app.py

import json
import os
from io import BytesIO

import gradio as gr
from huggingface_hub import upload_file

default_question = """
We're going to use the <a href="https://huggingface.co/datasets/wikitext" target="_blank"><code>wikitext (link)</a></code> dataset with the <code><a href="https://huggingface.co/distilbert-base-cased" target="_blank">distilbert-base-cased (link)</a></code> model checkpoint.

<br/><br/>

Start by loading the <code>wikitext-2-raw-v1</code> version of that dataset, and take the 11th example (index 10) of the <code>train</code> split.<br/>
We'll tokenize this using the appropriate tokenizer, and we'll mask the sixth token (index 5) the sequence.

<br/><br/>

When using the <code>distilbert-base-cased</code> checkpoint to unmask that (sixth token, index 5) token, what is the most probable predicted token (please provide the decoded token, and not the ID)?

<br/>
<br/>
Tips:
<br/>
- You might find the <a href="https://huggingface.co/docs/transformers/index" target="_blank">transformers docs (link)</a> useful.
<br/>
- You might find the <a href="https://huggingface.co/docs/datasets/index" target="_blank">datasets docs (link)</a> useful.
<br/>
- You might also be interested in the <a href="https://huggingface.co/course" target="_blank">Hugging Face course (link)</a>.
"""

skops_question = """
1. Create a python environment[1] and install `scikit-learn` version `1.0` in that environment.
<br/>
2. Using that environment, create a `LogisticRegression` model[2] and fit it on the Iris dataset[3].
<br/>
3. Save the trained model using `pickle`[4] or `joblib`[5].
<br/>
4. Create a second environment, and install `scikit-learn` version `1.1` in it.
<br/>
5. Try loading the model you saved in step 3 in this second environment.

<br/>
<br/>
Question:
<br/>
Is there a warning or error you receive while trying to load the model? If yes, what exactly is it.

<br/>
<br/>
References
<br/>
- [1] You can use any tool you want to create the environment. Two of the options are:
<br/>
- `venv`: https://docs.python.org/3/library/venv.html
<br/>
- `mamba`: https://github.com/mamba-org/mamba
<br/>
- [2] `LogisticRegression` API guide: https://scikit-learn.org/dev/modules/generated/sklearn.linear_model.LogisticRegression.html
<br/>
- [3] `load_iris` API guide: https://scikit-learn.org/dev/modules/generated/sklearn.datasets.load_iris.html
<br/>
- [4] `pickle`: https://docs.python.org/3/library/pickle.html
<br/>
- [5] - `joblib`: https://joblib.readthedocs.io/en/latest/
"""

code_question = """
You are probing your code generation model on a program synthesis benchmark and 
1 out of 4 the candidate solutions produced by your model pass the unit tests of a coding challenge. 
<br/>
<br/>
What’s the pass@2 metric (in percent) as introduced in the 
Codex paper (see section 2.1)?
<br/>
<br/>
References
<br/>
- Codex paper: https://arxiv.org/abs/2107.03374
"""

evaluate_question = """
Use the `evaluate` library to compute the BLEU score of the model generation `"Evaluate is a library to evaluate Machine Learning models"` and the reference solution `"Evaluate is a library to evaluate ML models"`. Round the result to two digits after the comma.
<br/>
<br/>
References
<br/>
- `evaluate` library: https://huggingface.co/docs/evaluate/index
- BLEU score: https://en.wikipedia.org/wiki/BLEU
"""

internships = {
    'Accelerate': default_question,
    'Diffusion distillation': default_question,
    'Skops & Scikit-Learn': skops_question,
    "Code Generation": code_question,
    "Document AI Democratization": default_question,
    "Evaluate": evaluate_question,
    "ASR": default_question,
    "Efficient video pretraining": default_question,
    "Embodied AI": default_question,
    "Emergence of scene and text understanding": default_question,
    "Everything is multimodal": default_question,
    "Everything is vision": default_question,
    "Retrieval augmentation as prompting": default_question,
    "Social impact evaluations": default_question,
    "Toolkit for detecting distribution shift": default_question,
    "AI Art Tooling Residency": default_question,
    "Gradio as an ecosystem": default_question,
    "Benchmarking transformers on various AI hardware accelerators": default_question,
}


with gr.Blocks() as demo:
    gr.Markdown(
        """
    # Internship introduction
    Please select the internship you would like to apply to and answer the question asked in the Answer box.
    """
    )

    internship_choice = gr.Dropdown(label='Internship', choices=list(internships.keys()))

    with gr.Column(visible=False) as details_col:
        summary = gr.HTML(label='Question')
        details = gr.Textbox(label="Answer")
        username = gr.Textbox(label="Hugging Face Username")
        comment = gr.Textbox(label="Any comment?")
        generate_btn = gr.Button("Submit")
        output = gr.Label()

    def filter_species(species):
        return gr.Label.update(
            internships[species]
        ), gr.update(visible=True)

    internship_choice.change(filter_species, internship_choice, [summary, details_col])

    def on_click(_details, _username, _internship_choice, _comment):
        response = {'response': _details, "internship": _internship_choice, "comment": _comment}
        upload_file(
            path_or_fileobj=BytesIO(bytes(json.dumps(response), 'utf-8')),
            path_in_repo=_username,
            repo_id='internships/internships-2023',
            repo_type='dataset',
            token=os.environ['HF_TOKEN']
        )
        return f"Submitted: '{_details}' for user '{_username}'"

    generate_btn.click(on_click, inputs=[details, username, internship_choice, comment], outputs=[output])


if __name__ == "__main__":
    demo.launch()