Notes on Mastering LLMs Maven course

• https://maven.com/parlance-labs/fine-tuning

Lecture 1

Most common tools for fine-tuning:

• Axolotl, huggingface’s transformer apply_chat() method — formats your fine-tune examples into templates that the LLM knows how to use
  • https://github.com/OpenAccess-AI-Collective/axolotl
  • https://huggingface.co/docs/transformers/main/en/chat_templating
• Many issues in development are due to templating inconsistencies (e.g. the special token that you put between prompt & response in the fine-tuning examples)
• (May 2024 statement) With larger context windows, bigger LLMs are able to take in more examples in the prompt itself, which makes fine-tuning less important
  • Is this true?
• You should prove to yourself that you need to fine-tune — and only do it so after you’ve tried using the base model, doing prompt eng, etc.
• A lot of reasons to fine-tune are...

Someone on one of my Philippine-based teams is going to be studying in the US for a quarter in 2024, and I offered to share some cultural advice/reflections to her prior to her departure; I figured it might be a fun post to open-source & share more widely, so here goes…

In no particular order — and being super explicit, even if some of this is obvious — as that’s kinda the whole point of this post:

• In American culture, it’s common to ask “How are you?” (e.g. when walking past each other on the street or seeing each other for the first time in the morning). If you respond with more than 3-5 words, the American will likely to start to feel uncomfortable, as they likely don’t really care how you are feeling. (That is, unless they ask a follow-up question to your initial response.) “How are you?” or questions about the weather are basically the American equivalent of saying “good morning”...

When I started getting excited about data science 7 years ago, I was also at the same time just learning how to program. Like…not how to program k-means from scratch, like how to draw circles in Javascript.

As such, I was learning multiple technical concepts at the same time:

Since then, I’ve collected some wins under my belt as a data scientist. But my best learnings actually came from making mistakes – writing inefficient Hive queries, making low-accuracy or over-engineered models, being overly academic about metric definitions, waiting too long to show business users intermediate progress, etc.

One of the biggest conceptual errors I made starting out in data science was thinking that data analysis was somehow a different, special, disjoint field from software engineering. I’m taking a few hours to write up – and draw up! – my past and current beliefs on this topic.

If you’re...

I love incentives. I enjoy thinking about what motivates people, what motivates myself, and, in general, how people make decisions – large and small. A special case in the study of incentives is game theory.

Many things that are not (literally) games – love, international relations, negotiations – can actually have their underlying mechanics modelled, and often accurately predicted, within a game-theoretic framework. Game theory is especially helpful to model situations in which individuals’ incentives are not atomic – that is, they interact with one another – which is quite often the case in the real world.

A fundamental building block of the field is a payoff matrix. A payoff matrix is where an individual’s decisions are mapped to “payoffs,” or return values, for an individual taking a certain decision in a moment in the game. For example, if you have $100 at the start of time, and...

Amidst the dramatic events of the pandemic, I’ve had one constant source of unexpected joy: the productivity, personal growth, and freedom I’ve discovered from working from home.¹ I wanted to write up my experience to put all my thoughts on it in one place.

Compared to working in an office, working from home feels like being liberated from a million nuisances and inconveniences of the day-to-day grind. I have so much more flexibility and control over my time now that I find it difficult to quantify. Since high school, my most productive times have always been late at night, when all is still and quiet in the house, and I’m able to deeply think about things uninterrupted for hours at a time.

I still remember fondly staying up late to write (and re-write) my college application essays, trying to fuse together ordinary public school experiences into a unique tapestry that a distant...

I recently read Y Combinator co-founder Jessica Livingston’s post on how to Grow The Puzzle Around You. It describes in beautiful detail her unique, often behind-the-scenes contributions to the enormously successful startup generator, and her journey to becoming a non-traditional technology investor.

In the conclusion, it implores the reader to “write down” your truly unique qualities – “and don’t edit [them].” That is, ignore the temptation to add positive/negative labels to each, and instead focus on the rock-bottom truth in the features themselves.

Because, what’s important is understanding yourself, so you’re then in a better position to leverage your unique strengths for the benefit of others. The overall goal is to "puzzle fit around you” as the title says, but to do that, in my reading, you first must have a solid understanding of yourself at the metaphorical center.

So, here...

Recently, I started having crippling lower back pain. When I woke up in the morning, I would often struggle to stand fully upright. If I were to walk for longer than a few blocks, I’d need to rest from the tightness/total exhaustion in my lower back.

After consulting with various doctors, I came to the conclusion that the simplest explanation was the best – through quarantine, I had ballooned to 106 kg (234 lbs). While I have an athletic build, I’m only 6 foot (183cm) and, from a typical college weight of 80 kg (180 lbs), this was quite a shock.

Over the past 2 months, I have started intermittent fasting (IF).¹ That is, I only eat from noon to around 8pm. The appeal to IF over other diet regimes to me has always been twofold:

1. The simplicity – the rules are so clear: eat at these times, don’t eat at these others. It doesn’t require a change in diet, which has downstream requirements...