train.py

# /// script
# dependencies = ["trl>=0.12.0", "peft>=0.7.0", "trackio", "datasets", "transformers", "accelerate", "jinja2"]
# ///

from datasets import load_dataset
from peft import LoraConfig
from trl import SFTTrainer, SFTConfig
from transformers import AutoTokenizer
import trackio
import os
import json

print("🚀 Starting FunctionGemma 270M Fine-tuning (V6 - Optimized with Sample Best Practices)")

model_id = "google/functiongemma-270m-it"
tokenizer = AutoTokenizer.from_pretrained(model_id)

# Ensure pad token is set
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

# Load dataset
dataset = load_dataset("epinfomax/vn-function-calling-dataset", split="train")

def format_conversation(example):
    # As per the sample notebook: separate prompt and completion for completion_only_loss
    # but TRL SFTTrainer can also handle a single 'text' field with completion_only_loss=True
    # by using a specific collator if needed. 
    # Here we will follow the sample's way of defining prompt and completion columns.
    
    full_text = tokenizer.apply_chat_template(
        example["messages"],
        tools=example["tools"],
        tokenize=False,
        add_generation_prompt=False
    )
    
    prompt_text = tokenizer.apply_chat_template(
        example["messages"][:-1], # Everything except the last assistant message
        tools=example["tools"],
        tokenize=False,
        add_generation_prompt=True # Include 'model' header
    )
    
    completion_text = full_text[len(prompt_text):]
    
    return {
        "prompt": prompt_text,
        "completion": completion_text
    }

print("🔄 Pre-processing dataset with prompt/completion split...")
dataset = dataset.map(format_conversation, remove_columns=dataset.column_names)

# LoRA configuration
peft_config = LoraConfig(
    r=16,
    lora_alpha=32,
    target_modules=["q_proj", "v_proj", "k_proj", "o_proj"],
    task_type="CAUSAL_LM",
)

# Training configuration (Optimized with Sample Best Practices)
config = SFTConfig(
    output_dir="vn-function-gemma-270m-finetuned",
    max_length=1024,
    push_to_hub=True,
    hub_model_id="epinfomax/vn-function-gemma-270m-finetuned",
    hub_strategy="every_save",
    num_train_epochs=5,
    per_device_train_batch_size=4,
    gradient_accumulation_steps=4,
    learning_rate=1e-5, # From sample: more conservative
    lr_scheduler_type="cosine", # From sample
    optim="adamw_torch_fused", # From sample
    logging_steps=5,
    save_strategy="steps",
    save_steps=50,
    report_to="trackio",
    project="vn-function-calling",
    run_name="function-gemma-270m-v6-optimized",
    completion_only_loss=True, # Focus on assistant responses
    packing=False
)

# Initialize and train
print("🎯 Initializing SFTTrainer with optimized configuration...")
trainer = SFTTrainer(
    model=model_id,
    train_dataset=dataset,
    peft_config=peft_config,
    args=config,
)

trainer.train()
trainer.push_to_hub()
print("✅ Training complete and pushed to Hub!")