|
import gradio as gr |
|
import torch |
|
import numpy as np |
|
from transformers import AutoModelForMultipleChoice, AutoTokenizer |
|
|
|
model_id = "microsoft/deberta-v2-xlarge" |
|
|
|
|
|
model = AutoModelForMultipleChoice.from_pretrained(model_id) |
|
tokenizer = AutoTokenizer.from_pretrained(model_id) |
|
|
|
|
|
def preprocess(text): |
|
|
|
lines = text.strip().split("\n") |
|
samples = [] |
|
|
|
|
|
for line in lines: |
|
parts = line.split("\t") |
|
if len(parts) >= 6: |
|
sample = { |
|
"prompt": parts[0], |
|
"A": parts[1], |
|
"B": parts[2], |
|
"C": parts[3], |
|
"D": parts[4], |
|
"E": parts[5] |
|
} |
|
samples.append(sample) |
|
|
|
return samples |
|
|
|
|
|
def predict(data): |
|
results = [] |
|
for sample in data: |
|
first_sentences = [sample["prompt"]] * 5 |
|
second_sentences = [sample[option] for option in "ABCDE"] |
|
tokenized_sentences = tokenizer(first_sentences, second_sentences, truncation=True, padding=True, return_tensors="pt") |
|
inputs = tokenized_sentences["input_ids"] |
|
masks = tokenized_sentences["attention_mask"] |
|
with torch.no_grad(): |
|
logits = model(inputs, attention_mask=masks).logits |
|
predictions_as_ids = torch.argsort(-logits, dim=1) |
|
answers = np.array(list("ABCDE"))[predictions_as_ids.tolist()] |
|
results.append(["".join(i) for i in answers[:, :3]]) |
|
return results |
|
|
|
|
|
iface = gr.Interface( |
|
fn=predict, |
|
inputs=gr.inputs.Textbox(placeholder="Paste multiple-choice questions (prompt and options separated by tabs, one question per line) ..."), |
|
outputs=gr.outputs.Label(num_top_classes=3), |
|
live=True, |
|
title="LLM Science Exam Demo", |
|
description="Enter multiple-choice questions (prompt and options) below and get predictions.", |
|
) |
|
|
|
|
|
iface.launch() |
|
iface.integrate(wandb=wandb) |