fine-tuning.llama2.py

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, Trainer, TrainingArguments
from datasets import load_dataset, Dataset
from peft import LoraConfig, get_peft_model
from transformers import BitsAndBytesConfig, DataCollatorForLanguageModeling
import evaluate
import pandas as pd
import time

# Quantization configuration
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=torch.float16
)

# Load model and tokenizer
model_name = "meta-llama/Llama-2-7b-chat-hf"
tokenizer = AutoTokenizer.from_pretrained(model_name, use_auth_token=True)

# Set the padding token
tokenizer.pad_token = tokenizer.eos_token

model = AutoModelForCausalLM.from_pretrained(
    model_name, 
    quantization_config=bnb_config, 
    device_map="auto",
    use_auth_token=True
)

print(f"Model {model_name} loaded and quantized in 4 bits")

# Load dataset
dataset = load_dataset("andreshere/counsel_chat", use_auth_token=True)
print(f"Dataset loaded")

# Configure LoRA
lora_config = LoraConfig(
    r=32,
    lora_alpha=32,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
    lora_dropout=0.1
)

model = get_peft_model(model, lora_config)

# Adjust dataset without weights
def preprocess_function(examples):
    inputs = examples['Context']
    targets = examples['Response']
    model_inputs = tokenizer(inputs, max_length=512, truncation=True, padding='max_length')
    labels = tokenizer(targets, max_length=512, truncation=True, padding='max_length').input_ids
    
    model_inputs["labels"] = labels
    model_inputs["Context"] = inputs
    model_inputs["Response"] = targets
    return model_inputs

tokenized_dataset = dataset.map(preprocess_function, batched=True)

train_dataset = tokenized_dataset['train']
eval_dataset = tokenized_dataset['test']

data_collator = DataCollatorForLanguageModeling(
    tokenizer=tokenizer,
    mlm=False,
)

# Prepare training arguments
training_args = TrainingArguments(
    output_dir='./',
    do_train=True,
    do_eval=True,
    eval_strategy='epoch',  # Update to use 'eval_strategy' instead of 'evaluation_strategy'
    gradient_accumulation_steps=2,
    auto_find_batch_size=True,
    weight_decay=0.01,
    num_train_epochs=4,
    learning_rate=1e-4,
    logging_dir='./logs',
    logging_strategy="steps",
    logging_steps=10,
    save_strategy="epoch",
    save_total_limit=2,
    save_safetensors=True,
    eval_steps=10,
    report_to="tensorboard",
    hub_strategy="every_save"
)

# Initialize trainer
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    data_collator=data_collator
)

# Train the model
trainer.train()

# Save the entire model
trainer.save_model()

# Push the entire model to the Hugging Face Hub
trainer.push_to_hub("andreshere/llama-2-7b-mental-health-counseler")

# Load evaluation metrics
rouge = evaluate.load('rouge')
bleu = evaluate.load('bleu')
bertscore = evaluate.load('bertscore')
meteor = evaluate.load('meteor')

# Function to compute metrics and log compute time
def compute_metrics_and_log_time(pred):
    labels_ids = pred.label_ids
    pred_ids = pred.predictions
    decoded_preds = tokenizer.batch_decode(pred_ids, skip_special_tokens=True)
    decoded_labels = tokenizer.batch_decode(labels_ids, skip_special_tokens=True)
    
    # Compute metrics
    rouge_result = rouge.compute(predictions=decoded_preds, references=decoded_labels)
    bleu_result = bleu.compute(predictions=decoded_preds, references=decoded_labels)
    bertscore_result = bertscore.compute(predictions=decoded_preds, references=decoded_labels)
    meteor_result = meteor.compute(predictions=decoded_preds, references=decoded_labels)
    
    result = {
        "rouge": rouge_result,
        "bleu": bleu_result,
        "bertscore": bertscore_result,
        "meteor": meteor_result,
    }
    
    # Log compute time for each response
    compute_times = []
    contexts = []
    original_responses = []
    model_responses = []

    for i, context in enumerate(pred.features['Context']):
        start_time = time.time()
        model_response = decoded_preds[i]
        end_time = time.time()
        compute_time = end_time - start_time

        compute_times.append(compute_time)
        contexts.append(context)
        original_responses.append(decoded_labels[i])
        model_responses.append(model_response)
    
    # Save the log to a DataFrame
    log_df = pd.DataFrame({
        "Context": contexts,
        "Original Response": original_responses,
        "Model Response": model_responses,
        "Compute Time": compute_times
    })
    
    log_df.to_csv("response_log.csv", index=False)

    return result

# Evaluate model
metrics = trainer.evaluate(eval_dataset=eval_dataset, metric_key_prefix="eval", compute_metrics=compute_metrics_and_log_time)
print(metrics)

# Export metrics to CSV
metrics_df = pd.DataFrame(metrics, index=[0])
metrics_df.to_csv("fine-tuned-metrics.csv", index=False)