app.py

"""
T5-Gemma-2 Fine-Tuning Space for Custom AI Specialists
========================================================
This Space fine-tunes google/t5gemma-2-270m-270m on custom training data.
Uses HuggingFace's free T4 GPU for training.

Specialist Models we can train:
1. UE5 Code Expert - Unreal Engine Python scripting
2. 3D Asset Expert - How to use 3D generation tools
3. Texture Expert - PBR materials, UV mapping guidance
4. Animation Expert - Rigging, animation workflows
"""

import gradio as gr
import json
import os
from pathlib import Path

# Training imports (loaded on demand to save memory)
def load_training_libs():
    global torch, AutoTokenizer, AutoModelForSeq2SeqLM, Trainer, TrainingArguments, Dataset
    import torch
    from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, Trainer, TrainingArguments
    from datasets import Dataset
    return True

# Constants
MODEL_ID = "google/t5gemma-2-270m-270m"
OUTPUT_DIR = "./fine-tuned-model"

# Sample training data for demo
DEMO_DATA = [
    {
        "input": "How do I get the current level name in UE5 Python?",
        "output": """import unreal
world = unreal.EditorLevelLibrary.get_editor_world()
level_name = world.get_name()
print(f"Current level: {level_name}")"""
    },
    {
        "input": "How do I spawn an actor at a location in UE5?",
        "output": """import unreal
# Define spawn location
location = unreal.Vector(0, 0, 100)
rotation = unreal.Rotator(0, 0, 0)

# Get the actor class (e.g., StaticMeshActor)
actor_class = unreal.StaticMeshActor

# Spawn the actor
actor = unreal.EditorLevelLibrary.spawn_actor_from_class(
    actor_class, location, rotation
)"""
    }
]


def prepare_dataset(training_data_text: str) -> dict:
    """Parse JSONL training data and prepare for training."""
    lines = training_data_text.strip().split('\n')
    data = []
    
    for line in lines:
        if line.strip():
            try:
                item = json.loads(line)
                if 'input' in item and 'output' in item:
                    data.append(item)
            except json.JSONDecodeError:
                continue
    
    return {
        "count": len(data),
        "sample": data[0] if data else None,
        "data": data
    }


def validate_data(training_data: str) -> str:
    """Validate training data format."""
    result = prepare_dataset(training_data)
    
    if result["count"] == 0:
        return "❌ No valid training examples found. Ensure each line is valid JSON with 'input' and 'output' fields."
    
    return f"""✅ **Validation Successful!**

**Examples found:** {result["count"]}

**Sample Entry:**
```json
{json.dumps(result["sample"], indent=2)}
```

Ready to start training!"""


def start_training(
    training_data: str,
    model_name: str,
    num_epochs: int,
    batch_size: int,
    learning_rate: float,
    progress=gr.Progress()
) -> str:
    """Fine-tune T5-Gemma-2 on the provided training data."""
    
    progress(0, desc="Loading training libraries...")
    try:
        load_training_libs()
    except Exception as e:
        return f"❌ Failed to load training libraries: {e}"
    
    progress(0.1, desc="Parsing training data...")
    dataset_info = prepare_dataset(training_data)
    
    if dataset_info["count"] < 5:
        return "❌ Need at least 5 training examples to fine-tune."
    
    progress(0.2, desc=f"Loading {MODEL_ID}...")
    try:
        tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
        model = AutoModelForSeq2SeqLM.from_pretrained(MODEL_ID)
    except Exception as e:
        return f"❌ Failed to load model: {e}"
    
    progress(0.3, desc="Tokenizing dataset...")
    
    def tokenize_function(examples):
        # T5-style input format
        inputs = [f"Question: {q}" for q in examples["input"]]
        targets = examples["output"]
        
        model_inputs = tokenizer(
            inputs,
            max_length=512,
            truncation=True,
            padding="max_length"
        )
        
        labels = tokenizer(
            targets,
            max_length=512,
            truncation=True,
            padding="max_length"
        )
        
        model_inputs["labels"] = labels["input_ids"]
        return model_inputs
    
    # Create HF Dataset
    dataset = Dataset.from_dict({
        "input": [d["input"] for d in dataset_info["data"]],
        "output": [d["output"] for d in dataset_info["data"]]
    })
    
    tokenized_dataset = dataset.map(tokenize_function, batched=True)
    
    progress(0.4, desc="Setting up training...")
    
    training_args = TrainingArguments(
        output_dir=OUTPUT_DIR,
        num_train_epochs=num_epochs,
        per_device_train_batch_size=batch_size,
        learning_rate=learning_rate,
        weight_decay=0.01,
        logging_steps=10,
        save_strategy="epoch",
        fp16=True,  # Use mixed precision on T4
        push_to_hub=False,
        report_to="none",
    )
    
    trainer = Trainer(
        model=model,
        args=training_args,
        train_dataset=tokenized_dataset,
    )
    
    progress(0.5, desc="Training in progress...")
    
    try:
        trainer.train()
        progress(0.9, desc="Saving model...")
        
