Spaces:
Sleeping
Sleeping
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,601 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#python app.py
|
2 |
+
import gradio as gr
|
3 |
+
import os
|
4 |
+
import pandas as pd
|
5 |
+
import requests
|
6 |
+
from pathlib import Path
|
7 |
+
import ctranslate2
|
8 |
+
import time
|
9 |
+
import logging
|
10 |
+
import transformers
|
11 |
+
import json
|
12 |
+
from tqdm import tqdm
|
13 |
+
import subprocess
|
14 |
+
from huggingface_hub import snapshot_download, upload_file
|
15 |
+
|
16 |
+
# Function to download a Parquet file from a specified URL
|
17 |
+
def download_parquet(url, local_path):
|
18 |
+
response = requests.get(url, stream=True)
|
19 |
+
if response.status_code == 200:
|
20 |
+
with open(local_path, 'wb') as file:
|
21 |
+
for chunk in response.iter_content(chunk_size=1024):
|
22 |
+
file.write(chunk)
|
23 |
+
print("File downloaded successfully.")
|
24 |
+
else:
|
25 |
+
print(f"Failed to download file, status code: {response.status_code}")
|
26 |
+
|
27 |
+
# Function to convert Parquet files to JSONL format
|
28 |
+
def convert_parquet_to_jsonl_polars(input_file, output_dir, override=False):
|
29 |
+
output_dir_path = Path(output_dir)
|
30 |
+
output_dir_path.mkdir(parents=True, exist_ok=True)
|
31 |
+
|
32 |
+
input_path = Path(input_file)
|
33 |
+
output_file_path = output_dir_path / input_path.with_suffix(".jsonl").name
|
34 |
+
|
35 |
+
if output_file_path.exists() and not override:
|
36 |
+
print(f"Skipping because output exists already: {output_file_path}")
|
37 |
+
else:
|
38 |
+
df = pl.read_parquet(input_path)
|
39 |
+
df.write_ndjson(output_file_path)
|
40 |
+
print(f"Data written to {output_file_path}")
|
41 |
+
|
42 |
+
def convert_parquet_to_jsonl(parquet_filename, jsonl_filename):
|
43 |
+
# Read the parquet file
|
44 |
+
df = pd.read_parquet(parquet_filename)
|
45 |
+
|
46 |
+
# Convert the dataframe to a JSON string and handle Unicode characters and forward slashes
|
47 |
+
json_str = df.to_json(orient='records', lines=True, force_ascii=False)
|
48 |
+
|
49 |
+
# Replace escaped forward slashes if needed
|
50 |
+
json_str = json_str.replace('\\/', '/')
|
51 |
+
|
52 |
+
# Write the modified JSON string to the JSONL file
|
53 |
+
with open(jsonl_filename, 'w', encoding='utf-8') as file:
|
54 |
+
file.write(json_str)
|
55 |
+
|
56 |
+
print(f"Data saved to {jsonl_filename}")
|
57 |
+
|
58 |
+
# Function to count lines in a JSONL file
|
59 |
+
def count_lines_in_jsonl(file_path):
|
60 |
+
with open(file_path, 'r', encoding='utf-8') as file:
|
61 |
+
line_count = sum(1 for _ in file)
|
62 |
+
return line_count
|
63 |
+
|
64 |
+
def parse_range_specification(range_specification, file_length):
|
65 |
+
line_indices = []
|
66 |
+
ranges = range_specification.split(',')
|
67 |
+
for r in ranges:
|
68 |
+
if '-' in r:
|
69 |
+
parts = r.split('-')
|
70 |
+
start = int(parts[0]) - 1 if parts[0] else 0
|
71 |
+
end = int(parts[1]) - 1 if parts[1] else file_length - 1
|
72 |
+
if start < 0 or end >= file_length:
|
73 |
+
logging.error(f"Range {r} is out of bounds.")
|
74 |
+
continue # Skip ranges that are out of bounds
|
75 |
+
line_indices.extend(range(start, end + 1))
|
76 |
+
else:
|
77 |
+
single_line = int(r) - 1
|
78 |
+
if single_line < 0 or single_line >= file_length:
|
79 |
+
logging.error(f"Line number {r} is out of bounds.")
