app.py

# import gradio as gr
# import torch
# from transformers import AutoTokenizer, AutoModelForCausalLM

# MODEL_ID = "AREEBAFATIMA12/SmolLM-135M-SFT-DPO"

# print("Loading model...")
# tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
# tokenizer.pad_token = tokenizer.eos_token
# model = AutoModelForCausalLM.from_pretrained(MODEL_ID, torch_dtype=torch.float32).eval()
# print("Model ready.")

# def respond(prompt):
#     if not prompt.strip():
#         return "Please enter a question."
#     formatted = f"<|user|>\n{prompt}</s>\n<|assistant|>\n"
#     inputs = tokenizer(formatted, return_tensors="pt", truncation=True, max_length=512)
#     with torch.no_grad():
#         out = model.generate(
#             **inputs,
#             max_new_tokens=200,
#             do_sample=False,
#             pad_token_id=tokenizer.eos_token_id,
#             eos_token_id=tokenizer.eos_token_id,
#         )
#     return tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True).strip()

# demo = gr.Interface(
#     fn=respond,
#     inputs=gr.Textbox(lines=3, placeholder="Ask me anything...", label="Your Question"),
#     outputs=gr.Textbox(label="Model Response", lines=6),
#     title="SmolLM-135M (SFT + DPO)",
#     description="SmolLM-135M fine-tuned on Dolly-15k (SFT) and Orca DPO pairs (DPO). Built for IBA NLP Assignment 04.",
#     examples=[
#         ["What causes seasons on Earth?"],
#         ["What is the capital of Australia?"],
#         ["Explain photosynthesis briefly."],
#         ["Give me 3 tips to reduce plastic waste."],
#     ],
#     theme=gr.themes.Soft()
# )

# demo.launch(server_name="0.0.0.0", server_port=7860)


import gradio as gr
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

BASE_ID  = "HuggingFaceTB/SmolLM-135M"
SFT_ID   = "AREEBAFATIMA12/SmolLM-135M-SFT-only"
DPO_ID   = "AREEBAFATIMA12/SmolLM-135M-SFT-DPO"

print("Loading all 3 models...")

# Base
base_tokenizer = AutoTokenizer.from_pretrained(BASE_ID)
base_tokenizer.pad_token = base_tokenizer.eos_token
base_model = AutoModelForCausalLM.from_pretrained(BASE_ID, torch_dtype=torch.float32).eval()

# SFT
sft_tokenizer = AutoTokenizer.from_pretrained(SFT_ID)
sft_tokenizer.pad_token = sft_tokenizer.eos_token
sft_model = AutoModelForCausalLM.from_pretrained(SFT_ID, torch_dtype=torch.float32).eval()

# SFT+DPO
dpo_tokenizer = AutoTokenizer.from_pretrained(DPO_ID)
dpo_tokenizer.pad_token = dpo_tokenizer.eos_token
dpo_model = AutoModelForCausalLM.from_pretrained(DPO_ID, torch_dtype=torch.float32).eval()

print("All models loaded.")

def generate(model, tokenizer, prompt, template="dpo"):
    if template == "base":
        formatted = prompt
    elif template == "sft":
        formatted = f"### Instruction:\n{prompt}\n\n### Response:\n"
    else:
        formatted = f"<|user|>\n{prompt}</s>\n<|assistant|>\n"
    
    inputs = tokenizer(formatted, return_tensors="pt", truncation=True, max_length=512)
    with torch.no_grad():
        out = model.generate(
            **inputs,
            max_new_tokens=150,
            do_sample=False,
            pad_token_id=tokenizer.eos_token_id,
            eos_token_id=tokenizer.eos_token_id,
        )
    return tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True).strip()

def compare(prompt):
    if not prompt.strip():
        return "", "", ""
    base_out = generate(base_model, base_tokenizer, prompt, template="base")
    sft_out  = generate(sft_model,  sft_tokenizer,  prompt, template="sft")
    dpo_out  = generate(dpo_model,  dpo_tokenizer,  prompt, template="dpo")
    return base_out, sft_out, dpo_out

with gr.Blocks(theme=gr.themes.Soft()) as demo:
    gr.Markdown("# SmolLM-135M: Base vs SFT vs SFT+DPO")
    gr.Markdown("Compare outputs across the full fine-tuning pipeline. Built for IBA NLP Assignment 04.")
    
    prompt_box = gr.Textbox(lines=2, placeholder="Ask a question...", label="Your Prompt")
    btn = gr.Button("Compare All 3 Models", variant="primary")
    
    with gr.Row():
        out_base = gr.Textbox(label="Base Model (no tuning)", lines=6)
        out_sft  = gr.Textbox(label="SFT Only (Trial 3)", lines=6)
        out_dpo  = gr.Textbox(label="SFT + DPO (Final)", lines=6)
    
    gr.Examples(
        examples=[
            ["What causes seasons on Earth?"],
            ["What is the capital of Australia?"],
            ["Explain photosynthesis briefly."],
            ["Give me 3 tips to reduce plastic waste."],
            ["What are the planets in our solar system?"],
        ],
        inputs=prompt_box
    )
    
    btn.click(fn=compare, inputs=prompt_box, outputs=[out_base, out_sft, out_dpo])

demo.launch(server_name="0.0.0.0", server_port=7860)