Upload app.py

app.py

@@ -0,0 +1,210 @@
+#!/usr/bin/env python3
+import gradio as gr
+import os
+
+# Model options
+MODELS = [
+    # Small (fast, <3B)
+    'Qwen/Qwen2.5-0.5B-Instruct',
+    'Qwen/Qwen2.5-1.5B-Instruct', 
+    'microsoft/Phi-3-mini-128k-instruct',
+    'google/gemma-2b-it',
+    'meta-llama/Llama-3.2-1B-Instruct',
+    # Medium (7-13B)
+    'Qwen/Qwen2.5-7B-Instruct',
+    'meta-llama/Llama-3.2-3B-Instruct',
+    'mistralai/Mistral-7B-Instruct-v0.3',
+    # Large (needs more GPU)
+    'Qwen/Qwen2.5-14B-Instruct',
+    'mistralai/Mixtral-8x7B-Instruct-v0.1',
+]
+
+# Training methods
+METHODS = ['SFT', 'DPO', 'GRPO', 'LoRA']
+
+# Public datasets
+DATASETS = [
+    'HuggingFaceH4/ultrachat_200k',
+    'openai/gsm8k',
+    'meta-math/MATH',
+    'anthropic/hh-rlhf',
+    'stanfordnlp/SHP',
+]
+
+def generate_training_script(model, method, dataset, epochs, lr, batch_size, max_length, output_name):
+    template = f'''#!/usr/bin/env python3
+# Auto-generated training script
+# Model: {model}
+# Method: {method}
+# Dataset: {dataset}
+
+from trl import SFTTrainer, SFTConfig
+from datasets import load_dataset
+
+dataset = load_dataset('{dataset}', split='train_sft')
+
+training_args = SFTConfig(
+    learning_rate={lr},
+    num_train_epochs={epochs},
+    per_device_train_batch_size={batch_size},
+    gradient_accumulation_steps=4,
+    max_seq_length={max_length},
+    gradient_checkpointing=True,
+    bf16=True,
+    output_dir='./{output_name}-output',
+    push_to_hub=True,
+    hub_model_id='YOUR_USERNAME/{output_name}',
+    logging_steps=10,
+    disable_tqdm=True,
+)
+
+trainer = SFTTrainer(
+    model='{model}',
+    args=training_args,
+    train_dataset=dataset,
+)
+
+trainer.train()
+trainer.push_to_hub()
+'''
+    return template
+
+def generate_dpo_script(model, dataset, epochs, lr, batch_size, output_name):
+    template = f'''#!/usr/bin/env python3
+# DPO Training Script
+# Model: {model}
+# Dataset: {dataset}
+
+from trl import DPOTrainer, DPOConfig
+from datasets import load_dataset
+
+dataset = load_dataset('{dataset}', split='train')
+
+training_args = DPOConfig(
+    learning_rate={lr},
+    num_train_epochs={epochs},
+    per_device_train_batch_size={batch_size},
+    max_seq_length={512},
+    bf16=True,
+    output_dir='./{output_name}-output',
+    push_to_hub=True,
+    hub_model_id='YOUR_USERNAME/{output_name}',
+)
+
+trainer = DPOTrainer(
+    model='{model}',
+    args=training_args,
+    train_dataset=dataset,
+)
+
+trainer.train()
+trainer.push_to_hub()
+'''
+    return template
+
+def generate_lora_script(model, dataset, epochs, lr, batch_size, output_name):
+    template = f'''#!/usr/bin/env python3
+# LoRA Fine-tuning Script
+# Model: {model}
+# Dataset: {dataset}
+
+from trl import SFTTrainer, SFTConfig
+from peft import LoraConfig
+from datasets import load_dataset
+
+dataset = load_dataset('{dataset}', split='train_sft')
+
+peft_config = LoraConfig(
+    r=16,
+    lora_alpha=32,
+    lora_dropout=0.05,
+    task_type='CAUSAL_LM',
+)
+
+training_args = SFTConfig(
+    learning_rate={lr * 10},  # LoRA needs higher LR
+    num_train_epochs={epochs},
+    per_device_train_batch_size={batch_size},
+    max_seq_length=2048,
+    bf16=True,