|
80 |
+
continue # Skip line numbers that are out of bounds
|
81 |
+
line_indices.append(single_line)
|
82 |
+
return line_indices
|
83 |
+
|
84 |
+
def translate_text(text, translator, tokenizer):
|
85 |
+
"""
|
86 |
+
Translates the given text from English to German using CTranslate2 and the WMT21 model,
|
87 |
+
with special handling for newlines and segmenting text longer than 500 characters.
|
88 |
+
Ensures sequences of newlines (\n\n, \n\n\n, etc.) are accurately reproduced.
|
89 |
+
"""
|
90 |
+
try:
|
91 |
+
segments = []
|
92 |
+
newline_sequences = [] # To store sequences of newlines
|
93 |
+
segment = ""
|
94 |
+
|
95 |
+
i = 0
|
96 |
+
while i < len(text):
|
97 |
+
# Collect sequences of newlines
|
98 |
+
if text[i] == '\n':
|
99 |
+
newline_sequence = '\n'
|
100 |
+
while i + 1 < len(text) and text[i + 1] == '\n':
|
101 |
+
newline_sequence += '\n'
|
102 |
+
i += 1
|
103 |
+
if segment:
|
104 |
+
segments.append(segment) # Add the preceding text segment
|
105 |
+
segment = ""
|
106 |
+
newline_sequences.append(newline_sequence) # Store the newline sequence
|
107 |
+
else:
|
108 |
+
segment += text[i]
|
109 |
+
# If segment exceeds 500 characters, or if we reach the end of the text, process it
|
110 |
+
if len(segment) >= 500 or i == len(text) - 1:
|
111 |
+
end_index = max(segment.rfind('.', 0, 500), segment.rfind('?', 0, 500), segment.rfind('!', 0, 500))
|
112 |
+
if end_index != -1 and len(segment) > 500:
|
113 |
+
# Split at the last punctuation within the first 500 characters
|
114 |
+
segments.append(segment[:end_index+1])
|
115 |
+
segment = segment[end_index+1:].lstrip()
|
116 |
+
else:
|
117 |
+
# No suitable punctuation or end of text, add the whole segment
|
118 |
+
segments.append(segment)
|
119 |
+
segment = ""
|
120 |
+
i += 1
|
121 |
+
|
122 |
+
# Translate the collected text segments
|
123 |
+
translated_segments = []
|
124 |
+
for segment in segments:
|
125 |
+
source = tokenizer.convert_ids_to_tokens(tokenizer.encode(segment))
|
126 |
+
target_prefix = [tokenizer.lang_code_to_token["de"]]
|
127 |
+
results = translator.translate_batch([source], target_prefix=[target_prefix])
|
128 |
+
target = results[0].hypotheses[0][1:]
|
129 |
+
translated_segment = tokenizer.decode(tokenizer.convert_tokens_to_ids(target))
|
130 |
+
translated_segments.append(translated_segment)
|
131 |
+
|
132 |
+
# Reassemble the translated text with original newline sequences
|
133 |
+
translated_text = ""
|
134 |
+
for i, segment in enumerate(translated_segments):
|
135 |
+
translated_text += segment
|
136 |
+
if i < len(newline_sequences):
|
137 |
+
translated_text += newline_sequences[i] # Insert the newline sequence
|
138 |
+
|
139 |
+
return translated_text.strip()
|
140 |
+
|
141 |
+
except Exception as e:
|
142 |
+
logging.error(f"An error occurred during translation: {e}")
|
143 |
+
return None
|
144 |
+
|
145 |
+
def translate_item_ufb(item, raw_file_path, translator, tokenizer):
|
146 |
+
try:
|
147 |
+
# Translate the prompt directly since it's a string
|
148 |
+
translated_prompt = translate_text(item['prompt'], translator, tokenizer)
|
149 |
+
|
150 |
+
# Translate the chosen and rejected contents
|
151 |
+
translated_chosen = []
|
152 |
+
for choice in item['chosen']:
|
153 |
+
translated_content = translate_text(choice['content'], translator, tokenizer)
|
154 |
+
translated_chosen.append({'content': translated_content, 'role': choice['role']})
|
155 |
+
|
156 |
+
translated_rejected = []
|
157 |
+
for choice in item['rejected']:
|
158 |
+
translated_content = translate_text(choice['content'], translator, tokenizer)
|
159 |
+
translated_rejected.append({'content': translated_content, 'role': choice['role']})
|
160 |
+
|
161 |
+
# Write the raw response to a backup file
|
162 |
+
with open(raw_file_path, 'a', encoding='utf-8') as raw_file:
|
163 |
+
raw_file.write(f"Prompt: {translated_prompt}\n")
|
164 |
+
raw_file.write(f"Chosen: {json.dumps(translated_chosen, ensure_ascii=False)}\n")
|
165 |
+
raw_file.write(f"Rejected: {json.dumps(translated_rejected, ensure_ascii=False)}\n\n")
|
166 |
+
|
167 |
+
logging.info("Translation request successful.")
