Upload train_qwen_codeforces.py with huggingface_hub

train_qwen_codeforces.py

@@ -6,21 +6,17 @@ from datasets import load_dataset
 from peft import LoraConfig
 from trl import SFTTrainer, SFTConfig
 import trackio
-import os
 
 # Load dataset - 1000 examples for ~20 min training
+print("📦 Loading dataset...")
 dataset = load_dataset(
     "open-r1/codeforces-cots",
     "solutions_w_editorials_py_decontaminated",
     split="train[:1000]"
 )
-
 print(f"📊 Training on {len(dataset)} examples for 3 epochs")
 
-# Get username for hub model id
-username = os.environ.get("HF_USERNAME", "papebaba")
-
-# Configure LoRA for efficient training on T4 small
+# LoRA configuration for efficient training
 peft_config = LoraConfig(
     r=8,
     lora_alpha=16,
@@ -30,46 +26,57 @@ peft_config = LoraConfig(
     target_modules=["q_proj", "k_proj", "v_proj", "o_proj"]
 )
 
-# Configure trainer - optimized for T4 small
+# Training configuration - optimized for T4 small
+config = SFTConfig(
+    # Hub settings - CRITICAL for saving results
+    output_dir="qwen-codeforces-finetuned",
+    push_to_hub=True,
+    hub_model_id="papebaba/qwen-codeforces-finetuned",
+    hub_strategy="end",
+    hub_private_repo=False,
+
+    # Training parameters
+    num_train_epochs=3,
+    per_device_train_batch_size=1,
+    gradient_accumulation_steps=8,  # Effective batch size = 8
+    learning_rate=2e-4,
+    max_length=512,  # Shorter sequences for T4 small
+
+    # Checkpointing
+    logging_steps=10,
+    save_strategy="epoch",
+    save_total_limit=1,
+
+    # Optimization for T4 small
+    gradient_checkpointing=True,
+    bf16=True,
+    max_grad_norm=1.0,
+    warmup_ratio=0.1,
+    lr_scheduler_type="cosine",
+    optim="adamw_torch",
+
+    # Trackio monitoring
+    report_to="trackio",
+    run_name="qwen-codeforces-sft-1k",
+)
+
+# Initialize trainer
+print("🎯 Initializing trainer...")
 trainer = SFTTrainer(
     model="Qwen/Qwen2.5-0.5B",
     train_dataset=dataset,
-    # No eval dataset to save ~40% memory on T4 small
+    args=config,
     peft_config=peft_config,
-    args=SFTConfig(
-        output_dir="qwen-codeforces-finetuned",
-        num_train_epochs=3,
-        per_device_train_batch_size=1,
-        gradient_accumulation_steps=8,  # Effective batch size = 8
-        gradient_checkpointing=True,
-        learning_rate=2e-4,
-        lr_scheduler_type="cosine",
-        warmup_ratio=0.1,
-        logging_steps=10,
-        save_strategy="epoch",
-        save_total_limit=1,
-        # Hub configuration
-        push_to_hub=True,
-        hub_model_id=f"{username}/qwen-codeforces-finetuned",
-        hub_strategy="end",
-        hub_private_repo=False,
-        # Trackio monitoring
-        report_to="trackio",
-        run_name="qwen-codeforces-sft-1k",
-        # Optimization for T4 small
-        bf16=True,
-        max_grad_norm=1.0,
-        optim="adamw_torch",
-        max_length=512,
-    )
 )
 
-# Train the model
+# Train
 print("🚀 Starting training on T4 small...")
 trainer.train()
 
-# Final push to hub
+# Push to Hub
 print("📤 Pushing final model to Hub...")
 trainer.push_to_hub()
 
 print("✅ Training complete!")
+print("📊 View metrics at: https://huggingface.co/spaces/papebaba/trackio")
+print("🤗 Model at: https://huggingface.co/papebaba/qwen-codeforces-finetuned")