Upload train_gemma_sportingbot.py with huggingface_hub

train_gemma_sportingbot.py

@@ -0,0 +1,88 @@
+# /// script
+# dependencies = ["trl>=0.12.0", "peft>=0.7.0", "trackio", "transformers>=4.36.0", "datasets>=2.16.0"]
+# ///
+
+from datasets import load_dataset
+from peft import LoraConfig
+from trl import SFTTrainer, SFTConfig
+import trackio
+
+# Load dataset
+dataset = load_dataset("machina-sports/sportingbot-classification", split="train")
+
+# Create train/eval split (10% eval)
+dataset_split = dataset.train_test_split(test_size=0.1, seed=42)
+
+print(f"✅ Dataset loaded: {len(dataset_split['train'])} train, {len(dataset_split['test'])} eval")
+
+# Configure LoRA
+peft_config = LoraConfig(
+    r=32,
+    lora_alpha=64,
+    target_modules="all-linear",
+    lora_dropout=0.05,
+    bias="none",
+    task_type="CAUSAL_LM"
+)
+
+# Configure training
+training_args = SFTConfig(
+    output_dir="sportingbot-gemma-classifier",
+
+    # Hub settings (CRITICAL - saves results)
+    push_to_hub=True,
+    hub_model_id="fernando-machina/sportingbot-gemma-classifier",
+    hub_strategy="every_save",
+    hub_private_repo=False,
+
+    # Training hyperparameters (from user's config)
+    num_train_epochs=5,
+    per_device_train_batch_size=2,
+    gradient_accumulation_steps=4,
+    learning_rate=0.0001,
+
+    # Optimization (bf16 for Gemma)
+    bf16=True,
+    gradient_checkpointing=True,
+
+    # Evaluation
+    eval_strategy="steps",
+    eval_steps=10,
+
+    # Checkpointing
+    save_strategy="steps",
+    save_steps=50,
+    save_total_limit=3,
+
+    # Logging
+    logging_steps=5,
+    report_to="trackio",
+
+    # Trackio monitoring
+    project="sportingbot-classification",
+    run_name="gemma-2-2b-it-v1",
+
+    # Sequence length
+    max_seq_length=512,
+)
+
+print("🚀 Starting training with Gemma 2-2B-it...")
+
+# Create trainer
+trainer = SFTTrainer(
+    model="google/gemma-2-2b-it",
+    train_dataset=dataset_split["train"],
+    eval_dataset=dataset_split["test"],
+    peft_config=peft_config,
+    args=training_args,
+)
+
+# Train
+trainer.train()
+
+print("✅ Training complete! Pushing to Hub...")
+
+# Push final model
+trainer.push_to_hub()
+
+print(f"🎉 Model saved to: https://huggingface.co/fernando-machina/sportingbot-gemma-classifier")