train_sft_qwen25.py

# /// script
# dependencies = [
#     "trl>=0.12.0",
#     "peft>=0.7.0",
#     "transformers>=4.44.0",
#     "accelerate>=0.34.0",
#     "datasets>=2.19.0",
#     "trackio",
# ]
# ///
"""
SFT demo for Qwen/Qwen2.5-0.5B on the Capybara dataset with LoRA and Trackio.
Designed for Hugging Face Jobs (uv) with Hub push enabled.
"""
import os
import random
from datasets import load_dataset
from transformers import AutoTokenizer
from peft import LoraConfig
from trl import SFTConfig, SFTTrainer
import trackio

MODEL_ID = os.environ.get("MODEL_ID", "Qwen/Qwen2.5-0.5B")
DATASET_ID = os.environ.get("DATASET_ID", "trl-lib/Capybara")
HUB_MODEL_ID = os.environ.get("HUB_MODEL_ID", "davidsmts/qwen2_5-0.5b-capybara-sft")
RUN_NAME = os.environ.get("RUN_NAME", "qwen25-0.5b-capybara-demo")
PROJECT = os.environ.get("TRACKIO_PROJECT", "qwen-sft-demo")
SPACE_ID = os.environ.get("TRACKIO_SPACE", "davidsmts/trackio")
MAX_TRAIN_SAMPLES = int(os.environ.get("MAX_TRAIN_SAMPLES", "200"))
SEED = int(os.environ.get("SEED", "42"))

random.seed(SEED)

print("Loading dataset...")
dataset = load_dataset(DATASET_ID, split="train")
print(f"Loaded {len(dataset)} examples")

if MAX_TRAIN_SAMPLES and len(dataset) > MAX_TRAIN_SAMPLES:
    dataset = dataset.shuffle(seed=SEED).select(range(MAX_TRAIN_SAMPLES))
    print(f"Subsampled to {len(dataset)} examples for quick demo")

print("Creating train/test split...")
split = dataset.train_test_split(test_size=0.1, seed=SEED)
train_dataset = split["train"]
eval_dataset = split["test"]

print("Loading tokenizer...")
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, use_fast=False)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

def formatting_func(example):
    # Capybara provides conversation-style "messages"; apply chat template.
    return tokenizer.apply_chat_template(example["messages"], tokenize=False, add_generation_prompt=False)

print("Initializing Trackio...")
trackio.init(
    project=PROJECT,
    name=RUN_NAME,
    space_id=SPACE_ID,
    config={
        "model": MODEL_ID,
        "dataset": DATASET_ID,
        "lr": 2e-5,
        "epochs": 1,
        "max_train_samples": MAX_TRAIN_SAMPLES,
    },
)

print("Building LoRA config...")
peft_config = LoraConfig(
    r=16,
    lora_alpha=32,
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM",
    target_modules=["q_proj", "v_proj"],
)

print("Setting trainer args...")
training_args = SFTConfig(
    output_dir="./outputs",
    push_to_hub=True,
    hub_model_id=HUB_MODEL_ID,
    hub_strategy="every_save",
    num_train_epochs=1,
    per_device_train_batch_size=2,
    gradient_accumulation_steps=8,
    learning_rate=2e-5,
    max_length=1024,
    logging_steps=5,
    save_strategy="steps",
    save_steps=50,
    save_total_limit=2,
    eval_strategy="steps",
    eval_steps=50,
    warmup_ratio=0.03,
    lr_scheduler_type="cosine",
    gradient_checkpointing=True,
    fp16=True,
    report_to="trackio",
    project=PROJECT,
    run_name=RUN_NAME,
)

print("Initializing trainer...")
trainer = SFTTrainer(
    model=MODEL_ID,
    tokenizer=tokenizer,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    formatting_func=formatting_func,
    peft_config=peft_config,
    args=training_args,
)

print("Starting training...")
trainer.train()

print("Saving and pushing to hub...")
trainer.push_to_hub()
trackio.finish()

print(f"Done! Model pushed to https://huggingface.co/{HUB_MODEL_ID}")
print(f"Track metrics at https://huggingface.co/spaces/{SPACE_ID}")