        # Save locally
        trainer.save_model(OUTPUT_DIR)
        tokenizer.save_pretrained(OUTPUT_DIR)
        
        progress(1.0, desc="Complete!")
        
        return f"""✅ **Training Complete!**

**Model:** {model_name}
**Base:** {MODEL_ID}
**Training Examples:** {dataset_info["count"]}
**Epochs:** {num_epochs}

Model saved to: `{OUTPUT_DIR}`

To push to HuggingFace Hub:
```python
from huggingface_hub import HfApi
api = HfApi()
api.upload_folder(folder_path="{OUTPUT_DIR}", repo_id="YOUR_USERNAME/{model_name}")
```
"""
    except Exception as e:
        return f"❌ Training failed: {e}"


def test_model(prompt: str) -> str:
    """Test the fine-tuned model."""
    model_path = Path(OUTPUT_DIR)
    
    if not model_path.exists():
        return "❌ No trained model found. Please train a model first."
    
    try:
        load_training_libs()
        tokenizer = AutoTokenizer.from_pretrained(OUTPUT_DIR)
        model = AutoModelForSeq2SeqLM.from_pretrained(OUTPUT_DIR)
        
        # Format input like training data
        input_text = f"Question: {prompt}"
        inputs = tokenizer(input_text, return_tensors="pt", max_length=512, truncation=True)
        
        outputs = model.generate(
            **inputs,
            max_length=512,
            num_beams=4,
            early_stopping=True
        )
        
        response = tokenizer.decode(outputs[0], skip_special_tokens=True)
        return response
    except Exception as e:
        return f"❌ Error: {e}"


def load_demo_data() -> str:
    """Load demo training data."""
    return "\n".join([json.dumps(d) for d in DEMO_DATA])


# Build the Gradio interface
with gr.Blocks(title="Custom AI Specialist Trainer") as demo:
    gr.Markdown("""
    # 🧠 Custom AI Specialist Trainer
    
    Fine-tune **T5-Gemma-2-270m** on your custom training data using HuggingFace's free T4 GPU!
    
    **Use Cases:**
    - 🎮 **UE5 Code Expert** - Unreal Engine Python scripting
    - 🏗️ **3D Asset Expert** - 3D generation tool guidance  
    - 🎨 **Texture Expert** - PBR materials, UV mapping
    - 🎬 **Animation Expert** - Rigging, motion workflows
    """)
    
    with gr.Tab("📝 Training Data"):
        gr.Markdown("""
        ### Training Data Format (JSONL)
        Each line should be a JSON object with `input` and `output` fields:
        ```json
        {"input": "How do I do X?", "output": "Here's how to do X..."}
        ```
        """)
        
        training_data = gr.Textbox(
            label="Training Data (JSONL)",
            placeholder='{"input": "How do I...", "output": "You can..."}',
            lines=15
        )
        
        with gr.Row():
            load_demo_btn = gr.Button("📥 Load Demo Data", variant="secondary")
            validate_btn = gr.Button("✅ Validate Data", variant="primary")
        
        validation_output = gr.Markdown()
        
        load_demo_btn.click(load_demo_data, outputs=training_data)
        validate_btn.click(validate_data, inputs=training_data, outputs=validation_output)
    
    with gr.Tab("🚀 Train Model"):
        gr.Markdown("### Training Configuration")
        
        with gr.Row():
            model_name = gr.Textbox(
                label="Model Name",
                value="ue5-code-expert",
                info="Name for your fine-tuned model"
            )
            num_epochs = gr.Slider(
                minimum=1, maximum=10, value=3, step=1,
                label="Training Epochs"
            )
        
        with gr.Row():
            batch_size = gr.Slider(
                minimum=1, maximum=8, value=4, step=1,
                label="Batch Size",
                info="Lower if out of memory"
            )
            learning_rate = gr.Number(
                value=5e-5,
                label="Learning Rate"
            )
        
        train_btn = gr.Button("🏋️ Start Training", variant="primary", size="lg")
        training_output = gr.Markdown()
        
        train_btn.click(
            start_training,
            inputs=[training_data, model_name, num_epochs, batch_size, learning_rate],
            outputs=training_output
        )
    
    with gr.Tab("🧪 Test Model"):
        gr.Markdown("### Test Your Fine-Tuned Model")
        
        test_prompt = gr.Textbox(
            label="Your Question",
            placeholder="How do I spawn an actor in UE5?",
            lines=3
        )
        
        test_btn = gr.Button("🔮 Generate Response", variant="primary")
        test_output = gr.Code(label="Model Response", language="python")
        
        test_btn.click(test_model, inputs=test_prompt, outputs=test_output)
    
    gr.Markdown("""
    ---
    ### 🔗 Related HF Spaces for Game Dev
    
    | Tool | Purpose | Link |
    |------|---------|------|
    | TRELLIS.2 | Image → 3D Mesh | [microsoft/TRELLIS.2](https://hf.co/spaces/microsoft/TRELLIS.2) |
    | Hunyuan3D | Text/Image → 3D | [tencent/Hunyuan3D-2.1](https://hf.co/spaces/tencent/Hunyuan3D-2.1) |
    | Stable Fast 3D | Quick 3D from Image | [stabilityai/stable-fast-3d](https://hf.co/spaces/stabilityai/stable-fast-3d) |
    | Chord PBR | PBR Material Gen | [ksangk/chord-demo](https://hf.co/spaces/ksangk/chord-demo) |
    | FLUX.2 | Image Generation | [black-forest-labs/FLUX.2-dev](https://hf.co/spaces/black-forest-labs/FLUX.2-dev) |
    """)

if __name__ == "__main__":
    demo.launch()