evalstate/demo-training-scripts

1# /// script2# dependencies = [3#     "trl>=0.12.0",4#     "transformers>=4.36.0",5#     "datasets>=2.14.0",6#     "peft>=0.7.0",7#     "accelerate>=0.24.0",8# ]9# ///10 11from datasets import load_dataset12from peft import LoraConfig13from trl import SFTTrainer, SFTConfig14 15# Load just 50 examples for quick demo16dataset = load_dataset("trl-lib/Capybara", split="train[:50]")17print(f"✅ Dataset loaded: {len(dataset)} examples")18 19# Training configuration20config = SFTConfig(21    # CRITICAL: Hub settings22    output_dir="qwen-demo-sft",23    push_to_hub=True,24    hub_model_id="evalstate/qwen-capybara-demo",25    26    # Quick training settings27    max_steps=20,  # Just 20 steps for demo28    per_device_train_batch_size=2,29    learning_rate=2e-5,30    31    # Logging32    logging_steps=5,33    save_strategy="steps",34    save_steps=10,35    36    # Optimization37    warmup_ratio=0.1,38)39 40# LoRA configuration for efficient training41peft_config = LoraConfig(42    r=16,43    lora_alpha=32,44    lora_dropout=0.05,45    bias="none",46    task_type="CAUSAL_LM",47    target_modules=["q_proj", "v_proj"],48)49 50# Initialize and train51trainer = SFTTrainer(52    model="Qwen/Qwen2.5-0.5B",53    train_dataset=dataset,54    args=config,55    peft_config=peft_config,56)57 58print("🚀 Starting training...")59trainer.train()60 61print("💾 Pushing to Hub...")62trainer.push_to_hub()63 64print("✅ Complete! Model at: https://huggingface.co/evalstate/qwen-capybara-demo")65