AI: LLMs, platforms (hardware) and projects

(AI in Business) - Applied Artificial Intelligence

https://www.bookdepository.com/Applied-Artificial-Intelligence-Where-AI-Can-Be-Used-Business-Francesco-Corea/9783319772516?ref=grid-view&qid=1592465308969&sr=1-1 

Web element detection

OpenDILabCommunity/webpage_element_detection · Hugging Face

Layout parser - doc ingester

https://layout-parser.github.io/

YOLO Object Detection
https://www.v7labs.com/blog/yolo-object-detection

Data generation - Langchain

Synthetic data generation | 🦜️🔗 Langchain

DistilBERT: Faster version of BERT

[1910.01108] DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter (arxiv.org)

DBRX - a 'mixture of experts' LLM

Introducing DBRX: A New State-of-the-Art Open LLM | Databricks

MoE in depth article 

https://cameronrwolfe.substack.com/p/moe-llms?utm_medium=android&triedRedirect=true

Deploying and Fine-tuning Deepseek

https://huggingface.co/blog/deepseek-r1-aws

Hermes-2 Mistral 7B - good for "Function Calling" (so LLM delegates creating/modifying an app, to the client)

https://huggingface.co/NousResearch/Hermes-2-Pro-Mistral-7B

multimodal LLMs

HuggingFaceM4/idefics2-8b-chatty · Hugging Face

HuggingFaceM4/idefics2-8b · Hugging Face

HPT 1.5 Air: A New Open-Sourced 8B Multimodal LLM with Llama 3

Datasets - Roboflow (like HuggingFace for data)

https://public.roboflow.com/

Accelerating Hugging Face Transformers with AWS Inferentia2

AWS Inferentia2 (inf2 AWS instances - cheaper than g5)

Prompt hacking and testing

HackAPrompt

https://github.com/Giskard-AI/giskard

Token counts - for cost estimates

https://huggingface.co/Xenova/claude-tokenizer

https://huggngface.co/spaces/Xenova/the-tokenizer-playground

https://docs.aws.amazon.com/bedrock/latest/userguide/model-customization-prepare.html

LLM powered autonomous agents

https://lilianweng.github.io/posts/2023-06-23-agent/

Building a NN from scratch

MNIST Neural Network From Scratch (kaggle.com)

Data science and data trends (via ML) Python libraries

https://www.marktechpost.com/2024/05/15/10-python-packages-revolutionizing-data-science-workflow/

evaluating LLMs

prometheus-2

https://github.com/prometheus-eval/prometheus-eval

LLM GitHub accelerator projects 

https://github.blog/2024-05-23-2024-github-accelerator-meet-the-11-projects-shaping-open-source-ai/

distilbert - can train it to classify intent or toxic

https://huggingface.co/distilbert/distilbert-base-uncased

Prompt engineering guides

- The Prompt Report - https://arxiv.org/pdf/2406.06608

- https://www.promptingguide.ai

- Anthropic - [Answer Key] Anthropic's Prompt Engineering Interactive Tutorial [PUBLIC ACCESS] - Google Spreadsheets

- Claude Prompt Engineering & Function Calls

- Claude - Tool Use

Tiny Multi-modal LLM - regions + OCR

microsoft/Florence-2-large · Hugging Face

AWS Bedrock Tips

- batch inference

https://docs.aws.amazon.com/bedrock/latest/userguide/batch-inference.html

AWS Inferentia - speculative decoding

Faster LLMs with speculative decoding and AWS Inferentia2 | AWS Machine Learning Blog

AWS Prompt Routing and prompt caching

- can auto pick an LLM?

Reduce costs and latency with Amazon Bedrock Intelligent Prompt Routing and prompt caching (preview) | AWS News Blog

SOTA LLM by Richard

https://huggingface.co/BasiliskLabs/Sovereign-0.1-72B

AI Papers, Books and Datasets

Learning Transferable Visual Models From Natural Language Supervision = 
https://arxiv.org/abs/2103.00020

-

Datasets

from AI Engineering:

Resources for Publicly Available Datasets

Here are a few resources where you can look for publicly available datasets. While you should take advantage of available data, you should never fully trust it. Data needs to be thoroughly inspected and validated.

Always check a dataset's license before using it. Try your best to understand where the data comes from. Even if a dataset has a license that allows commercial use, it's possible that part of it comes from a source that doesn't:

1. Hugging Face (https://oreil.ly/tlt5h) and Kaggle (https://oreil.ly/g8A4a) each host
   hundreds of thousands of datasets.
2. Google has a wonderful and underrated Dataset Search (https://oreil.ly/TgOaR).
3. Governments are often great providers of open data. Data.gov (https://data.gov) hosts hundreds of thousands of datasets, and data.gov.in (https://data.gov.in) hosts tens of thousands.
4. University of Michigan's Institute for Social Research (https://oreil.ly/VhVzp)
   ICPSR has data from tens of thousands of social studies.
5. UC Irvine's Machine Learning Repository (https://oreil.ly/jAR9e) and OpenML (https://oreil.ly/d-Yty) are two older dataset repositories, each hosting several thousand datasets.
6. The Open Data Network (https://oreil.ly/_tW6P) lets you search among tens of thousands of datasets.
7. Cloud service providers often host a small collection of open datasets; the most notable one is AWS's Open Data (https://oreil.ly/DZ5uV).
8. ML frameworks often have small pre-built datasets that you can load while using the framework, such as TensorFlow datasets (https://oreil.ly/HMJX_).
9. Some evaluation harness tools host evaluation benchmark datasets that are suff-ciently large for PEFT finetuning. For example, Eleuther AI's Im-evaluation-harness (https://github.com/EleutherAl/m-evaluation-harness) hosts 400+ benchmark datasets, averaging 2,000+ examples per dataset.

10. The Stanford Large Network Dataset Collection (hts://oreilye_B) is a great

repository for graph datasets.

Free Datasets

https://www.v7labs.com/blog/best-free-datasets-for-machine-learning

UI Modelling - annotated UIs

https://uimodeling.github.io/

Datasets - Roboflow (like HuggingFace for data)

https://public.roboflow.com/

Dataset - Anthropic use of AI by job category 

https://huggingface.co/datasets/Anthropic/EconomicIndex

-

MoE (Mixture of Experts)

LLM Mixture of Experts Explained (tensorops.ai)

An Introduction to Vision-Language Modeling

Vision AI - VLMs and CNNs
- Vision Language Models
- Convolutional Neural Networks

https://arxiv.org/pdf/2405.17247

Multi-modal LLMs

Cambrian-1: A Fully Open, Vision-Centric Exploration of Multimodal LLMs (cambrian-mllm.github.io)

Movies: rated [for collaborative filtering]

MovieLens Latest Datasets | GroupLens

The Society of Mind

Book by Marvin Minsky

AI Engineering[Book] [O'Reilly] [Zeki recommends]

same author: Designing Machine Learning Systems- DMLS focuses on building applications on top of traditional ML models, which involves more tabular data annotations, feature engineering, and model training