app.py

import os
from datetime import datetime
from pathlib import Path

import numpy as np
import pandas as pd
import requests
import streamlit as st
from datasets import get_dataset_config_names
from dotenv import load_dotenv

if Path(".env").is_file():
    load_dotenv(".env")

auth_token = os.getenv("HF_HUB_TOKEN")
header = {"Authorization": "Bearer " + auth_token}

TASKS = sorted(get_dataset_config_names("ought/raft"))
# Split and capitalize the task names, e.g. banking_77 => Banking 77
FORMATTED_TASK_NAMES = sorted([" ".join(t.capitalize() for t in task.split("_")) for task in TASKS])


def extract_tags(dataset):
    tags = {}
    for tag in dataset["tags"]:
        k, v = tuple(tag.split(":", 1))
        tags[k] = v
    return tags


def download_submissions():
    response = requests.get("http://huggingface.co/api/datasets", headers=header)
    all_datasets = response.json()

    submissions = []

    for dataset in all_datasets:
        tags = extract_tags(dataset)
        if tags.get("benchmark") == "raft" and tags.get("type") == "evaluation":
            submissions.append(dataset)
    return submissions


def format_submissions(submissions):
    submission_data = {**{"Team": []}, **{"Model": []}, **{"Date": []}, **{t: [] for t in TASKS}}

    # TODO(lewtun): delete / filter all the junk repos from development
    # The following picks the latest submissions which adhere to the model card schema
    for submission in submissions:
        submission_id = submission["id"]
        response = requests.get(
            f"http://huggingface.co/api/datasets/{submission_id}?full=true",
            headers=header,
        )
        data = response.json()
        card_data = data["card_data"]
        username = card_data["submission_dataset"].split("/")[0]
        submission_data["Team"].append(username)
        submission_id = card_data["submission_id"]
        submission_name, sha, timestamp = submission_id.split("__")
        submission_data["Model"].append(submission_name)
        timestamp = pd.to_datetime(int(timestamp))
        submission_data["Date"].append(datetime.date(timestamp).strftime("%b %d, %Y"))

        for task in card_data["results"]:
            task_data = task["task"]
            task_name = task_data["name"]
            score = task_data["metrics"][0]["value"]
            submission_data[task_name].append(score)

    df = pd.DataFrame(submission_data)
    df.insert(3, "Overall", df[TASKS].mean(axis=1))
    df = df.copy().sort_values("Overall", ascending=False)
    df.rename(columns={k: v for k, v in zip(TASKS, FORMATTED_TASK_NAMES)}, inplace=True)
    # Start ranking from 1
    df.insert(0, "Rank", np.arange(1, len(df) + 1))
    return df


###########
### APP ###
###########
st.set_page_config(layout="wide")
st.title("RAFT: Real-world Annotated Few-shot Tasks")
st.markdown(
    """
Large pre-trained language models have shown promise for few-shot learning, completing text-based tasks given only a few task-specific examples. Will models soon solve classification tasks that have so far been reserved for human research assistants? 

[RAFT](https://raft.elicit.org) is a few-shot classification benchmark that tests language models:

- across multiple domains (lit review, tweets, customer interaction, etc.)
- on economically valuable classification tasks (someone inherently cares about the task)
- in a setting that mirrors deployment (50 examples per task, info retrieval allowed, hidden test set)

To submit to RAFT, follow the instruction posted on [this page](https://huggingface.co/datasets/ought/raft-submission).
"""
)
submissions = download_submissions()
df = format_submissions(submissions)
styler = df.style.set_precision(3).set_properties(**{"white-space": "pre-wrap", "text-align": "center"})
# hack to remove index column: https://discuss.streamlit.io/t/questions-on-st-table/6878/3
st.markdown(
    """
<style>
table td:nth-child(1) {
    display: none
}
table th:nth-child(1) {
    display: none
}
</style>
""",
    unsafe_allow_html=True,
)
st.table(styler)