Spaces:
Sleeping
Sleeping
feat: add new slides on LLM training paradigms and further reading resources
Browse files
content/slides/ch01-what-is-ai.tsx
CHANGED
|
@@ -394,6 +394,13 @@ export const ch01: Chapter = {
|
|
| 394 |
detection head β which is why it converges fast on the small
|
| 395 |
datasets we collect at competitions.
|
| 396 |
</p>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 397 |
</div>
|
| 398 |
),
|
| 399 |
viz: <FoundationModels />,
|
|
|
|
| 394 |
detection head β which is why it converges fast on the small
|
| 395 |
datasets we collect at competitions.
|
| 396 |
</p>
|
| 397 |
+
<Callout label="Further reading" tone="accent">
|
| 398 |
+
<a className="underline" href="https://huggingface.co/blog/samuellimabraz/peft-methods" target="_blank" rel="noreferrer">
|
| 399 |
+
PEFT: Parameter-Efficient Fine-Tuning Methods
|
| 400 |
+
</a>
|
| 401 |
+
{" β "}LoRA, adapters, prefix-tuning and friends: how to fine-tune
|
| 402 |
+
a foundation model with < 1 % of its parameters.
|
| 403 |
+
</Callout>
|
| 404 |
</div>
|
| 405 |
),
|
| 406 |
viz: <FoundationModels />,
|
content/slides/ch03-paradigms.tsx
CHANGED
|
@@ -195,6 +195,56 @@ export const ch03: Chapter = {
|
|
| 195 |
),
|
| 196 |
viz: <AgentEnvLoop />,
|
| 197 |
},
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 198 |
{
|
| 199 |
id: "ch03-05",
|
| 200 |
title: "Where each fits in Black Bee",
|
|
|
|
| 195 |
),
|
| 196 |
viz: <AgentEnvLoop />,
|
| 197 |
},
|
| 198 |
+
{
|
| 199 |
+
id: "ch03-llm",
|
| 200 |
+
title: "Modern LLMs use all three",
|
| 201 |
+
eyebrow: "Pretrain Β· fine-tune Β· align",
|
| 202 |
+
layout: "split",
|
| 203 |
+
content: (
|
| 204 |
+
<div className="space-y-4">
|
| 205 |
+
<p>
|
| 206 |
+
The training pipeline behind GPT-4, LLaMA, and Gemini composes the
|
| 207 |
+
paradigms we just saw, in this order:
|
| 208 |
+
</p>
|
| 209 |
+
<ol className="space-y-2 text-[14px] text-ink/85">
|
| 210 |
+
<li>
|
| 211 |
+
<strong>1. Self-supervised pretraining</strong> β predict the
|
| 212 |
+
next token on trillions of tokens of unlabelled text. Produces a
|
| 213 |
+
general-purpose base model.
|
| 214 |
+
</li>
|
| 215 |
+
<li>
|
| 216 |
+
<strong>2. Supervised fine-tuning (SFT)</strong> β train on a
|
| 217 |
+
much smaller set of high-quality instruction / answer pairs.
|
| 218 |
+
Teaches the format.
|
| 219 |
+
</li>
|
| 220 |
+
<li>
|
| 221 |
+
<strong>3. RLHF</strong> β reinforcement learning from human
|
| 222 |
+
feedback. A reward model learns from human preferences; the
|
| 223 |
+
policy is then optimised against it (PPO or DPO).
|
| 224 |
+
</li>
|
| 225 |
+
</ol>
|
| 226 |
+
<p className="text-muted">
|
| 227 |
+
Yann LeCun's 2016 "cake" analogy survives in this
|
| 228 |
+
form: most of the calories come from unsupervised pretraining;
|
| 229 |
+
SFT and RLHF are the icing and the cherry.
|
| 230 |
+
</p>
|
| 231 |
+
</div>
|
| 232 |
+
),
|
| 233 |
+
viz: (
|
| 234 |
+
<figure className="mx-auto flex w-full max-w-[760px] flex-col items-center">
|
| 235 |
+
<div className="w-full overflow-hidden rounded-md border border-stroke bg-surface">
|
| 236 |
+
<img
|
| 237 |
+
src="/figures/llm-paradigms.jpg"
|
| 238 |
+
alt="Unsupervised pretraining, supervised fine-tuning, and RLHF, drawn as a monstrous creature with a cherry on top."
|
| 239 |
+
className="block w-full"
|
| 240 |
+
/>
|
| 241 |
+
</div>
|
| 242 |
+
<figcaption className="mt-3 text-center font-mono text-[11px] uppercase tracking-[0.12em] text-muted">
|
| 243 |
+
modern reimagining of LeCun's "cake" analogy
|
| 244 |
+
</figcaption>
|
| 245 |
+
</figure>
|
| 246 |
+
),
|
| 247 |
+
},
|
| 248 |
{
|
| 249 |
id: "ch03-05",
|
| 250 |
title: "Where each fits in Black Bee",
|
content/slides/ch05-deep-learning.tsx
CHANGED
|
@@ -244,6 +244,15 @@ export const ch05: Chapter = {
|
|
| 244 |
differentiate it automatically β autograd. We only have to write
|
| 245 |
the forward pass.
