train_without_trackio.py

# /// script
# requires-python = ">=3.10"
# dependencies = ["trl", "datasets"]
# ///
"""Simple SFT training WITHOUT trackio - logs should flow normally."""

import sys
sys.stdout.reconfigure(line_buffering=True)

from datasets import load_dataset
from trl import SFTTrainer, SFTConfig

print("=" * 50)
print("Training WITHOUT Trackio")
print("=" * 50)

dataset = load_dataset("trl-lib/Capybara", split="train[:1000]")
print(f"Dataset loaded: {len(dataset)} samples")

config = SFTConfig(
    output_dir="./output",
    max_steps=50,
    logging_steps=10,
    report_to="none",
)

trainer = SFTTrainer(
    model="Qwen/Qwen2.5-0.5B",
    train_dataset=dataset,
    args=config,
)

print("Starting training...")
trainer.train()
print("Training complete!")
print("=" * 50)