|
168 |
+
# Update the original item with the translated fields
|
169 |
+
item['prompt'] = translated_prompt
|
170 |
+
item['chosen'] = translated_chosen
|
171 |
+
item['rejected'] = translated_rejected
|
172 |
+
return item
|
173 |
+
|
174 |
+
except Exception as e:
|
175 |
+
logging.error(f"An error occurred during translation: {e}")
|
176 |
+
return None
|
177 |
+
|
178 |
+
def validate_item_ufb(item):
|
179 |
+
# Check basic required fields including 'prompt' as a simple string
|
180 |
+
required_fields = ['source', 'prompt', 'chosen', 'rejected']
|
181 |
+
for field in required_fields:
|
182 |
+
if field not in item:
|
183 |
+
logging.warning(f"Missing required field: {field}")
|
184 |
+
return False
|
185 |
+
if field == 'prompt' and not isinstance(item['prompt'], str):
|
186 |
+
logging.warning("Prompt must be a string.")
|
187 |
+
return False
|
188 |
+
|
189 |
+
# Check 'chosen' and 'rejected' which should be lists of dictionaries
|
190 |
+
for field in ['chosen', 'rejected']:
|
191 |
+
if not isinstance(item[field], list) or not item[field]:
|
192 |
+
logging.warning(f"No entries or incorrect type for section: {field}")
|
193 |
+
return False
|
194 |
+
for idx, message in enumerate(item[field]):
|
195 |
+
if 'content' not in message or 'role' not in message:
|
196 |
+
logging.warning(f"Missing 'content' or 'role' field in {field} at index {idx}")
|
197 |
+
return False
|
198 |
+
if not isinstance(message['content'], str) or not isinstance(message['role'], str):
|
199 |
+
logging.warning(f"Invalid type for 'content' or 'role' field in {field} at index {idx}")
|
200 |
+
return False
|
201 |
+
|
202 |
+
return True
|
203 |
+
|
204 |
+
|
205 |
+
|
206 |
+
def translate_item_mix(item, raw_file_path, translator, tokenizer):
|
207 |
+
"""
|
208 |
+
Translates the relevant fields in the given item from English to German using CTranslate2 and the WMT21 model,
|
209 |
+
and saves the raw response to a backup file.
|
210 |
+
"""
|
211 |
+
#print ("translating:", item)
|
212 |
+
try:
|
213 |
+
# Translate each part of the prompt separately and preserve the order
|
214 |
+
translated_prompts = []
|
215 |
+
for message in item['prompt']:
|
216 |
+
translated_content = translate_text(message['content'], translator, tokenizer)
|
217 |
+
translated_prompts.append({'content': translated_content, 'role': message['role']})
|
218 |
+
|
219 |
+
# Translate the chosen and rejected contents
|
220 |
+
translated_chosen_content = translate_text(item['chosen'][0]['content'], translator, tokenizer)
|
221 |
+
translated_rejected_content = translate_text(item['rejected'][0]['content'], translator, tokenizer)
|
222 |
+
|
223 |
+
# Write the raw response to a backup file
|
224 |
+
with open(raw_file_path, 'a', encoding='utf-8') as raw_file:
|
225 |
+
raw_file.write("Prompt content:\n")
|
226 |
+
for translated_prompt in translated_prompts:
|
227 |
+
raw_file.write(f"{translated_prompt['role']}: {translated_prompt['content']}\n")
|
228 |
+
raw_file.write(f"Chosen content: {translated_chosen_content}\n")
|
229 |
+
raw_file.write(f"Rejected content: {translated_rejected_content}\n\n")
|
230 |
+
|
231 |
+
logging.info("Translation request successful.")
