scripts/train_dolphin_phi.py

import torch, os
from datasets import load_dataset
from transformers import EarlyStoppingCallback
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig, TrainingArguments
from peft import LoraConfig, get_peft_model
from trl import SFTTrainer, SFTConfig, setup_chat_format
import torch

print("Is a CUDA GPU available? ", torch.cuda.is_available())
print("The CUDA version is: ", torch.version.cuda)

NAME_OF_MODEL = "microsoft/phi-2"
DATASET_PATH = "data/data_set1.jsonl"
OUTPUT_DIR = "/model_output/dolphi_round_1"

os.makedirs(OUTPUT_DIR, exist_ok=True)

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


lora_config = LoraConfig(
    r=32,
    lora_alpha=64,
    bias='none',
    target_modules=["q_proj", "k_proj", "v_proj"],
    lora_dropout=0.15,
    task_type="CAUSAL_LM"
)

try:
    # Load dataset with your 'prompt' and 'response' keys
    dataset = load_dataset("json", data_files=DATASET_PATH)
    split_dataset = dataset["train"].train_test_split(test_size=0.1, seed=42)
    train_dataset = split_dataset["train"]
    eval_dataset = split_dataset["test"]
    print("Dataset loaded and split successfully!")

    train_dataset = train_dataset.rename_column("response", "completion")
    eval_dataset = eval_dataset.rename_column("response", "completion")
    print("Renamed 'response' column to 'completion' in datasets.")
except Exception as e:
    print(f"Error loading dataset from {DATASET_PATH}: {e}")
    exit(1)

def formatting_func(example):
    text = f"### System Prompt:\nSummarize the following log entry in the specified format.\n\n### Log Entry:\n{example['prompt']}\n\n### Summary:\n{example['completion']}"
    return text


try:
    # Use setup_chat_format to automatically configure the tokenizer and model.
    # This prevents manual syntax errors and resizes the embedding layer.
    model = AutoModelForCausalLM.from_pretrained(
        NAME_OF_MODEL,
        quantization_config=bnb_config,
        device_map="auto",
        trust_remote_code=True,
        torch_dtype=torch.float16,
        attn_implementation="eager"
    )
    tokenizer = AutoTokenizer.from_pretrained(NAME_OF_MODEL, trust_remote_code=True)
    model, tokenizer = setup_chat_format(
        model,
        tokenizer,
        resize_to_multiple_of=8
    )

    # Note: When passing the model object directly to SFTTrainer,
    # the model_init_kwargs in SFTConfig are ignored.
    # The setup_chat_format function also correctly sets the chat template,
    # making the manual definition unnecessary.
    print("Model and Tokenizer loaded and configured successfully!")

except Exception as e:
    print(f'ERROR LOADING MODEL OR TOKENIZER: {e}')
    exit(1)



sft_config = SFTConfig(
    output_dir=OUTPUT_DIR,
    per_device_train_batch_size=4,
    gradient_accumulation_steps=16,
    learning_rate=1e-4,
    weight_decay=0.001,
    bf16=True,
    warmup_ratio=0.03,
    group_by_length=True,
    lr_scheduler_type='cosine',
    num_train_epochs=2,
    logging_steps=10,
    save_steps=25,
    fp16=False,
    optim="paged_adamw_8bit",
    report_to=["tensorboard"],
    eval_strategy="steps",
    eval_steps=25,
    packing=False,
    completion_only_loss=False,
    max_length=2048,
    load_best_model_at_end=True,
    metric_for_best_model="eval_loss",
    greater_is_better=False
)

trainer=SFTTrainer(
    model=model,
    processing_class=tokenizer,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    peft_config=lora_config,
    args=sft_config,
    formatting_func=formatting_func,
    callbacks=[EarlyStoppingCallback(early_stopping_patience=7)]
)

print("training started")

trainer.train()

print("fine tuning complete")

trainer.save_model(OUTPUT_DIR, merge_adapter_layers=True)