Upload train_sft_qwen.py with huggingface_hub

train_sft_qwen.py

@@ -1,101 +1,110 @@
+#!/usr/bin/env python3
 # /// script
 # dependencies = [
-#   "trl>=0.12.0",
-#   "peft>=0.7.0",
-#   "trackio",
-#   "transformers>=4.40.0",
-#   "datasets>=2.18.0",
-#   "torch>=2.0.0",
+#     "trl>=0.12.0",
+#     "peft>=0.7.0",
+#     "transformers>=4.36.0",
+#     "accelerate>=0.24.0",
+#     "trackio",
 # ]
 # ///
 
 """
-SFT (Supervised Fine-Tuning) training script for Qwen/Qwen2.5-0.5B
-Uses TRL with LoRA, Trackio monitoring, and automatic Hub push
+SFT training script for Qwen/Qwen2.5-0.5B model.
+
+This script demonstrates:
+- Trackio integration for real-time monitoring
+- LoRA/PEFT for efficient training
+- Proper Hub saving configuration
+- Train/eval split for monitoring progress
+- Optimized training parameters for small model testing
 """
 
+import trackio
 from datasets import load_dataset
 from peft import LoraConfig
 from trl import SFTTrainer, SFTConfig
-import trackio
 
-# Load a high-quality instruction dataset
+# Load dataset
+print("📦 Loading dataset...")
 dataset = load_dataset("trl-lib/Capybara", split="train")
+print(f"✅ Dataset loaded: {len(dataset)} examples")
 
-# Create train/eval split for monitoring training progress
+# Create train/eval split for monitoring
+print("🔀 Creating train/eval split...")
 dataset_split = dataset.train_test_split(test_size=0.1, seed=42)
-
-# Configure LoRA for efficient fine-tuning
-peft_config = LoraConfig(
-    r=16,                      # LoRA rank
-    lora_alpha=32,             # LoRA alpha scaling
-    lora_dropout=0.05,         # Dropout for regularization
-    bias="none",               # Don't train bias parameters
-    task_type="CAUSAL_LM",     # Causal language modeling
-    target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],  # Qwen attention layers
-)
-
-# Configure trainer
-training_config = SFTConfig(
-    # Model and output
-    output_dir="qwen-sft-capybara",
-
-    # Hub configuration - CRITICAL for saving results
+train_dataset = dataset_split["train"]
+eval_dataset = dataset_split["test"]
+print(f"   Train: {len(train_dataset)} examples")
+print(f"   Eval: {len(eval_dataset)} examples")
+
+# Training configuration
+print("⚙️  Configuring training parameters...")
+config = SFTConfig(
+    # CRITICAL: Hub settings - Save model to Hugging Face Hub
+    output_dir="qwen-0.5b-sft-capybara",
     push_to_hub=True,
-    hub_model_id="qwen-sft-capybara-demo",  # Will use format: username/qwen-sft-capybara-demo
-    hub_strategy="every_save",              # Push checkpoints during training
-    hub_private_repo=False,
+    hub_model_id="vgtomahawk/qwen-0.5b-sft-capybara",
+    hub_strategy="every_save",  # Push checkpoints to Hub
 
     # Training parameters
     num_train_epochs=3,
-    per_device_train_batch_size=2,
-    gradient_accumulation_steps=4,  # Effective batch size: 2 * 4 = 8
+    per_device_train_batch_size=4,
+    gradient_accumulation_steps=4,  # Effective batch size = 4 * 4 = 16
+    learning_rate=2e-5,
 
-    # Optimization
-    learning_rate=2e-4,
-    lr_scheduler_type="cosine",
-    warmup_ratio=0.1,
+    # Logging & checkpointing
+    logging_steps=10,
+    save_strategy="steps",
+    save_steps=100,
+    save_total_limit=2,  # Keep only last 2 checkpoints
 
     # Evaluation
     eval_strategy="steps",
-    eval_steps=50,
-    per_device_eval_batch_size=2,
+    eval_steps=100,
 
-    # Checkpointing
-    save_strategy="steps",
-    save_steps=100,
-    save_total_limit=3,  # Keep only last 3 checkpoints
+    # Optimization
+    warmup_ratio=0.1,
+    lr_scheduler_type="cosine",
 
-    # Logging - Trackio integration
-    logging_steps=10,
+    # Monitoring with Trackio
     report_to="trackio",
-    run_name="qwen-0.5b-sft-demo",
-
-    # Performance optimization
-    bf16=True,  # Use bfloat16 for better performance on modern GPUs
-    gradient_checkpointing=True,  # Reduce memory usage
+    project="qwen-sft-demo",
+    run_name="qwen-0.5b-baseline",
+)
 
-    # Misc
-    seed=42,
-    dataloader_num_workers=4,
+# LoRA configuration for efficient training
+print("🔧 Setting up LoRA configuration...")
+peft_config = LoraConfig(
+    r=16,
+    lora_alpha=32,
+    lora_dropout=0.05,
+    bias="none",
+    task_type="CAUSAL_LM",
+    target_modules=["q_proj", "v_proj"],
 )
 
 # Initialize trainer
+print("🎯 Initializing SFT trainer...")
 trainer = SFTTrainer(
     model="Qwen/Qwen2.5-0.5B",
-    train_dataset=dataset_split["train"],
-    eval_dataset=dataset_split["test"],
+    train_dataset=train_dataset,
+    eval_dataset=eval_dataset,
+    args=config,
     peft_config=peft_config,
-    args=training_config,
 )
 
-# Train the model
-print("Starting training...")
+# Start training
+print("🚀 Starting training...")
+print("=" * 60)
 trainer.train()
 
-# Final push to Hub
-print("Training complete! Pushing final model to Hub...")
+# Push final model to Hub
+print("=" * 60)
+print("💾 Pushing final model to Hub...")
 trainer.push_to_hub()
 
-print("✅ Training complete and model saved to Hub!")
-print(f"Model available at: https://huggingface.co/{trainer.hub_model_id}")
+# Complete
+print("✅ Training complete!")
+print(f"📊 Model available at: https://huggingface.co/vgtomahawk/qwen-0.5b-sft-capybara")
+print(f"📈 View training metrics at: https://huggingface.co/spaces/vgtomahawk/trackio")