|
232 |
+
except Exception as e:
|
233 |
+
logging.error(f"An error occurred during translation: {e}")
|
234 |
+
return None
|
235 |
+
|
236 |
+
# Update the original item with the translated fields
|
237 |
+
item['prompt'] = translated_prompts
|
238 |
+
item['chosen'][0]['content'] = translated_chosen_content
|
239 |
+
item['rejected'][0]['content'] = translated_rejected_content
|
240 |
+
|
241 |
+
logging.info("Translation processing successful.")
|
242 |
+
return item
|
243 |
+
|
244 |
+
def validate_item_mix(item):
|
245 |
+
"""
|
246 |
+
Validates the structure, presence, and content of required fields in the given item,
|
247 |
+
allowing for multiple elements in the 'prompt' field for multi-turn conversations.
|
248 |
+
"""
|
249 |
+
required_fields = ['dataset', 'prompt', 'chosen', 'rejected']
|
250 |
+
for field in required_fields:
|
251 |
+
if field not in item:
|
252 |
+
logging.warning(f"Missing required field: {field}")
|
253 |
+
return False
|
254 |
+
|
255 |
+
# Check for at least one element in 'prompt' and exactly one element in 'chosen' and 'rejected'
|
256 |
+
if len(item['prompt']) < 1 or len(item['chosen']) != 1 or len(item['rejected']) != 1:
|
257 |
+
logging.warning("Invalid number of elements in 'prompt', 'chosen', or 'rejected' field.")
|
258 |
+
return False
|
259 |
+
|
260 |
+
# Validate 'content' and 'role' fields in all messages of 'prompt', and single elements of 'chosen' and 'rejected'
|
261 |
+
for choice in item['prompt'] + item['chosen'] + item['rejected']:
|
262 |
+
if 'content' not in choice or 'role' not in choice:
|
263 |
+
logging.warning("Missing 'content' or 'role' field in choice.")
|
264 |
+
return False
|
265 |
+
if not isinstance(choice['content'], str) or not isinstance(choice['role'], str):
|
266 |
+
logging.warning("Invalid type for 'content' or 'role' field in choice.")
|
267 |
+
return False
|
268 |
+
|
269 |
+
return True
|
270 |
+
|
271 |
+
def translate_item_orpo(item, raw_file_path, translator, tokenizer):
|
272 |
+
try:
|
273 |
+
translated_texts = {} # Cache to store translated texts
|
274 |
+
|
275 |
+
# Translate the prompt if necessary (which is a user input and can appear again)
|
276 |
+
if item['prompt'] not in translated_texts:
|
277 |
+
translated_prompt = translate_text(item['prompt'], translator, tokenizer)
|
278 |
+
translated_texts[item['prompt']] = translated_prompt
|
279 |
+
else:
|
280 |
+
translated_prompt = translated_texts[item['prompt']]
|
281 |
+
|
282 |
+
# Helper function to handle content translation with caching
|
283 |
+
def get_translated_content(content):
|
284 |
+
if content not in translated_texts:
|
285 |
+
translated_texts[content] = translate_text(content, translator, tokenizer)
|
286 |
+
return translated_texts[content]
|
287 |
+
|
288 |
+
# Process translations for chosen and rejected sections
|
289 |
+
def translate_interactions(interactions):
|
290 |
+
translated_interactions = []
|
291 |
+
for interaction in interactions:
|
292 |
+
translated_content = get_translated_content(interaction['content'])
|
293 |
+
translated_interactions.append({'content': translated_content, 'role': interaction['role']})
|
294 |
+
return translated_interactions
|
295 |
+
|
296 |
+
translated_chosen = translate_interactions(item['chosen'])
|
297 |
+
translated_rejected = translate_interactions(item['rejected'])
|
298 |
+
|
299 |
+
# Write the raw response to a backup file
|
300 |
+
with open(raw_file_path, 'a', encoding='utf-8') as raw_file:
|
301 |
+
raw_file.write(f"Prompt: {translated_prompt}\n")
|
302 |
+
raw_file.write(f"Chosen: {json.dumps(translated_chosen, ensure_ascii=False)}\n")
|
303 |
+
raw_file.write(f"Rejected: {json.dumps(translated_rejected, ensure_ascii=False)}\n\n")
|
304 |
+
|
305 |
+
logging.info("Translation request successful.")
