Upload train_qwen_codeforces.py with huggingface_hub

train_qwen_codeforces.py

@@ -0,0 +1,75 @@
+# /// script
+# dependencies = ["trl>=0.12.0", "peft>=0.7.0", "trackio>=0.1.0", "datasets>=2.0.0"]
+# ///
+
+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
+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
+peft_config = LoraConfig(
+    r=8,
+    lora_alpha=16,
+    lora_dropout=0.05,
+    bias="none",
+    task_type="CAUSAL_LM",
+    target_modules=["q_proj", "k_proj", "v_proj", "o_proj"]
+)
+
+# Configure trainer - optimized for T4 small
+trainer = SFTTrainer(
+    model="Qwen/Qwen2.5-0.5B",
+    train_dataset=dataset,
+    # No eval dataset to save ~40% memory on T4 small
+    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_seq_length=512,
+    )
+)
+
+# Train the model
+print("🚀 Starting training on T4 small...")
+trainer.train()
+
+# Final push to hub
+print("📤 Pushing final model to Hub...")
+trainer.push_to_hub()
+
+print("✅ Training complete!")