Model-Training Micro-Courses
A structured course on LoRA, QLoRA, distillation, and RLHF, built from the training-the-delta field note, with one free sample module you can try right now.
Free preview, no payment flow — this is a waitlist, not a checkout
This page is a preview of a planned course, not a live paid product: there is no payment processing here. Module 1 is fully built and free to try, including AI-generated feedback on your own explanation. Joining the waitlist below stores your email so the full course can be announced when it ships; it does not charge you anything or promise a ship date.
How it works
The curriculum below expands the training mechanics covered in the training-the-delta field note ↗ into a four-module course. Module 1 (LoRA) is fully built as a free sample: read the real explanation, then write your own in the box below and a Foundry prompt agent — a supportive tutor, not an exam grader — checks your understanding against a concept rubric and tells you what to revisit.
Try it
Free sample — Module 1
LoRA: adapting a model without touching its weights
Full fine-tuning updates every parameter in a model, which is expensive to store and
serve for every task variant you need. LoRA freezes the base model entirely and learns
only a pair of small low-rank matrices, A and B, inserted
alongside a target weight matrix. The rank r — often as small as 8 or 16 —
controls how much capacity that update has. At inference, the learned A·B
update can be merged directly into the frozen weights, so there is no extra latency
compared to the original model. The result: a small fraction of the trainable
parameters and optimizer memory of full fine-tuning, with most of the task performance.
Module 2
QLoRA: fitting a bigger base model into less memory
Adds 4-bit quantization of the frozen base model plus paged optimizers, so a single GPU can fine-tune a model that would otherwise need several — and what you trade away to get there.
Module 3
Distillation and model merging
Teach a small model from a large one's outputs, and combine several fine-tuned checkpoints into one model without further training — and when each technique earns its complexity.
Module 4
RLHF, DPO, and Constitutional AI
How preference data turns into an aligned model, through the reward-model path, the simpler direct-preference path, and the AI-feedback path.
Try the free sample: explain LoRA in your own words
No wrong answers here — this is practice, not a test.
Write a few sentences to begin.
Join the waitlist for modules 2-4
Get notified when the rest of the course ships. No spam, no payment now.
Why this exists
This product extends the BETA-MIND project record's commitment to evidence-led, rigorous educational design, and directly reuses the mechanics already researched and cited in the training-the-delta field note. Rather than fabricate a checkout flow this site cannot honestly support, the free sample module and waitlist are the genuinely working parts of this preview.
Production checklist
- No fabricated payment flow — a free sample module and an honest waitlist, clearly labeled.
- Tutor agent grades against a fixed concept rubric and is instructed to be supportive, not harsh.
- Waitlist emails are hashed for the storage key and independently rate-limited against spam.
- Dedicated Foundry prompt agent, versioned and isolated from the portfolio chatbot.
- Automated backend tests cover the waitlist, understanding-check, and idempotent resubmission.