|
306 |
+
# Update the original item with the translated fields
|
307 |
+
item['prompt'] = translated_prompt
|
308 |
+
item['chosen'] = translated_chosen
|
309 |
+
item['rejected'] = translated_rejected
|
310 |
+
return item
|
311 |
+
|
312 |
+
except Exception as e:
|
313 |
+
logging.error(f"An error occurred during translation: {e}")
|
314 |
+
return None
|
315 |
+
|
316 |
+
def validate_item_orpo(item):
|
317 |
+
# Check basic required fields
|
318 |
+
required_fields = ['source', 'prompt', 'chosen', 'rejected']
|
319 |
+
for field in required_fields:
|
320 |
+
if field not in item:
|
321 |
+
logging.warning(f"Missing required field: {field}")
|
322 |
+
return False
|
323 |
+
|
324 |
+
# Ensure 'prompt' is a string
|
325 |
+
if not isinstance(item['prompt'], str):
|
326 |
+
logging.warning("Prompt must be a string.")
|
327 |
+
return False
|
328 |
+
|
329 |
+
# Check 'chosen' and 'rejected' which should be lists of dictionaries
|
330 |
+
for field in ['chosen', 'rejected']:
|
331 |
+
if not isinstance(item[field], list) or not item[field]:
|
332 |
+
logging.warning(f"No entries or incorrect type for section: {field}")
|
333 |
+
return False
|
334 |
+
for idx, message in enumerate(item[field]):
|
335 |
+
if 'content' not in message or 'role' not in message:
|
336 |
+
logging.warning(f"Missing 'content' or 'role' field in {field} at index {idx}")
|
337 |
+
return False
|
338 |
+
if not isinstance(message['content'], str) or not isinstance(message['role'], str):
|
339 |
+
logging.warning(f"Invalid type for 'content' or 'role' field in {field} at index {idx}")
|
340 |
+
return False
|
341 |
+
|
342 |
+
return True
|
343 |
+
|
344 |
+
def process_file(input_file_path, output_file_path, raw_file_path, line_indices, translator, tokenizer, model_type):
|
345 |
+
try:
|
346 |
+
# Assigning validation and translation functions based on model_type
|
347 |
+
if model_type == "mix":
|
348 |
+
print ("translating a mix-style model...")
|
349 |
+
validate_item = validate_item_mix
|
350 |
+
translate_item = translate_item_mix
|
351 |
+
elif model_type == "orpo":
|
352 |
+
print ("translating an orpo-style model...")
|
353 |
+
validate_item = validate_item_orpo
|
354 |
+
translate_item = translate_item_orpo # def translate_item_ufb(item, raw_file_path, translator, tokenizer):
|
355 |
+
elif model_type == "ufb":
|
356 |
+
print ("translating an ultrafeedback-style model...")
|
357 |
+
validate_item = validate_item_ufb
|
358 |
+
translate_item = translate_item_ufb # def translate_item_ufb(item, raw_file_path, translator, tokenizer):
|
359 |
+
else:
|
360 |
+
raise ValueError(f"Unsupported model_type: {model_type}")
|
361 |
+
|
362 |
+
with open(input_file_path, 'r', encoding='utf-8') as file:
|
363 |
+
data_points = [json.loads(line) for line in file]
|
364 |
+
|
365 |
+
failed_items = []
|
366 |
+
failed_items_indices = []
|
367 |
+
|
368 |
+
for index in tqdm(line_indices, desc="Processing lines", unit="item"):
|
369 |
+
item = data_points[index]
|
370 |
+
|
371 |
+
# Validate the item structure
|
372 |
+
if not validate_item(item):
|
373 |
+
logging.warning("Skipping item due to invalid structure.")