+    output_dir='./{output_name}-output',
+    push_to_hub=True,
+    hub_model_id='YOUR_USERNAME/{output_name}',
+)
+
+trainer = SFTTrainer(
+    model='{model}',
+    args=training_args,
+    train_dataset=dataset,
+    peft_config=peft_config,
+)
+
+trainer.train()
+trainer.push_to_hub()
+'''
+    return template
+
+def generate_script(model, method, dataset, epochs, lr, batch_size, max_length, output_name):
+    if method == 'DPO':
+        return generate_dpo_script(model, dataset, epochs, lr, batch_size, output_name)
+    elif method == 'LoRA':
+        return generate_lora_script(model, dataset, epochs, lr, batch_size, output_name)
+    else:  # SFT, GRPO
+        return generate_training_script(model, method, dataset, epochs, lr, batch_size, max_length, output_name)
+
+def get_hardware_requirement(model):
+    if '0.5B' in model or '1B' in model:
+        return 'a10g-small (24GB VRAM)'
+    elif '3B' in model or '7B' in model:
+        return 'a10g-large or a100-small (24-80GB VRAM)'
+    else:
+        return 'a100-large (80GB VRAM)'
+
+# Gradio UI
+with gr.Blocks(title='ML Model Trainer', theme=gr.themes.Soft()) as demo:
+    gr.Markdown('# 🤖 ML Model Trainer')
+    gr.Markdown('Generate training scripts for SFT, DPO, LoRA fine-tuning')
+    
+    with gr.Row():
+        with gr.Column():
+            model = gr.Dropdown(MODELS, label='Model', value='Qwen/Qwen2.5-0.5B-Instruct')
+            method = gr.Dropdown(METHODS, label='Training Method', value='SFT')
+            dataset = gr.Dropdown(DATASETS, label='Dataset', value='HuggingFaceH4/ultrachat_200k')
+            output_name = gr.Textbox(label='Output Model Name', value='my-finetuned-model')
+        
+        with gr.Column():
+            epochs = gr.Slider(1, 10, value=3, step=1, label='Epochs')
+            lr = gr.Number(label='Learning Rate', value=2e-5)
+            batch_size = gr.Slider(1, 16, value=4, step=1, label='Batch Size')
+            max_length = gr.Slider(256, 8192, value=4096, step=256, label='Max Sequence Length')
+    
+    generate_btn = gr.Button('🔧 Generate Training Script', variant='primary')
+    
+    output_code = gr.Code(label='Training Script', language='python', lines=20)
+    
+    hardware_info = gr.Markdown('')
+    
+    def on_generate(model, method, dataset, epochs, lr, batch_size, max_length, output_name):
+        script = generate_script(model, method, dataset, epochs, lr, batch_size, max_length, output_name)
+        hw = get_hardware_requirement(model)
+        return script, f'**Recommended Hardware:** {hw} | **Timeout:** ~{int(epochs * 2)}h'
+    
+    generate_btn.click(
+        on_generate,
+        inputs=[model, method, dataset, epochs, lr, batch_size, max_length, output_name],
+        outputs=[output_code, hardware_info]
+    )
+    
+    gr.Markdown('---')
+    gr.Markdown('### 📋 How to Use')
+    gr.Markdown('''
+    1. Configure your training parameters above
+    2. Click **Generate Training Script**
+    3. Copy the script to a file (e.g., `train.py`)
+    4. Install dependencies: `pip install transformers trl torch datasets accelerate peft`
+    5. Run: `python train.py`
+    
+    **Note:** You'll need Hugging Face Pro or compute credits for cloud training.
+    ''')
+
+demo.launch()