app.py

import vllm
import torch
import gradio
import huggingface_hub
import os

huggingface_hub.login(token=os.environ["HF_TOKEN"])
hf_writer = gradio.HuggingFaceDatasetSaver(os.environ["HF_WRITE_TOKEN"], "fava-flagged-demo")

# Fava prompt
INPUT = "Read the following references:\n{evidence}\nPlease identify all the errors in the following text using the information in the references provided and suggest edits if necessary:\n[Text] {output}\n[Edited] "

model = vllm.LLM(model="fava-uw/fava-model") 
def result(passage, reference):
    prompt = [INPUT.format_map({"evidence":reference, "output":passage})]
    print(prompt)
    sampling_params = vllm.SamplingParams(
        temperature=0,
        top_p=1.0,
        max_tokens=500,
    )
    outputs = model.generate(prompt, sampling_params)
    outputs = [it.outputs[0].text for it in outputs]
    output = outputs[0].replace("<mark>", "<span style='color: green; font-weight: bold;'> ")
    output = output.replace("</mark>", " </span>")
    output = output.replace("<delete>", "<span style='color: red; text-decoration: line-through;'>")
    output = output.replace("</delete>", "</span>")
    output = output.replace("<entity>", "<span style='background-color: #E9A2D9; border-bottom: 1px dotted;'>entity</span>")
    output = output.replace("<relation>", "<span style='background-color: #F3B78B; border-bottom: 1px dotted;'>relation</span>")
    output = output.replace("<contradictory>", "<span style='background-color: #FFFF9B; border-bottom: 1px dotted;'>contradictory</span>")
    output = output.replace("<unverifiable>", "<span style='background-color: #D3D3D3; border-bottom: 1px dotted;'>unverifiable</span><u>")
    output = output.replace("<invented>", "<span style='background-color: #BFE9B9; border-bottom: 1px dotted;'>invented</span>")
    output = output.replace("<subjective>", "<span style='background-color: #D3D3D3; border-bottom: 1px dotted;'>subjective</span><u>")
    output = output.replace("</entity>", "")
    output = output.replace("</relation>", "")
    output = output.replace("</contradictory>", "")
    output = output.replace("</unverifiable>", "</u>")
    output = output.replace("</invented>", "")
    output = output.replace("</subjective>", "</u>")
    output = output.replace("Edited:", "")
    return f'<div style="font-weight: normal;">{output}</div>'; #output;

if __name__ == "__main__":
    article = """<center><img src='https://github.com/abhika-m/researchpapers/blob/main/fava.png?raw=true' width="650px"'><img src='https://github.com/abhika-m/researchpapers/blob/main/taxonomy.png?raw=true' width="850px"></center>"""
    description = """Given a passage and a reference, FAVA will detect and edit any hallucinations present in the passage. If you find any errors with FAVA's output, please flag it. For more information, check out our <a href="https://arxiv.org/abs/2401.06855" target='_blank'>paper</a>."""
    examples = [["Canada's oldest national park, Banff, was established in 1886. It recently won a Nature's Choice 2023 award for its beautiful mountainous terrain. It's the best national park ever.", 
                 "Banff National Park is Canada's oldest national park, established in 1885 as Rocky Mountains Park. Located in Alberta's Rocky Mountains, 110–180 kilometres (68–112 mi) west of Calgary, Banff encompasses 6,641 square kilometres (2,564 sq mi) of mountainous terrain."]]
    demo = gradio.Interface(fn=result, inputs=["text", "text"], outputs="html", title="Fine-grained Hallucination Detection and Editing (FAVA)", 
                            description=description, article=article,
                           examples=examples, allow_flagging="manual", flagging_options=["wrong detection", "wrong edit", "both wrong", "other"], flagging_callback=hf_writer)
    demo.launch(share=True)