|
374 |
+
failed_items.append(item)
|
375 |
+
continue
|
376 |
+
|
377 |
+
# Translate the relevant fields in the item
|
378 |
+
translated_item = None
|
379 |
+
retry_count = 0
|
380 |
+
while translated_item is None and retry_count < 3:
|
381 |
+
print ("going to translate the item...")
|
382 |
+
translated_item = translate_item(item, raw_file_path, translator, tokenizer)
|
383 |
+
retry_count += 1
|
384 |
+
if translated_item is None:
|
385 |
+
logging.warning(f"Translation failed for item. Retry attempt: {retry_count}")
|
386 |
+
time.sleep(1)
|
387 |
+
|
388 |
+
if translated_item is not None:
|
389 |
+
translated_item['index'] = index
|
390 |
+
with open(output_file_path, 'a', encoding='utf-8') as file:
|
391 |
+
file.write(json.dumps(translated_item, ensure_ascii=False) + "\n")
|
392 |
+
else:
|
393 |
+
failed_items_indices.append(index)
|
394 |
+
failed_items.append(item)
|
395 |
+
logging.error("Translation failed after multiple attempts. Skipping item.")
|
396 |
+
|
397 |
+
# Validate the translated item structure
|
398 |
+
if not validate_item(translated_item):
|
399 |
+
logging.warning("Skipping translated item due to invalid structure.")
|
400 |
+
failed_items.append(item)
|
401 |
+
continue
|
402 |
+
|
403 |
+
with open('failed_items.jsonl', 'w', encoding='utf-8') as file:
|
404 |
+
for item in failed_items:
|
405 |
+
file.write(json.dumps(item, ensure_ascii=False) + "\n")
|
406 |
+
|
407 |
+
failed_items_str = generate_failed_items_str(failed_items_indices)
|
408 |
+
with open('failed_items_index.txt', 'w', encoding='utf-8') as f:
|
409 |
+
f.write(failed_items_str)
|
410 |
+
|
411 |
+
logging.info("Translation completed successfully.")
|
412 |
+
|
413 |
+
except Exception as e:
|
414 |
+
logging.error(f"An error occurred: {e}")
|
415 |
+
|
416 |
+
def generate_failed_items_str(indices):
|
417 |
+
"""
|
418 |
+
Converts a list of failed item indices into a string.
|
419 |
+
"""
|
420 |
+
if not indices:
|
421 |
+
return ""
|
422 |
+
|
423 |
+
# Sort the list of indices and initialize the first range
|
424 |
+
indices.sort()
|
425 |
+
range_start = indices[0]
|
426 |
+
current = range_start
|
427 |
+
ranges = []
|
428 |
+
|
429 |
+
for i in indices[1:]:
|
430 |
+
if i == current + 1:
|
431 |
+
current = i
|
432 |
+
else:
|
433 |
+
if range_start == current:
|
434 |
+
ranges.append(f"{range_start}")
|
435 |
+
else:
|
436 |
+
ranges.append(f"{range_start}-{current}")
|
437 |
+
range_start = current = i
|
438 |
+
|
439 |
+
# Add the last range
|
440 |
+
if range_start == current:
|
441 |
+
ranges.append(f"{range_start}")
|
442 |
+
else:
|
443 |
+
ranges.append(f"{range_start}-{current}")
|
444 |
+
|
445 |
+
return ",".join(ranges)
|
446 |
+
|
447 |
+
# Function to upload the output file to Hugging Face
|
448 |
+
def upload_output_to_huggingface(output_file_path, repo_name, token):
|
449 |
+
upload_file(
|
450 |
+
path_or_fileobj=output_file_path,
|
451 |
+
path_in_repo=output_file_path,
|
452 |
+
repo_id=repo_name,
|
453 |
+
repo_type="dataset",
|
454 |
+
token=token
|
455 |
+
)
|
456 |
+
print(f"Uploaded {output_file_path} to Hugging Face repository: {repo_name}")
|
457 |
+
|
458 |
+
def translate_dataset(train_url, local_parquet_path, input_file_path, output_file_path, raw_file_path, range_specification, model_type, output_dir, output_repo_name, token, translator, tokenizer):
|
459 |
