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 wikitext (link) dataset with the distilbert-base-cased (link) model checkpoint.
Start by loading the wikitext-2-raw-v1 version of that dataset, and take the 11th example (index 10) of the train split.
We'll tokenize this using the appropriate tokenizer, and we'll mask the sixth token (index 5) the sequence.
When using the distilbert-base-cased 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)?
Tips:
- You might find the transformers docs (link) useful.
- You might find the datasets docs (link) useful.
- You might also be interested in the Hugging Face course (link).
"""
skops_question = """
1. Create a python environment[1] and install `scikit-learn` version `1.0` in that environment.
2. Using that environment, create a `LogisticRegression` model[2] and fit it on the Iris dataset[3].
3. Save the trained model using `pickle`[4] or `joblib`[5].
4. Create a second environment, and install `scikit-learn` version `1.1` in it.
5. Try loading the model you saved in step 3 in this second environment.
Question:
Is there a warning or error you receive while trying to load the model? If yes, what exactly is it.
References
- [1] You can use any tool you want to create the environment. Two of the options are:
- `venv`: https://docs.python.org/3/library/venv.html
- `mamba`: https://github.com/mamba-org/mamba
- [2] `LogisticRegression` API guide: https://scikit-learn.org/dev/modules/generated/sklearn.linear_model.LogisticRegression.html
- [3] `load_iris` API guide: https://scikit-learn.org/dev/modules/generated/sklearn.datasets.load_iris.html
- [4] `pickle`: https://docs.python.org/3/library/pickle.html
- [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.
What’s the pass@2 metric (in percent) as introduced in the
Codex paper (see section 2.1)?
References
- 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.
References
- `evaluate` library: https://huggingface.co/docs/evaluate/index
- BLEU score: https://en.wikipedia.org/wiki/BLEU
"""
embodied_question = """
We are going to use Simulate to create a basic RL environment.
Instructions:
pip install simulate
create a scene with the unity engine
add a box to the scene at position [0, 0, 1], add a camera named "cam" at default position
show the scene, step the scene once
what is the mean pixel value from the frames from "cam".
For some resources, you may want to check out:
* Simulate quick start for installation,
* for running the simulation.
"""
fast_distributed_framework_question = """
We are going to understand how many operations does a matrix multiplication hold using the simplest algorithm.
Let A,B two matrices of size 256x64 and 128x64 respectively. When computing the matrix multiplication of A and (B^T), how many scalar multiplications are done? How many scalar additions are done? Please answer in the following format:
multiplications: {YOUR_ANSWER_AS_A_SINGLE_NUMBER}
additions: {YOUR_ANSWER_AS_A_SINGLE_NUMBER}
"""
internships = {
'Accelerate': default_question,
'Skops & Scikit-Learn': skops_question,
'Diffusion distillation': default_question,
# "Evaluate": evaluate_question,
"Speech": default_question,
"TTS": default_question,
"ML for Code/Code Generation": code_question,
# "Model forgetting": default_question,
# "Multimodal AI": default_question,
# "OCR": default_question,
# "Efficient video pretraining": default_question,
# "Retrieval augmentation as prompting": default_question,
"Embodied AI": embodied_question,
# "Toolkit for detecting distribution shift/Robustness": default_question,
"Social impact evaluations": default_question,
"Gradio as an ecosystem": default_question,
"Benchmarking transformers on various AI hardware accelerators": default_question,
"AI Art Tooling Residency": default_question,
"Datasets for Large Language Models": default_question,
"Fast Distributed Training Framework": fast_distributed_framework_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()