The purpose of this post is to give you fundamental insights for reinforcement learning (RL) adoption in LLM’s training – such as, “what is RL for ?” and “how RL matters in LLM’s training ?” – with theoretical background.
Custom Metrics for Evaluation in LLM (Azure AI Studio)
In this post, I’ll show you how to use your own custom metrics (both statistical code metrics and LLM-measured prompt metrics) working with Azure AI Studio.
Improve accuracy for vision tasks with OCR assisted in multimodal models
Unfortunately the quality of text extraction in multimodal models (such as, GPT-4v, GPT-4o, Claude, etc) is not as well as today’s state-of-the-art (SOTA) OCR model. This post introduce how to improve text extraction quality in these models with the help of OCR models.
Improve AI Safety in LLM Apps – Prompt example, Red teaming, etc
In this post, I’ll briefly show you risk mitigation architecture for adversarial prompting and prompt’s example for safety.
Implement Model Parallelism in LLMs
In this post, I will break down the techniques for scaling of large model’s training in a step-by-step manner.
Implement Advanced Reasoning in Semantic Kernel
In this blog post, I’ll show you how to implement custom Planner for advanced reasoning in Semantic Kernel. Reasoning will be a key to create an intelligent autonomous agent.
Image Processing in LLMs – TaskMatrix (Visual ChatGPT)
Visual ChatGPT (TaskMatrix) is one of interesting examples, in which visual information can be generated or replaced by interacting with OpenAI ChatGPT (LLM). This post shows you how it’s built on ReAct chain and reasoning.
ReAct (Reason+Act) prompting in LLMs
ReAct (Reasoning + Acting) is a flexible LLM chain framework and essential for today’s advanced LLM reasoning. LangChain helps you compose ReAct framework.
This post will give you the answer for the questions: “What is ReAct?”, “How ReAct works?”, and “How to build ReAct?”.
Hugging Face Fine-tune for Multilingual Question Answering (Japanese Example)
In this post, I’ll show you multilingual (Japanese) example for question-answering in Hugging Face.
This also describes how to configure practical QA system with the fine-tuned extractive QA models.
Hugging Face Fine-tune for Multilingual Summarization (Japanese Example)
In this post, I’ll show you multilingual (Japanese) example for text summarization in Hugging Face.
You can learn how to fine-tune multilingual transformer models in sequence-to-sequence tasks.
tsmatz