+
try:
|
460 |
+
# Download the Parquet file
|
461 |
+
download_parquet(train_url, local_parquet_path)
|
462 |
+
except Exception as e:
|
463 |
+
logging.error(f"Failed to download the Parquet file from {train_url}: {e}")
|
464 |
+
return
|
465 |
+
|
466 |
+
try:
|
467 |
+
# Convert the downloaded Parquet file to JSONL
|
468 |
+
convert_parquet_to_jsonl(local_parquet_path, output_dir)
|
469 |
+
except Exception as e:
|
470 |
+
logging.error(f"Failed to convert Parquet to JSONL: {e}")
|
471 |
+
return
|
472 |
+
|
473 |
+
try:
|
474 |
+
# Rename the JSONL file using subprocess to ensure correct handling
|
475 |
+
subprocess.run(["mv", f"{output_dir}/train.jsonl", input_file_path], check=True)
|
476 |
+
except subprocess.CalledProcessError as e:
|
477 |
+
logging.error(f"Failed to rename the file from 'train.jsonl' to {input_file_path}: {e}")
|
478 |
+
return
|
479 |
+
|
480 |
+
try:
|
481 |
+
# Count lines in the JSONL file to validate contents
|
482 |
+
line_count = count_lines_in_jsonl(input_file_path)
|
483 |
+
logging.info(f"Number of lines in the file: {line_count}")
|
484 |
+
except Exception as e:
|
485 |
+
logging.error(f"Failed to count lines in {input_file_path}: {e}")
|
486 |
+
return
|
487 |
+
|
488 |
+
try:
|
489 |
+
# Parse the range specification for processing specific lines
|
490 |
+
line_indices = parse_range_specification(range_specification, file_length=line_count)
|
491 |
+
if not line_indices:
|
492 |
+
logging.error("No valid line indices to process. Please check the range specifications.")
|
493 |
+
return
|
494 |
+
except Exception as e:
|
495 |
+
logging.error(f"Error parsing range specification '{range_specification}': {e}")
|
496 |
+
return
|
497 |
+
|
498 |
+
try:
|
499 |
+
# Process the file with specified model type and line indices
|
500 |
+
process_file(input_file_path, output_file_path, raw_file_path, line_indices, translator, tokenizer, model_type)
|
501 |
+
except Exception as e:
|
502 |
+
logging.error(f"Failed to process the file {input_file_path}: {e}")
|
503 |
+
return
|
504 |
+
|
505 |
+
try:
|
506 |
+
# Upload the output file to Hugging Face repository
|
507 |
+
upload_output_to_huggingface(output_file_path, output_repo_name, token)
|
508 |
+
except Exception as e:
|
509 |
+
logging.error(f"Failed to upload {output_file_path} to Hugging Face: {e}")
|
510 |
+
|
511 |
+
# Setup logging configuration
|
512 |
+
logging.basicConfig(level=logging.INFO, filename='translation.log', filemode='a',
|
513 |
+
format='%(asctime)s - %(levelname)s - %(message)s')
|
514 |
+
|
515 |
+
def main(model_id, dataset_url, model_type, output_dataset_name):
|
516 |
+
try:
|
517 |
+
# Login to Hugging Face
|
518 |
+
token = login()
|
519 |
+
if token:
|
520 |
+
logging.info("Logged in to Hugging Face")
|
521 |
+
|
522 |
+
# Configuration and paths
|
523 |
+
tokenizer_name = "facebook/wmt21-dense-24-wide-en-x"
|
524 |
+
model_repo_name = "cstr/wmt21ct2_int8" # Repository to download the model from
|
525 |
+
|
526 |
+
# Download the model snapshot from Hugging Face
|
527 |
+
model_path = snapshot_download(repo_id=model_repo_name, token=token)
|
528 |
+
logging.info(f"Model downloaded to: {model_path}")
|
529 |
+
|
530 |
+
# Load the CTranslate2 model
|
531 |
+
translator = ctranslate2.Translator(model_path, device="auto")
|
532 |
+
logging.info("CTranslate2 model loaded successfully.")
