qwen2.5-3b-feynman-explainer

"You don't understand something unless you can explain it to your grandmother." — Feynman

A LoRA fine-tune of Qwen2.5-3B-Instruct trained to explain anything the way Richard Feynman did — from the ground up, with vivid analogies, no jargon until it's earned, and pure flowing prose.

Built by Shabul Abdul, Sr. Data Scientist.
Trained on Apple M5 MacBook Pro · 24 GB unified memory · Apple MLX.

---

Try it right here — no download needed

The difference is immediate. Same model weights, same question. Just a style shift.

---

"How does gradient descent work?"

Base model (Qwen2.5-3B-Instruct, no fine-tuning):

Gradient descent is an optimization algorithm used in machine learning to minimize a loss function. It works by computing the gradient of the loss with respect to the model parameters and updating the parameters in the opposite direction of the gradient, scaled by a learning rate. This process is repeated iteratively until convergence...

This model (Feynman fine-tuned):

Imagine you're blindfolded somewhere on a hilly landscape, and your only goal is to reach the lowest valley. You can't see anything — but you can feel the slope of the ground under your feet.

Here's what you do: take a small step in whichever direction feels most downhill. Then stop. Feel the slope again. Take another step downhill. Repeat.

That's gradient descent. The "gradient" is just the slope — which direction is up, how steep it is. You always step opposite to it, because you're trying to go down. The "learning rate" is how big each step is. Too big and you overshoot the valley and bounce around the hillside forever. Too small and you're shuffling along for eternity.

The key insight: you never need to see the whole landscape. You only ever need to know the slope exactly where you're standing.

---

"What is entropy?"

Base model:

Entropy is a thermodynamic quantity representing the amount of disorder or randomness in a system. In information theory, entropy measures the average amount of information or uncertainty in a random variable. Higher entropy indicates greater disorder...

This model:

Here's something that bothered physicists for a long time: heat always flows from hot things to cold things. Never the other way. You've never seen a warm cup of coffee spontaneously get hotter while the air around it gets colder. But why not? The laws of physics don't actually forbid it.

The answer is entropy, and it's not really about disorder — that's a metaphor that confuses more than it helps. It's about counting.

Take a room with all the air molecules crammed in one corner. That's one arrangement. Now consider all the arrangements where molecules are spread out evenly — there are astronomically more of them. So when molecules move randomly, they almost certainly end up spread out, simply because there are so many more ways to be spread out than bunched up.

Entropy is just a measure of how many arrangements look like what you've got. High entropy means lots of arrangements produce this same picture. The reason entropy always increases is brutally simple: there are more ways to be disordered than ordered, so randomness always drifts toward disorder.

---

"What is a p-value?"

Base model:

A p-value is the probability of obtaining results at least as extreme as the observed results, assuming the null hypothesis is true. If the p-value is below the significance threshold (typically 0.05), the result is considered statistically significant...

This model:

Scientists get this wrong all the time, so let's go slowly.

You've run an experiment. You gave half the patients a drug and half a sugar pill, and the drug group got better faster. Great — but here's the question you should be asking: could this difference have shown up just by chance, even if the drug does absolutely nothing?

The p-value answers that question. It says: if the drug had zero effect, what's the probability I'd still see a difference this big, just from the randomness of who got sick and who got better?

If that probability is tiny — say, 2% — we say p = 0.02. That's suspicious. It's hard to believe chance alone did this.

Here's what p-value does NOT mean, and this is where everyone goes wrong: it is not the probability that the drug works. It's not even close to that. It's only the probability of your data, assuming the drug doesn't work.

The difference matters enormously. One is a statement about the drug. The other is a statement about your data.

---

Training details

Base model	Qwen/Qwen2.5-3B-Instruct
Method	LoRA (rank 16, alpha 32)
Dataset	shabul/feynman-explainer-dataset · 575 synthetic prompts, chat-formatted into 517 train + 58 validation rows
Hardware	Apple M5 MacBook Pro · 24 GB unified memory
Framework	mlx-lm (Apple MLX)
Iterations	1,500 steps
Learning rate	2e-4

Full loss curve and throughput stats will be added after training.

The full training dataset is published here: shabul/feynman-explainer-dataset

---

How to run

from mlx_lm import load, generate

model, tokenizer = load("shabul/qwen2.5-3b-feynman-explainer")

question = "Why does ice float on water?"
messages = [
    {"role": "system", "content": (
        "You are a Feynman-style explainer. Build intuition from the ground up "
        "using concrete analogies. No jargon until it's earned. Flowing prose only."
    )},
    {"role": "user", "content": question},
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
print(generate(model, tokenizer, prompt=prompt, max_tokens=400))

---

Why this works

Style transfer via LoRA is a different beast from knowledge fine-tuning. The base model already knows what gradient descent is. We're teaching it how to talk about it — the rhythm, the analogy-first structure, the short declarative sentences, the moment of "here's where most people get confused."

Rank 16 (vs. rank 8 for a knowledge fine-tune) gives the adapter enough capacity to shift the generative distribution meaningfully. Higher learning rate (2e-4) pushes the style harder in fewer steps.

---

Author

Shabul Abdul — Sr. Data Scientist
huggingface.co/shabul

---

No cloud GPUs. No PhD required. Just a MacBook and a good idea.