#fine-tuning
73 episodes
#5485: Fine-Tuning Parakeet for Hebrew and Your Own Jargon
NVIDIA's Parakeet beats Whisper on Android — but can you teach it Hebrew, or just your own jargon? Two answers, one much happier.
#5443: Heads vs Layers: How Model Merging Actually Works
Heads aren't the Lego bricks of model merging — layers are. Here's what heads really do and how frankenmerges get built.
#5436: Small Models as Rewriters, Not Writers
Why "don't say X" prompts backfire, and how a tiny grammar-constrained model can scrub a script without breaking its grammar.
#5431: The Other Half of Hugging Face: Why BERT Still Out-Downloads Llama
Encoder models pull over a billion downloads a month. Decoder models pull 397 million. The AI conversation and the download counter are describing ...
#5412: Editing vs. Note-Taking for AI Fine-Tunes
Hand-editing a model's output gives three training signals at once. Writing notes gives one — and a weaker one at that.
#5411: Fine-Tuning at 4-Bit vs 16-Bit: What It Really Costs
QLoRA cuts fine-tuning VRAM 15x and cost up to 85% — but you pay in training time, quality, and safety alignment.
#5410: Adapters: 102KB That Reshapes a 403GB Model
A 102KB adapter file changes how a 403GB base model behaves — without ever merging into it. Here's how model adapters actually work.
#5408: When Small NLP Models Beat the LLM
Feature extraction, fill-mask, token classification — the classic NLP tasks still have a job. Here's when a small model beats a frontier API.
#5407: Hemmingway-1 and the War on Waffle
A 27B model promises answers without the preamble. Its benchmark is homegrown — and the behavior it targets has a paper trail.
#5396: Teaching a Small Model to Stop Spelling Out Numbers
Your ASR pipeline is fine until someone dictates "three point two" and gets "three point two" spelled out. Here's how inverse text normalization ac...
#5393: Fine-Tuning a Model on 100 Hand-Edited Answers
You don't need 10,000 examples to make a model sound like you. The real number is closer to 100 — if the edits are opinionated.
#5245: Training a Model on 100 Phone Photos
One drawer of screws, a hundred phone photos, and a schema. When does a small custom model beat a fine-tuned VLM?
#5169: Fine-Tuning vs From-Scratch for Minor Languages
What 774 experiments and a new Armenian model reveal about the tradeoff between fluency and knowledge in low-resource languages.
#5116: Beyond Common Crawl: The Datasets Shaping AI
Common Crawl is just the bulk flour. We explore the curated datasets that actually shape what AI models know and do.
#4971: Fine-Tune vs Prompt: Taming Hallucinated Deadlines
Big models add fake deadlines. Small models miss tasks. Which wrongness can you build a system around?
#4969: How to Train a One-Class Detector from 300 Photos
From photo folder to production YOLO model — the real walkthrough for custom object detection.
#4809: Can a Podcast Train Itself?
Can a podcast become a self-developing project? We explore RAG vs. fine-tuning for lore memory.
#4723: Are Model Families Actually Different Models?
Claude Opus, Sonnet, and Haiku aren't trim levels — they're different models sharing a brand name.
#4057: How to Fix AI's Bullet Point Addiction
Why AI models default to bullet points and how textual LoRAs achieve 94% prose adherence.
#3596: Why an AI Model Kept Calling Itself Sonnet 4.6
When a Chinese model insists it's "Sonnet 4.6," is it theft, sloppy training, or something stranger?
#3406: LoRA Isn’t Just for Image Generation
LoRA lets you fine-tune an LLM’s behavior with a 50MB file. Here’s how it works and why it matters.
#3283: Fine-Tuning DeepSeek for One Podcast
Can a purpose-specific fine-tune fix a model's stubborn writing tics? We explore the practical engineering behind it.
#3171: How to Break an LLM's Bad Verbal Habits
Blacklists fail and regex inverts meaning. Here's what actually works to clean up AI writing tics.
#3170: Pharmacokinetics vs Neural Nets: Two Meanings of "Model
Two things called "models" that work completely differently — and why the confusion matters for patient safety.