|
533 |
+
|
534 |
+
# Load the tokenizer
|
535 |
+
tokenizer = transformers.AutoTokenizer.from_pretrained(tokenizer_name)
|
536 |
+
tokenizer.src_lang = "en"
|
537 |
+
logging.info("Tokenizer loaded successfully.")
|
538 |
+
|
539 |
+
# Define the task based on user input
|
540 |
+
task = {
|
541 |
+
"url": dataset_url,
|
542 |
+
"local_path": "train.parquet",
|
543 |
+
"input_file": f"{model_type}_en.jsonl",
|
544 |
+
"output_file": f"{model_type}_de.jsonl",
|
545 |
+
"raw_file": f"{model_type}_de_raw.jsonl",
|
546 |
+
"range_spec": "1-",
|
547 |
+
"model_type": model_type
|
548 |
+
}
|
549 |
+
|
550 |
+
# Call the translate_dataset function with the provided parameters
|
551 |
+
translate_dataset(
|
552 |
+
train_url=task["url"],
|
553 |
+
local_parquet_path=task["local_path"],
|
554 |
+
input_file_path=task["input_file"],
|
555 |
+
output_file_path=task["output_file"],
|
556 |
+
output_dir=".",
|
557 |
+
output_repo_name=output_dataset_name,
|
558 |
+
raw_file_path=task["raw_file"],
|
559 |
+
token=token,
|
560 |
+
range_specification=task["range_spec"],
|
561 |
+
model_type=task["model_type"],
|
562 |
+
translator=translator,
|
563 |
+
tokenizer=tokenizer,
|
564 |
+
)
|
565 |
+
return "Dataset translation completed!"
|
566 |
+
else:
|
567 |
+
return "Login failed. Please try again."
|
568 |
+
except Exception as e:
|
569 |
+
logging.error(f"An error occurred in the main function: {e}")
|
570 |
+
return f"An error occurred: {e}"
|
571 |
+
|
572 |
+
# Gradio interface setup
|
573 |
+
gradio_title = "π§ WMT21 Dataset Translation"
|
574 |
+
gradio_desc = """This tool translates datasets using the WMT21 translation model.
|
575 |
+
## π What Does This Tool Do:
|
576 |
+
- Translates datasets based on the selected model type.
|
577 |
+
- Uploads the translated dataset to Hugging Face.
|
578 |
+
## π οΈ Backend:
|
579 |
+
The translation backend runs on the Hugging Face Hub API.
|
580 |
+
"""
|
581 |
+
|
582 |
+
with gr.Blocks() as demo:
|
583 |
+
gr.HTML(f"""<h1 align="center" id="space-title">{gradio_title}</h1>""")
|
584 |
+
gr.Markdown(gradio_desc)
|
585 |
+
|
586 |
+
with gr.Row(equal_height=False):
|
587 |
+
with gr.Column():
|
588 |
+
model_id = gr.Textbox(label="Model ID or URL", lines=1)
|
589 |
+
dataset_url = gr.Textbox(label="Dataset URL", lines=1)
|
590 |
+
model_type = gr.Dropdown(choices=["mix", "orpo", "ufb"], label="Model Type")
|
591 |
+
output_dataset_name = gr.Textbox(label="Output Dataset Name", lines=1)
|
592 |
+
login_button = gr.Button("Login to Hugging Face")
|
593 |
+
|
594 |
+
with gr.Column():
|
595 |
+
output = gr.Textbox(label="Output", lines=1)
|
596 |
+
logout_button = gr.Button("Logout")
|
597 |
+
|
598 |
+
submit_btn = gr.Button("Translate Dataset", variant="primary")
|
599 |
+
submit_btn.click(main, inputs=[model_id, dataset_url, model_type, output_dataset_name], outputs=output)
|
600 |
+
|
601 |
+
demo.launch()
|