# Extract true and predicted classes
|
275 |
+
true_classes = df[true_classes_column]
|
276 |
+
predicted_classes = df['predicted_class']
|
277 |
+
|
278 |
+
# Compute metrics
|
279 |
+
accuracy = accuracy_score(true_classes, predicted_classes)
|
280 |
+
f1 = f1_score(true_classes, predicted_classes, average="macro")
|
281 |
+
cm = confusion_matrix(true_classes, predicted_classes, labels=self.candidate_labels)
|
282 |
+
if plot_conf_matrix:
|
283 |
+
disp = ConfusionMatrixDisplay(confusion_matrix=cm, display_labels=self.candidate_labels)
|
284 |
+
fig, ax = plt.subplots(figsize=(4, 4))
|
285 |
+
disp.plot(cmap=plt.cm.Blues, ax=ax, colorbar=False)
|
286 |
+
ax.set_title(f"Zero-shot classification with {self.model_path}", fontsize=10)
|
287 |
+
ax.set_xlabel("Predicted label", fontsize=8)
|
288 |
+
ax.set_ylabel("True label", fontsize=8)
|
289 |
+
|
290 |
+
ax.set_xticklabels(ax.get_xticklabels(), rotation=45, ha="right", fontsize=8)
|
291 |
+
ax.set_yticklabels(ax.get_yticklabels(), fontsize=8)
|
292 |
+
|
293 |
+
fig.text(
|
294 |
+
0.5, 0.01,
|
295 |
+
f"Accuracy: {accuracy:.4f} | F1 Score: {f1:.4f}",
|
296 |
+
ha="center",
|
297 |
+
fontsize=10
|
298 |
+
)
|
299 |
+
plt.tight_layout(rect=[0, 0.05, 1, 1]) # Adjust bottom margin
|
300 |
+
plt.show()
|
301 |
+
|
302 |
+
return {
|
303 |
+
"accuracy": accuracy,
|
304 |
+
"f1_score": f1,
|
305 |
+
"confusion_matrix": cm,
|
306 |
+
"detailed_results": df.to_dict(), # Full dataframe with predictions
|
307 |
+
}
|
308 |
+
|
309 |
+
def test_zs_predictions_with_dataset(self, df, target_column='text', true_classes_column='category', plot_conf_matrix=True):
|
310 |
+
dataset = Dataset.from_pandas(df)
|
311 |
+
def classify_text(batch):
|
312 |
+
classification_output = self.classifier(
|
313 |
+
batch[target_column],
|
314 |
+
self.candidate_labels,
|
315 |
+
multi_label=False,
|
316 |
+
clean_up_tokenization_spaces=True,
|
317 |
+
)
|
318 |
+
return {
|
319 |
+
"predicted_class": classification_output["labels"][0],
|
320 |
+
"predicted_scores": classification_output["scores"],
|
321 |
+
}
|
322 |
+
|
323 |
+
# Apply classification to the dataset
|
324 |
+
classified_dataset = dataset.map(classify_text, batched=False)
|
325 |
+
# classified_dataset = dataset.map(classify_text, batched=True, batch_size=16)
|
326 |
+
|
327 |
+
# Extract true and predicted classes
|
328 |
+
true_classes = classified_dataset[true_classes_column]
|
329 |
+
predicted_classes = classified_dataset["predicted_class"]
|
330 |
+
|
331 |
+
# Compute metrics
|
332 |
+
accuracy = accuracy_score(true_classes, predicted_classes)
|
333 |
+
f1 = f1_score(true_classes, predicted_classes, average="macro")
|
334 |
+
|
335 |
+
# Print metrics
|
336 |
+
print(f"Accuracy: {accuracy:.4f}")
|
337 |
+
print(f"F1 Score: {f1:.4f}")
|
338 |
+
|
339 |
+
# Generate confusion matrix:
|
340 |
+
cm = confusion_matrix(true_classes, predicted_classes, labels=self.candidate_labels)
