app.py

import streamlit as st
import os

st.set_page_config(page_title='ML Model Trainer', page_icon='🤖')

st.title('🤖 ML Model Trainer')
st.markdown('Generate training scripts for SFT, DPO, LoRA fine-tuning')

# Model options
MODELS = [
    'Qwen/Qwen2.5-0.5B-Instruct',
    'Qwen/Qwen2.5-1.5B-Instruct',
    'Qwen/Qwen2.5-7B-Instruct',
    'meta-llama/Llama-3.2-1B-Instruct',
    'meta-llama/Llama-3.2-3B-Instruct',
    'microsoft/Phi-3-mini-128k-instruct',
    'google/gemma-2b-it',
    'mistralai/Mistral-7B-Instruct-v0.3',
]

METHODS = ['SFT', 'DPO', 'LoRA']

DATASETS = [
    'HuggingFaceH4/ultrachat_200k',
    'openai/gsm8k',
    'meta-math/MATH',
    'anthropic/hh-rlhf',
]

col1, col2 = st.columns(2)

with col1:
    model = st.selectbox('Model', MODELS)
    method = st.selectbox('Training Method', METHODS)
    dataset = st.selectbox('Dataset', DATASETS)
    output_name = st.text_input('Output Model Name', 'my-finetuned-model')

with col2:
    epochs = st.slider('Epochs', 1, 10, 3)
    lr = st.number_input('Learning Rate', value=2e-5, format='%.0e')
    batch_size = st.slider('Batch Size', 1, 16, 4)
    max_length = st.slider('Max Sequence Length', 256, 8192, 4096, 256)

if st.button('🔧 Generate Training Script', type='primary'):
    if method == 'SFT':
        script = f'''#!/usr/bin/env python3
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()
'''
    elif method == 'DPO':
        script = f'''#!/usr/bin/env python3
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()
'''
    else:  # LoRA
        script = f'''#!/usr/bin/env python3
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},
    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()
'''
    
    st.code(script, language='python')
    
    # Hardware info
    if '0.5B' in model or '1B' in model:
        hw = 'a10g-small (24GB VRAM) - ~2h training'
    elif '3B' in model or '7B' in model:
        hw = 'a10g-large or a100-small (24-80GB VRAM) - ~4h training'
    else:
        hw = 'a100-large (80GB VRAM) - ~6h training'
    
    st.info(f'**Recommended:** {hw}')

st.markdown('---')
st.markdown('### 📋 How to Use')
st.markdown('''
1. Configure parameters above
2. Click **Generate Training Script**
3. Copy the script to `train.py`
4. Install: `pip install transformers trl torch datasets accelerate peft`
5. Run: `python train.py`

**Note:** Need Hugging Face Pro or compute credits for cloud training.
''')