Upload train_qwen_codeforces_v2.py with huggingface_hub

train_qwen_codeforces_v2.py

@@ -0,0 +1,114 @@
+# /// script
+# dependencies = [
+#     "trl>=0.12.0",
+#     "peft>=0.7.0",
+#     "transformers>=4.45.0",
+#     "accelerate>=0.24.0",
+#     "trackio",
+#     "datasets",
+# ]
+# ///
+
+"""
+Fine-tune Qwen3-0.6B on open-r1/codeforces-cots for competitive programming.
+"""
+
+import trackio
+from datasets import load_dataset
+from peft import LoraConfig
+from trl import SFTTrainer, SFTConfig
+from transformers import AutoTokenizer
+
+# Load dataset - using the solutions config with messages column
+print("Loading dataset...")
+dataset = load_dataset("open-r1/codeforces-cots", "solutions", split="train")
+print(f"Dataset loaded: {len(dataset)} examples")
+
+# The dataset has a 'messages' column in chat format
+# We need to keep only the 'messages' column for SFT training
+print("Preparing dataset - keeping only messages column...")
+dataset = dataset.select_columns(["messages"])
+
+# Create train/eval split
+print("Creating train/eval split...")
+dataset_split = dataset.train_test_split(test_size=0.05, seed=42)
+train_dataset = dataset_split["train"]
+eval_dataset = dataset_split["test"]
+print(f"Train: {len(train_dataset)} examples")
+print(f"Eval: {len(eval_dataset)} examples")
+
+# Load tokenizer for chat template
+print("Loading tokenizer...")
+tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B")
+if tokenizer.pad_token is None:
+    tokenizer.pad_token = tokenizer.eos_token
+
+# Processing function to convert messages to text using chat template
+def formatting_func(example):
+    return tokenizer.apply_chat_template(example["messages"], tokenize=False)
+
+# Training configuration
+config = SFTConfig(
+    # Hub settings - CRITICAL for saving results
+    output_dir="qwen3-codeforces-sft",
+    push_to_hub=True,
+    hub_model_id="kintopp/qwen3-0.6b-codeforces-cots",
+    hub_strategy="every_save",
+
+    # Training parameters
+    num_train_epochs=1,
+    per_device_train_batch_size=2,
+    gradient_accumulation_steps=8,
+    learning_rate=2e-5,
+    max_length=2048,
+
+    # Logging & checkpointing
+    logging_steps=25,
+    save_strategy="steps",
+    save_steps=500,
+    save_total_limit=2,
+
+    # Evaluation
+    eval_strategy="steps",
+    eval_steps=500,
+
+    # Optimization
+    warmup_ratio=0.1,
+    lr_scheduler_type="cosine",
+    gradient_checkpointing=True,
+    bf16=True,
+
+    # Monitoring with Trackio
+    report_to="trackio",
+    project="qwen3-codeforces",
+    run_name="qwen3-0.6b-codeforces-sft",
+)
+
+# LoRA configuration for efficient training
+peft_config = LoraConfig(
+    r=16,
+    lora_alpha=32,
+    lora_dropout=0.05,
+    bias="none",
+    task_type="CAUSAL_LM",
+    target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
+)
+
+# Initialize trainer
+print("Initializing trainer...")
+trainer = SFTTrainer(
+    model="Qwen/Qwen3-0.6B",
+    train_dataset=train_dataset,
+    eval_dataset=eval_dataset,
+    args=config,
+    peft_config=peft_config,
+    formatting_func=formatting_func,
+)
+
+print("Starting training...")
+trainer.train()
+
+print("Pushing final model to Hub...")
+trainer.push_to_hub()
+
+print("Complete! Model at: https://huggingface.co/kintopp/qwen3-0.6b-codeforces-cots")