Upload training/evez-colab-headless.py with huggingface_hub

training/evez-colab-headless.py

@@ -0,0 +1,97 @@
+#!/usr/bin/env python3
+"""EVEZ Colab Training - Headless Runner
+Usage: Run on any machine with GPU access (Colab, Kaggle, local)
+1. Upload this script + training data
+2. python3 evez-colab-headless.py
+3. Adapter weights saved to ./evez-adapter/
+"""
+import json
+import os
+import sys
+
+def main():
+    print("🧬 EVEZ Self-Training Pipeline - Headless Mode")
+    
+    # Install deps
+    os.system("pip install -q transformers datasets peft trl torch accelerate")
+    
+    import torch
+    if not torch.cuda.is_available():
+        print("❌ No GPU detected. This needs CUDA.")
+        sys.exit(1)
+    print(f"✅ GPU: {torch.cuda.get_device_name(0)}")
+    
+    from transformers import AutoModelForCausalLM, AutoTokenizer
+    from peft import LoraConfig, get_peft_model
+    from trl import SFTTrainer, SFTConfig
+    from datasets import Dataset
+    
+    # Load training data
+    data_url = "https://raw.githubusercontent.com/EVEZX/neuros/main/training/evez-alpaca.json"
+    try:
+        import urllib.request
+        urllib.request.urlretrieve(data_url, "evez-alpaca.json")
+        with open("evez-alpaca.json") as f:
+            data = json.load(f)
+        print(f"✅ Loaded {len(data)} instruction pairs")
+    except Exception as e:
+        print(f"❌ Failed to load data: {e}")
+        sys.exit(1)
+    
+    # Format dataset
+    formatted = []
+    for d in data:
+        text = f"### Instruction:\n{d['instruction']}\n### Input:\n{d.get('input','')}\n### Response:\n{d['output']}"
+        formatted.append({"text": text})
+    dataset = Dataset.from_list(formatted)
+    
+    # Load model
+    model_name = "HuggingFaceTB/SmolLM2-135M"  # Tiny model for free tier
+    tokenizer = AutoTokenizer.from_pretrained(model_name)
+    tokenizer.pad_token = tokenizer.eos_token
+    model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16, device_map="auto")
+    
+    # LoRA config
+    lora_config = LoraConfig(
+        r=8, lora_alpha=16, lora_dropout=0.05,
+        target_modules=["q_proj", "v_proj"],
+        task_type="CAUSAL_LM"
+    )
+    model = get_peft_model(model, lora_config)
+    model.print_trainable_parameters()
+    
+    # Train
+    training_args = SFTConfig(
+        output_dir="./evez-adapter",
+        num_train_epochs=3,
+        per_device_train_batch_size=4,
+        learning_rate=2e-4,
+        logging_steps=10,
+        save_steps=100,
+        max_seq_length=512,
+        dataset_text_field="text",
+    )
+    
+    trainer = SFTTrainer(
+        model=model,
+        args=training_args,
+        train_dataset=dataset,
+        processing_class=tokenizer,
+    )
+    
+    print("🚀 Starting training...")
+    trainer.train()
+    
+    # Save adapter
+    model.save_pretrained("./evez-adapter")
+    tokenizer.save_pretrained("./evez-adapter")
+    print("✅ Adapter saved to ./evez-adapter/")
+    
+    # Optional: push to HF Hub
+    if os.environ.get("PUSH_TO_HUB"):
+        model.push_to_hub("evez420/EVEZ", token=os.environ.get("HF_TOKEN"))
+        tokenizer.push_to_hub("evez420/EVEZ", token=os.environ.get("HF_TOKEN"))
+        print("✅ Pushed to HF Hub: evez420/EVEZ")
+
+if __name__ == "__main__":
+    main()