rockdich / translate_final.py
sudoaza
adding final dictionary
3fd70ad
raw
history blame contribute delete
No virus
4.25 kB
"""Translate a password dictionary using a finetuned model."""
from unsloth import FastLanguageModel
import torch
import argparse
from tqdm import tqdm
max_seq_length = 2048 # Choose any! We auto support RoPE Scaling internally!
dtype = torch.float16 # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
load_in_4bit = True # Use 4bit quantization to reduce memory usage. Can be False.
model, tokenizer = FastLanguageModel.from_pretrained(
model_name = "unsloth/llama-3-8b-bnb-4bit",
max_seq_length = max_seq_length,
dtype = dtype,
load_in_4bit = load_in_4bit,
)
model = FastLanguageModel.get_peft_model(
model,
r = 16, # Choose any number > 0 ! Suggested 8, 16, 32, 64, 128
target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
"gate_proj", "up_proj", "down_proj",],
lora_alpha = 16,
lora_dropout = 0, # Supports any, but = 0 is optimized
bias = "none", # Supports any, but = "none" is optimized
# [NEW] "unsloth" uses 30% less VRAM, fits 2x larger batch sizes!
use_gradient_checkpointing = "unsloth", # True or "unsloth" for very long context
random_state = 3407,
use_rslora = False, # We support rank stabilized LoRA
loftq_config = None, # And LoftQ
)
import re
def extract_response(text):
# Define a regular expression to find the content after "### Response:"
# Using non-greedy matching to stop at the first potential end-of-text token or excessive newlines
match = re.search(r"### Response:\n(.*?)$", text, re.DOTALL)
if match:
response = match.group(1)
response = response.replace("<|end_of_text|>", "")
if response[-1] != "\n":
response = response + "\n"
return response
else:
raise "No response found in the text."
from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained(
model_name = "lora_model", # YOUR MODEL YOU USED FOR TRAINING
max_seq_length = max_seq_length,
dtype = dtype,
load_in_4bit = load_in_4bit,
)
FastLanguageModel.for_inference(model) # Enable native 2x faster inference
tokenizer.padding_side = "left"
alpaca_prompt = """Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.
### Instruction:
{}
### Input:
{}
### Response:
{}"""
def process_batch(batch):
inputs = []
chunk_size = 10 # You can adjust the chunk size based on your needs
for i in range(0, len(batch), chunk_size):
chunk = ''.join(batch[i:i+chunk_size])
inputs.append(alpaca_prompt.format(
"Translate this passwords while keeping the original format.", # instruction
chunk, # input
"", # output - leave this blank for generation!
))
input_tokens = tokenizer(inputs, return_tensors = "pt", padding=True).to("cuda")
outputs = model.generate(**input_tokens, max_new_tokens = 64, use_cache = True)
return [extract_response(response) for response in tokenizer.batch_decode(outputs)]
BATCH_SIZE = 1000
def process_file(infile, outfile):
try:
with open(infile, 'r', encoding='latin1') as file:
lines = file.readlines()
translated_lines = []
# use tqdm for progress bar
for i in tqdm(range(0, len(lines), BATCH_SIZE)):
translated_batch = process_batch(lines[i:i+BATCH_SIZE])
translated_lines.extend(translated_batch)
# Write the translated text to another file
with open(outfile, 'w', encoding='latin1') as file:
file.writelines(translated_lines)
except FileNotFoundError:
print("The input file was not found.")
def main():
parser = argparse.ArgumentParser(description="Translate text file content to German.")
parser.add_argument("-i", "--input_file", required=True, help="Path to the input text file")
parser.add_argument("-o", "--output_file", required=True, help="Path to the output text file where translated text will be saved")
args = parser.parse_args()
process_file(args.input_file, args.output_file)
if __name__ == "__main__":
main()