app.py

import gradio as gr
import torch
from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM, AutoTokenizer

# Repos
BASE_MODEL = "deepseek-ai/DeepSeek-R1-Distill-Qwen-7B"
ADAPTER_REPO = "cheberle/autotrain-35swc-b4r9z"

# 1. Load the PEFT config to confirm the base model
peft_config = PeftConfig.from_pretrained(ADAPTER_REPO)
print("PEFT Base Model:", peft_config.base_model_name_or_path)

# 2. Load the tokenizer & base model
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, trust_remote_code=True)
base_model = AutoModelForCausalLM.from_pretrained(BASE_MODEL, trust_remote_code=True)

# 3. Load your LoRA adapter weights onto the base model
model = PeftModel.from_pretrained(base_model, ADAPTER_REPO)

def classify_text(text):
    """
    Simple prompting approach: we ask the model to return a single classification label 
    (e.g., 'positive', 'negative', etc.).
    You can refine this prompt, add chain-of-thought, or multiple classes as needed.
    """
    prompt = f"Below is some text.\nText: {text}\nPlease classify the sentiment (positive or negative):"
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    with torch.no_grad():
        outputs = model.generate(**inputs, max_new_tokens=64)
    answer = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return answer

with gr.Blocks() as demo:
    gr.Markdown("## Qwen + LoRA Adapter: Text Classification Demo")
    input_box = gr.Textbox(lines=3, label="Enter text")
    output_box = gr.Textbox(lines=3, label="Model's generated output (classification)")

    classify_btn = gr.Button("Classify")
    classify_btn.click(fn=classify_text, inputs=input_box, outputs=output_box)

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