Giters

mapmeld / use-this-now

Links to new technologies which improve on the tech which I used in old posts

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

I wrote multiple blog posts in 2019-present which relied on older methods of NLP, older Solidity tools, etc. My goal is to link the old blog posts to a header in this README and update one central repo as these changes occur.

Anti-Explanations

No new updates

Arabic NLP

Datasets (various uses)

https://twitter.com/zaidalyafeai/status/1448667731675398145

Dialect + Generative Model

Pre-trained Language Models

Sentiment Analysis

TBD

Blockchain

Census API

I wrote an update about using the 2020 Census API and calculating 6+ race population by state and county https://blog.georeactor.com/census-1

Crowdfunded AI

No updates

Facial recognition

  • No updates on head coverings datasets

Gender bias

Shown that embedding bias was not reflected in model output biases https://twitter.com/seraphinagt/status/1400769594005000192

Gender re-inflection

My Arabic seq2seq model: https://huggingface.co/monsoon-nlp/ar-seq2seq-gender-decoder

My Spanish seq2seq model: https://huggingface.co/monsoon-nlp/es-seq2seq-gender-decoder

Updated and expanded Arabic dataset from NYU AD: https://camel.abudhabi.nyu.edu/arabic-parallel-gender-corpus/

Large-scale Facebook / Meta project: https://ai.facebook.com/blog/measure-fairness-and-mitigate-ai-bias/

I18n

JAX tutorials

A video tutorial: https://www.youtube.com/watch?v=SstuvS-tVc0

Notebook tutorials: https://github.com/AakashKumarNain/TF_JAX_tutorials

Model Editing

  • Deleting / forgetting information from models continues to be active research. -- Counterfactuals in avoiding memorized sequences in models https://arxiv.org/abs/2112.12938

  • Patching / updating models continues to be active research.

Model Introspection

TBD

Model training

For large transformer models, use Transformers / Trainer

Negative results

No updates

NLP + AAVE

Object recognition / YOLO

OpenStreetMap

Reddit datasets

  • PushShift.io is the best place to download Reddit data
  • Using Reddit as a source was highly criticized in Delphi moral AI project

South Asian Language Model Projects

  • Use Google's MuRIL model for Hindi, Tamil, Bangla, and other main languages of India (original and transliterated to Latin alphabet). HuggingFace and TFHub links.
  • I have a few transfer learning experiments for Dhivehi , and local developers created TTS https://huggingface.co/models?filter=dv

Text Augmentation and Attack Libraries

No new changes

Text To Speech

  • Kinyarwanda and Luganda are success stories on Mozilla CommonVoice
  • TTS library continues to be developed by Coqui
  • Google mSLAM multilingual model for text and speech https://arxiv.org/abs/2202.01374

Thai NLP

Pretrained models:

Or largest pretrained ByT5 model you can use: https://huggingface.co/models?search=byt5

Tokenization

Toxicity / Hate Speech in NLP

Latest overview post: https://mapmeld.medium.com/its-not-easy-being-clean-ee217ed4825c

Perspective API is available for more languages: https://developers.perspectiveapi.com/s/about-the-api-attributes-and-languages

Video thesis defense around 'finding and fixing undesirable behaviors' of language models https://www.youtube.com/watch?v=BgcU_kytMf8

Annotator bias when labeling toxicity of AAE/AAVE language https://arxiv.org/abs/2111.07997

New effort to create a dataset: https://github.com/surge-ai/toxicity

A "Red Team" language model helping find problem responses in other language models https://deepmind.com/research/publications/2022/Red-Teaming-Language-Models-with-Language-Models

Request permission to access this dataset: https://huggingface.co/datasets/irlab-udc/metahate

About

Links to new technologies which improve on the tech which I used in old posts