app.py

import json
import os
import shutil
from datetime import datetime
from pathlib import Path

import streamlit as st
from dotenv import load_dotenv
from huggingface_hub import HfApi, Repository

from utils import http_post, validate_json

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

HF_TOKEN = os.getenv("HF_TOKEN")
AUTONLP_USERNAME = os.getenv("AUTONLP_USERNAME")
HF_AUTONLP_BACKEND_API = os.getenv("HF_AUTONLP_BACKEND_API")
LOCAL_REPO = "submission_repo"

## TODO ##
# 1. Add check that fields are nested under `tasks` field correctly
# 2. Add check that names of tasks and datasets are valid


###########
### APP ###
###########
st.title("GEM Submissions")
st.markdown(
    """
    Welcome to the [GEM benchmark](https://gem-benchmark.com/)! GEM is a benchmark
    environment for Natural Language Generation with a focus on its Evaluation, both
    through human annotations and automated Metrics.

    GEM aims to:

    - measure NLG progress across many NLG tasks across languages.
    - audit data and models and present results via data cards and model robustness
    reports.
    - develop standards for evaluation of generated text using both automated and
    human metrics.

    Use this page to submit your system's predictions to the benchmark.
    """
)

with st.form(key="form"):
    # Flush local repo
    shutil.rmtree(LOCAL_REPO, ignore_errors=True)
    submission_errors = 0
    uploaded_file = st.file_uploader("Upload submission.json file", type=["json"])

    if uploaded_file:
        if uploaded_file.name != "submission.json":
            st.error(f"⛔ Invalid filename. Please upload a submission.json file.")
            submission_errors += 1
        else:
            data = str(uploaded_file.read(), "utf-8")
            json_data = json.loads(data)
            is_valid, message = validate_json(json_data)
            if is_valid:
                st.success(message)
            else:
                st.error(message)
                submission_errors += 1

    with st.expander("Submission format"):
        st.markdown(
            """
        Please follow this JSON format for your `submission.json` file:

        ```json
        {
        "submission_name": "An identifying name of your system",
        "param_count": 123, # The number of parameters your system has.
        "description": "An optional brief description of the system that will be shown on the results page",
        "tasks":
            {
            "dataset_identifier": {
                "values": ["output-0", "output-1", "..."], # A list of system outputs.
                "keys": ["gem_id-0", "gem_id-1", ...] # A list of GEM IDs.
                }
            }
        }
        ```
        Here, `dataset_identifier` is the identifier of the dataset followed by
        an identifier of the set the outputs were created from, for example
        `_validation` or `_test`. For example, the `mlsum_de` test set has the
        identifier `mlsum_de_test`. The `keys` field is needed to avoid
        accidental shuffling that will impact your metrics. Simply add a list of
        IDs from the `gem_id` column of each evaluation dataset in the same
        order as your values. Please see the sample submission below:
        """
        )
        with open("sample-submission.json", "r") as f:
            example_submission = json.load(f)
            st.json(example_submission)

    token = st.text_input(
        "Enter 🤗 Hub access token",
        type="password",
        help="You can generate an access token via your 🤗 Hub settings. See the [docs](https://huggingface.co/docs/hub/security#user-access-tokens) for more details",
    )
    if token:
        try:
            user_info = HfApi().whoami(token)
        except Exception as e:
            st.error("⛔ Invalid access token")
            submission_errors += 1

    submit_button = st.form_submit_button("Make Submission")

if submit_button and submission_errors == 0:
    st.write("⏳ Preparing submission for evaluation ...")
    user_name = user_info["name"]
    submission_name = json_data["submission_name"]

    # Create submission dataset under benchmarks ORG
    dataset_repo_url = f"https://huggingface.co/datasets/GEM-submissions/gem-{user_name}"
    repo = Repository(
        local_dir=LOCAL_REPO,
        clone_from=dataset_repo_url,
        repo_type="dataset",
        private=True,
        use_auth_token=HF_TOKEN,
    )
    submission_metadata = {"benchmark": "gem", "type": "prediction", "submission_name": submission_name}
    repo.repocard_metadata_save(submission_metadata)

    with open(f"{LOCAL_REPO}/submission.json", "w", encoding="utf-8") as f:
        json.dump(json_data, f)

    # TODO: add informative commit msg
    commit_url = repo.push_to_hub()
    if commit_url is not None:
        commit_sha = commit_url.split("/")[-1]
    else:
        commit_sha = repo.git_head_commit_url().split("/")[-1]

    submission_time = str(int(datetime.now().timestamp()))
    submission_id = submission_name + "__" + commit_sha + "__" + submission_time

    payload = {
        "username": AUTONLP_USERNAME,
        "dataset": "GEM/references",
        "task": 1,
        "model": "gem",
        "submission_dataset": f"GEM-submissions/gem-{user_name}",
        "submission_id": submission_id,
        "col_mapping": {},
        "split": "test",
        "config": None,
    }
    json_resp = http_post(
        path="/evaluate/create", payload=payload, token=HF_TOKEN, domain=HF_AUTONLP_BACKEND_API
    ).json()

    if json_resp["status"] == 1:
        st.success(f"✅ Submission {submission_name} was successfully submitted for evaluation!")
    else:
        st.error("🙈 Oh noes, there was an error submitting your submission! Please contact the organisers")

    # Flush local repo
    shutil.rmtree(LOCAL_REPO, ignore_errors=True)