Spaces:
Sleeping
Sleeping
File size: 3,076 Bytes
fdd2607 ea92c48 f0d69bf ea92c48 f0d69bf c79c478 f0d69bf c79c478 f0d69bf ea92c48 f0d69bf ea92c48 f0d69bf ea92c48 f0d69bf ea92c48 f0d69bf ea92c48 f0d69bf ea92c48 f0d69bf ea92c48 f0d69bf |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 |
import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer, Trainer, TrainingArguments
from datasets import load_dataset, Dataset
from peft import LoraConfig, get_peft_model
import torch
# Load GPT-2 model and tokenizer
model_name = "gpt2"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)
# Custom Dataset (Improved format)
custom_data = [
{"text": "Prompt: Who are you?\nResponse: I am Eva, a virtual voice assistant."},
{"text": "Prompt: What is your name?\nResponse: I am Eva, how can I help you?"},
{"text": "Prompt: What can you do?\nResponse: I can assist with answering questions, searching the web, and much more!"},
{"text": "Prompt: Who invented the computer?\nResponse: Charles Babbage is known as the father of the computer."},
{"text": "Prompt: Tell me a joke.\nResponse: Why don’t scientists trust atoms? Because they make up everything!"},
{"text": "Prompt: Who is the Prime Minister of India?\nResponse: The current Prime Minister of India is Narendra Modi."},
{"text": "Prompt: Who created you?\nResponse: I was created by an expert team specializing in AI fine-tuning and web development."},
{"text": "Prompt: Can you introduce yourself?\nResponse: I am Eva, your AI assistant, designed to assist and provide information."}
]
# Convert custom data to a Dataset
dataset_custom = Dataset.from_list(custom_data)
def tokenize_function(examples):
return tokenizer(examples["text"], truncation=True, padding="max_length", max_length=512)
tokenized_datasets = dataset_custom.map(tokenize_function, batched=True)
# Apply LoRA for efficient fine-tuning
lora_config = LoraConfig(
r=4, # Reduced r for stability
lora_alpha=16,
lora_dropout=0.1,
bias="none",
target_modules=["c_attn", "c_proj"] # LoRA targets attention layers
)
model = get_peft_model(model, lora_config)
model.gradient_checkpointing_enable()
# Training arguments
training_args = TrainingArguments(
output_dir="gpt2_finetuned",
auto_find_batch_size=True,
gradient_accumulation_steps=4,
learning_rate=3e-5, # Lowered learning rate for improved stability
num_train_epochs=5, # Increased epochs for better training
save_strategy="epoch",
logging_dir="logs",
bf16=True,
push_to_hub=True
)
# Trainer setup
trainer = Trainer(
model=model,
args=training_args,
train_dataset=tokenized_datasets
)
trainer.train()
# Save and push the model
trainer.save_model("gpt2_finetuned")
tokenizer.save_pretrained("gpt2_finetuned")
trainer.push_to_hub()
# Gradio Interface for Responses
def generate_response(prompt):
inputs = tokenizer(f"Prompt: {prompt}\nResponse:", return_tensors="pt")
outputs = model.generate(**inputs, max_length=150, num_return_sequences=1, temperature=0.7, top_p=0.9)
return tokenizer.decode(outputs[0], skip_special_tokens=True).split("Response:")[-1].strip()
demo = gr.Interface(fn=generate_response, inputs="text", outputs="text")
if __name__ == "__main__":
demo.launch()
|