- 👉 Register: https://forms.gle/2d5kabsuvnUx6xX76
- 📂 Download the Data: Check the "data" folder
- 🏗️ Build your model: A quick recipe to get you upto speed
- 🏟️ Submit your predictions at: https://codalab.lisn.upsaclay.fr/competitions/14977
- 🗓️ Competition Deadline: 12th October
- 🏆 Win prizes worth INR 1 Lakhs
- 🎥 https://www.youtube.com/watch?v=AY9qM_LFNHA
- 🎥 A more longish tutorial - https://www.youtube.com/watch?v=L3fa8Frmqro
- Get a model finetuned for intent detection, say on the MASSIVE dataset ➞ https://huggingface.co/models?dataset=dataset:AmazonScience%2Fmassive&sort=trending
- Finetune the model on our training dataset ➞
- Obtain the predictions and submit
- Data folder: https://github.com/krishnamrith12/2023-IndoML-Datathon-Starter-kit/tree/main/data
- Training data: surprise_data.zip
- Test data: massive_test.data
- Use your favourite text editor or IDE to see the data
- Discord https://discord.com/invite/zqp7Pma3TA
- Tutorials playlist: https://www.youtube.com/playlist?list=PLvzThk3qRkmVe1v2wRRKnQeV7BHIY5xQW
- Datathon website - https://sites.google.com/view/datathon-indoml23
- IndoML website - https://indoml.in/
- Linktree - - https://linktr.ee/datathon
In the online tutorial we will go over the notebooks of this repo. We focus on the AmazonScience/MASSIVE multilingual intent-detection.
There are few different parts to the tutorial:
- EDA.ipynb: We first explore the dataset a bit to understand the task and various stats about the data. We also get to see how to use the
datasets
library from HuggingFace, what metadata information is available, and how to apply preprocessing to the data. - FinetuneTransformer.ipynb: We then finetune a transformer model on the dataset. We use the
Trainer
class from HuggingFace to do this. - LLM_PromptEngineering.ipynb: We then explore the use of LLMs for intent detection. We explore an in-context learning method to solve the task using a instruction-finetuned large-language model (LLM), without any finetuning.
There are also a few other bonus scripts available that showcases the use of other advanced techniques like parameter efficient tuning and low-rank adaptation methods. These methods will be useful for the datathon, to tackle the few-shot challenge in later phases and running on GPU devices available through Colab.
Best options are Kaggle Notebooks and Google Colab.
##IndoML-2023 #Datathon
based on materials from the tutorial by @bsantraigi, titled Intent Detection: From Sesame Street to LLMs. Tutorial video available here: https://www.youtube.com/watch?v=L3fa8Frmqro
- Nice blog about training very large models (Falcon-7b!) on Colab: The Falcon has landed in the Hugging Face ecosystem
- Colab Link: Finetuning Falcon-7b