train_qwen3_codeforces_test.py

# /// script
# dependencies = ["trl>=0.12.0", "peft>=0.7.0", "trackio", "transformers>=4.44.0", "datasets"]
# ///

import sys
from datasets import load_dataset
from peft import LoraConfig
from trl import SFTTrainer, SFTConfig
import trackio

print("="*60)
print("🚀 STARTING TRAINING JOB - VERBOSE MODE")
print("="*60)

# Step 1: Load dataset
print("\n📥 Step 1/5: Loading dataset...")
try:
    dataset = load_dataset(
        "open-r1/codeforces-cots",
        name="solutions_w_editorials_decontaminated",
        split="train[:500]"  # Small subset for quick testing
    )
    print(f"✅ Dataset loaded: {len(dataset)} examples")
    print(f"   Columns: {dataset.column_names}")
    print(f"   First example keys: {list(dataset[0].keys())}")
except Exception as e:
    print(f"❌ FAILED to load dataset: {e}")
    sys.exit(1)

# Step 2: Create train/eval split
print("\n📊 Step 2/5: Creating train/eval split...")
try:
    dataset_split = dataset.train_test_split(test_size=0.1, seed=42)
    print(f"✅ Split created:")
    print(f"   Train: {len(dataset_split['train'])} examples")
    print(f"   Eval: {len(dataset_split['test'])} examples")
except Exception as e:
    print(f"❌ FAILED to create split: {e}")
    sys.exit(1)

# Step 3: Configure LoRA
print("\n🔧 Step 3/5: Configuring LoRA...")
try:
    peft_config = LoraConfig(
        r=16,
        lora_alpha=32,
        lora_dropout=0.05,
        target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
        task_type="CAUSAL_LM"
    )
    print(f"✅ LoRA configured: r={peft_config.r}, alpha={peft_config.lora_alpha}")
except Exception as e:
    print(f"❌ FAILED to configure LoRA: {e}")
    sys.exit(1)

# Step 4: Configure training
print("\n⚙️  Step 4/5: Configuring training...")
try:
    training_args = SFTConfig(
        output_dir="qwen3-0.6b-codeforces-test",
        
        # Quick training for testing
        num_train_epochs=1,  # Just 1 epoch for quick test
        per_device_train_batch_size=2,
        per_device_eval_batch_size=2,
        gradient_accumulation_steps=2,
        gradient_checkpointing=True,
        
        # Learning rate
        learning_rate=2e-4,
        lr_scheduler_type="cosine",
        warmup_ratio=0.1,
        optim="paged_adamw_8bit",
        
        # Frequent logging for visibility
        eval_strategy="steps",
        eval_steps=20,
        logging_steps=5,  # Log every 5 steps
        save_strategy="steps",
        save_steps=50,
        save_total_limit=2,
        
        # Hub integration
        push_to_hub=True,
        hub_model_id="kneeraj/qwen3-0.6b-codeforces-test",
        hub_strategy="every_save",
        hub_private_repo=False,
        
        # Trackio monitoring
        report_to="trackio",
        project="codeforces-finetuning-test",
        run_name="qwen3-quick-test",
        
        # Performance
        bf16=True,
        max_grad_norm=1.0,
        
        # Data processing
        max_seq_length=1024,  # Shorter for faster processing
        dataset_text_field="messages",
        packing=False,
    )
    print(f"✅ Training config created")
    print(f"   Epochs: {training_args.num_train_epochs}")
    print(f"   Batch size: {training_args.per_device_train_batch_size}")
    print(f"   Output: {training_args.hub_model_id}")
except Exception as e:
    print(f"❌ FAILED to configure training: {e}")
    sys.exit(1)

# Step 5: Initialize trainer and train
print("\n🏋️  Step 5/5: Initializing trainer and starting training...")
try:
    print("   Loading model: Qwen/Qwen2.5-0.5B-Instruct...")
    trainer = SFTTrainer(
        model="Qwen/Qwen2.5-0.5B-Instruct",
        train_dataset=dataset_split["train"],
        eval_dataset=dataset_split["test"],
        peft_config=peft_config,
        args=training_args,
    )
    print(f"✅ Trainer initialized")
    print(f"   Training samples: {len(dataset_split['train'])}")
    print(f"   Evaluation samples: {len(dataset_split['test'])}")
    
    print("\n" + "="*60)
    print("🎯 STARTING TRAINING...")
    print("="*60 + "\n")
    
    trainer.train()
    
    print("\n" + "="*60)
    print("💾 Pushing final model to Hub...")
    trainer.push_to_hub()
    
    print("\n" + "="*60)
    print("✅ TRAINING COMPLETE!")
    print("="*60)
    print(f"Model saved to: kneeraj/qwen3-0.6b-codeforces-test")
    print(f"View at: https://huggingface.co/kneeraj/qwen3-0.6b-codeforces-test")
    
except Exception as e:
    print(f"\n❌ TRAINING FAILED: {e}")
    import traceback
    traceback.print_exc()
    sys.exit(1)