- Notifications
You must be signed in to change notification settings - Fork2
This repository provides a selection of very basic and minimal notebooks for various NLP tasks in written in Jax.
License
deterministic-algorithms-lab/NLP-Journey
Folders and files
Name | Name | Last commit message | Last commit date | |
---|---|---|---|---|
Repository files navigation
A follow up repository ofJax-Journey. This repository provides a selection of notebooks for various NLP tasks, which are completely see-through (i.e., you can see the implementation till the basic Jax/Haiku modules, in a single notebook). These were meant to be used as further tutorials in Jax for NLP, and as a guide for the coding style followed in thisawesome article by Madison May.
These notebooks, although mostly code, also mention the nuanced features, often missed when using off-the-shelf models. Moreover, they allow you to optimize everything right to the innermost modules. Also, we mention how to adapt the model to your use case, in each notebook.
A basic introductory notebook consisting of the originalRoBERTa initialized version andrandomly initialized version .
Here we realise the need for restructuring the code, and correspondingly, place all the code component-wise insrc/
. The new things we code over the original implementation are:
- The masking function for MLMhere,
- AHuggingFace Tokenizers based tokenizer,here
- A Language Embedding for TLM task,here.
- Additionally, we include an option to make the transformer auto-regressive and add a mask for the same,here. This is needed for CLM.
The final notebook can be foundhere.
Hoping to create a kind, giving, an open community that forms deep connetions by working together. Join us here :https://discord.gg/s6xSHG94u5
About
This repository provides a selection of very basic and minimal notebooks for various NLP tasks in written in Jax.
Topics
Resources
License
Uh oh!
There was an error while loading.Please reload this page.
Stars
Watchers
Forks
Releases
Packages0
Uh oh!
There was an error while loading.Please reload this page.