|
341 |
+
if plot_conf_matrix:
|
342 |
+
disp = ConfusionMatrixDisplay(confusion_matrix=cm, display_labels=self.candidate_labels)
|
343 |
+
fig, ax = plt.subplots(figsize=(6, 6))
|
344 |
+
disp.plot(cmap=plt.cm.Blues, ax=ax)
|
345 |
+
plt.xticks(rotation=45, ha="right")
|
346 |
+
plt.show()
|
347 |
+
|
348 |
+
# Return metrics for further inspection
|
349 |
+
return {
|
350 |
+
"accuracy": accuracy,
|
351 |
+
"f1_score": f1,
|
352 |
+
"confusion_matrix": cm,
|
353 |
+
"detailed_results": classified_dataset.to_dict(),
|
354 |
+
}
|
355 |
+
|
356 |
+
class MetricsComparison:
|
357 |
+
def __init__(self, base_classifier, fine_tuned_classifier, base_metrics, fine_tuned_metrics):
|
358 |
+
self.base_classifier = base_classifier
|
359 |
+
self.fine_tuned_classifier = fine_tuned_classifier
|
360 |
+
self.base_metrics = base_metrics
|
361 |
+
self.fine_tuned_metrics = fine_tuned_metrics
|
362 |
+
|
363 |
+
def compare_conf_matrices(self):
|
364 |
+
fig, axes = plt.subplots(1, 2, figsize=(12, 6))
|
365 |
+
# Plot for base_classifier (left)
|
366 |
+
disp1 = ConfusionMatrixDisplay(confusion_matrix=self.base_metrics["confusion_matrix"],
|
367 |
+
display_labels=self.base_classifier.candidate_labels)
|
368 |
+
disp1.plot(cmap=plt.cm.Blues, ax=axes[0], colorbar=False)
|
369 |
+
axes[0].set_title(f"Zero-shot classification with {self.base_classifier.model_path}", fontsize=10)
|
370 |
+
axes[0].set_xlabel("Predicted class", fontsize=8)
|
371 |
+
axes[0].set_ylabel("True class", fontsize=8)
|
372 |
+
axes[0].set_xticklabels(axes[0].get_xticklabels(), rotation=45, ha="right", fontsize=8)
|
373 |
+
axes[0].set_yticklabels(axes[0].get_yticklabels(), fontsize=8)
|
374 |
+
|
375 |
+
fig.text(
|
376 |
+
0.25, 0.01,
|
377 |
+
f"Accuracy: {self.base_metrics['accuracy']:.4f} | F1 Score: {self.base_metrics['f1_score']:.4f}",
|
378 |
+
ha="center",
|
379 |
+
fontsize=10
|
380 |
+
)
|
381 |
+
|
382 |
+
# Plot for zs_classifier (fine-tuned) (right)
|
383 |
+
disp2 = ConfusionMatrixDisplay(confusion_matrix=self.fine_tuned_metrics["confusion_matrix"],
|
384 |
+
display_labels=self.fine_tuned_classifier.candidate_labels)
|
385 |
+
disp2.plot(cmap=plt.cm.Blues, ax=axes[1], colorbar=False)
|
386 |
+
axes[1].set_title(f"ZS classification with {self.fine_tuned_classifier.model_path}", fontsize=10)
|
387 |
+
axes[1].set_xlabel("Predicted class", fontsize=8)
|
388 |
+
axes[1].set_ylabel("True class", fontsize=8)
|
389 |
+
axes[1].set_xticklabels(axes[1].get_xticklabels(), rotation=45, ha="right", fontsize=8)
|
390 |
+
axes[1].set_yticklabels(axes[1].get_yticklabels(), fontsize=8)
|
391 |
+
|
392 |
+
fig.text(
|
393 |
+
0.75, 0.01,
|
394 |
+
f"Accuracy: {self.fine_tuned_metrics['accuracy']:.4f} | F1 Score: {self.fine_tuned_metrics['f1_score']:.4f}",
|
395 |
+
ha="center",
|
396 |
+
fontsize=10
|
397 |
+
)
|
398 |
+
|
399 |
+
plt.tight_layout(rect=[0, 0.05, 1, 0.95])
|
400 |
+
plt.show()
|
401 |
+
|