Our Mission¶

The LLM FOMO and Our Answer To It !!¶

TLDR

Keeping pace with the rapid advancements in Large Language Models (LLMs) can be a daunting task. This project aims to address this "Fear Of Missing Out" (FOMO) by organizing information through a comprehensive mind map.

The LLM Boom and Knowledge Explosion

Since OpenAI's release of ChatGPT, the LLM field has undergone explosive growth. The research community is constantly innovating, with new ideas emerging on a daily basis. This rapid pace can be overwhelming, with several new LLMs and associated research areas like prompting techniques, fine-tuning methods, evaluation benchmarks, and context window solutions appearing frequently. While exciting for the AI community, it can also lead to a sense of FOMO.

Traditionally, keeping up with every new method can be impractical. We propose a more efficient approach: a hierarchical mind map.

Organizing the LLM Landscape with a Mind Map

Despite the constant stream of new methods, they often fall under existing categories. By organizing this information hierarchically, understanding new methods and their purpose becomes significantly easier.

For example, RLHF gained popularity after ChatGPT. Recently other terms such PPO, DPO, and ORPO have also emerged. Can you guess their purpose?

The mind map below illustrates these methods as advancements upon each other for fine-tuning LLMs to human preferences.

LLM
 └── Training_LLM
     └── Supervised_Fine-tuning
          └── Fine-tuning_LLMs_to_Human_Preferences
               ├── RLHF
               ├── DDPG
               ├── PPO
               └── ORPO

Previously seemingly random names can be instantly understandable when organized hierarchically. Here, all methods represent advancements in fine-tuning LLMs based on human preference.

Our Goal: A Continuously Updated LLM Knowledge Base

Our objective is to curate and maintain an up-to-date mind map of emerging LLM techniques. This allows you to encounter a new concept (through a LinkedIn post or tweet) and instantly grasp its essence within the context of existing techniques in the mind map.

Note

As maintainers, we ensure the accuracy and freshness of the mind map and session outline.

Teaching is Learning Twice: We need your support!!¶

TLDR

We've combined relevant topics from the mind map and tuned them into multi part sessions. We welcome your contributions to create detailed articles and videos for each part.

Organizing new research through a mind map is just the first step. True understanding requires deeper exploration. We believe that teaching a concept reinforces learning it. Here's where you come in.

On the repository's homepage, you'll find related topics from the mind map organized into session series, each potentially divided into multiple parts. We aim to publish technical blog posts on the website and explanatory videos on YouTube channel for each part, and we highly encourage your contributions.

How You Benefit ?

Let's say you choose to write an article on "Semantic Caching for LLMs" for session 5, part 3. This process involves in-depth independent learning followed by article creation.
The article will be added to the repository and published on the website. Additionally, we'll host a live seminar where you present the article. The live stream will be recorded on YouTube for future reference.
By participating in this learn-and-publish cycle, you can continuously update your knowledge with the latest llm work.

How Others Benefit ?

Imagine a larger community contributing in this way. Collectively, these smaller contributions can cover a vast range of LLM topics in one central location. This will establish a single, authoritative source for understanding the vast LLM landscape.

Practitioners can leverage this knowledge base when building LLM-based tools. For instance, an engineer considering local LLM deployment can find clear outlines of available options and detailed explanations within relevant sessions. This collaborative approach saves everyone valuable time.
Researchers can use the mind map to track progress within specific areas and identify potential areas for improvement in the hierarchy, such as developing novel prompting techniques that complement existing methods.

If our vision aligns with yours, join us! Pick a topic that interests you and become part of this generous effort.

Please follow the contribution guide for more details.