|
| 246 |
</p>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 247 |
</div>
|
| 248 |
),
|
| 249 |
viz: <BackpropChain />,
|
|
|
|
| 244 |
differentiate it automatically β autograd. We only have to write
|
| 245 |
the forward pass.
|
| 246 |
</p>
|
| 247 |
+
<Callout label="Implemented from scratch">
|
| 248 |
+
<a className="underline" href="https://github.com/samuellimabraz/cafedl" target="_blank" rel="noreferrer">
|
| 249 |
+
cafedl
|
| 250 |
+
</a>{" "}
|
| 251 |
+
β a small deep-learning library I wrote in Java to internalise
|
| 252 |
+
backprop, layer composition, and OOP design patterns. Useful if
|
| 253 |
+
you want to see every line of an autograd engine before reaching
|
| 254 |
+
for PyTorch.
|
| 255 |
+
</Callout>
|
| 256 |
</div>
|
| 257 |
),
|
| 258 |
viz: <BackpropChain />,
|
content/slides/ch11-nectar.tsx
CHANGED
|
@@ -246,6 +246,16 @@ ObjectDetectionEvaluator(detector.model, cfg).evaluate()`}
|
|
| 246 |
</tbody>
|
| 247 |
</table>
|
| 248 |
</div>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 249 |
</div>
|
| 250 |
),
|
| 251 |
},
|
|
|
|
| 246 |
</tbody>
|
| 247 |
</table>
|
| 248 |
</div>
|
| 249 |
+
<Callout label="Worked example, published" tone="accent">
|
| 250 |
+
For a full, public end-to-end detection project that follows the
|
| 251 |
+
same pipeline (data β architecture comparison β training β
|
| 252 |
+
evaluation β deployment), see{" "}
|
| 253 |
+
<a className="underline" href="https://huggingface.co/blog/samuellimabraz/signature-detection-model" target="_blank" rel="noreferrer">
|
| 254 |
+
Open-Source Handwritten Signature Detection Model
|
| 255 |
+
</a>{" "}
|
| 256 |
+
β YOLO vs DETR vs RF-DETR with real numbers and an open model on
|
| 257 |
+
the Hub.
|
| 258 |
+
</Callout>
|
| 259 |
</div>
|
| 260 |
),
|
| 261 |
},
|
content/slides/ch12-closing.tsx
CHANGED
|
@@ -36,6 +36,37 @@ export const ch12: Chapter = {
|
|
| 36 |
<li>Β· COCO mAP definition (cocodataset.org)</li>
|
| 37 |
</ul>
|
| 38 |
</div>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 39 |
</div>
|
| 40 |
),
|
| 41 |
},
|
|
|
|
| 36 |
<li>Β· COCO mAP definition (cocodataset.org)</li>
|
| 37 |
</ul>
|
| 38 |
</div>
|
| 39 |
+
<div className="md:col-span-2">
|
| 40 |
+
<div className="mb-3 font-mono text-[11px] uppercase tracking-[0.12em] text-muted">
|
| 41 |
+
Worked examples and side projects
|
| 42 |
+
</div>
|
| 43 |
+
<ul className="space-y-2 text-[15px]">
|
| 44 |
+
<li>
|
| 45 |
+
Β·{" "}
|
| 46 |
+
<a className="underline" href="https://huggingface.co/blog/samuellimabraz/signature-detection-model" target="_blank" rel="noreferrer">
|
| 47 |
+
Signature detection blog
|
| 48 |
+
</a>{" "}
|
| 49 |
+
β full detection pipeline, YOLO vs DETR vs RF-DETR, with
|
| 50 |
+
published model and dataset.
|
| 51 |
+
</li>
|
| 52 |
+
<li>
|
| 53 |
+
Β·{" "}
|
| 54 |
+
<a className="underline" href="https://huggingface.co/blog/samuellimabraz/peft-methods" target="_blank" rel="noreferrer">
|
| 55 |
+
PEFT methods
|
| 56 |
+
</a>{" "}
|
| 57 |
+
β LoRA, adapters, prefix-tuning: fine-tuning foundation models
|
| 58 |
+
cheaply.
|
| 59 |
+
</li>
|
| 60 |
+
<li>
|
| 61 |
+
Β·{" "}
|
| 62 |
+
<a className="underline" href="https://github.com/samuellimabraz/cafedl" target="_blank" rel="noreferrer">
|
| 63 |
+
cafedl
|
| 64 |
+
</a>{" "}
|
| 65 |
+
β Java deep-learning library written from scratch; an autograd
|
| 66 |
+
engine you can read end-to-end.
|
| 67 |
+
</li>
|
| 68 |
+
</ul>
|
| 69 |
+
</div>
|
| 70 |
</div>
|
| 71 |
),
|
| 72 |
},
|
public/figures/llm-paradigms.jpg
ADDED
|
Git LFS Details
|