Indie Degree
GitHubSign in
← Programme

Fine-tuning and Post-training

AIE-205

3 credits · 60h required · after AIE-204, AIE-102

Where a base model becomes a product, and where the evaluation discipline from AIE-102 stops being optional. The single question this course exists to answer honestly is the one most fine-tuning write-ups avoid: did it beat a well-prompted baseline on the same eval set? Frequently it does not, and a course that cannot produce that answer has taught a technique rather than a judgement.

The lab. Career Side Quests' evidence extraction, because the eval already exists. You built a 30-case gold set for it in AIE-102 and the harness that runs it in tools/eval-harness — which means the measurement predates the model. That is the correct order and almost never how it happens in practice, and having it means every claim in this course is checkable rather than asserted.

By the end you can

  • Curate a fine-tuning dataset and defend every inclusion rule
  • Run LoRA fine-tuning and measure whether it actually helped
  • Explain preference optimisation and implement DPO
  • Detect and report regression on capabilities you did not intend to change

What this builds

Next assessment. Inclusion rules, written before the data — 180 min, tier 1. Machine-checkable, so you can attempt it before reading anything and find out whether you need to.

0/17 assessed items demonstrated · 0/13 read or watched. The second number is how you close a gap; only the first is evidence that it closed.

0 of 17 demonstrated · 0/13 read

0m of 60h

M1 · The dataset is the model

0/6 · 13h

  • lecture · 3h · tier 0 self-marked

    Deep Dive into LLMs like ChatGPT — Andrej Karpathy

  • reading · 2h 30m · tier 0 self-marked

    Training language models to follow instructions with human feedback

  • assignment · 3h · tier 1 machine-verified

  • assignment · 2h 30m · tier 2 panel-assessed

  • reading · 1h 45m · tier 0 self-marked

    AI Engineering

  • retention · 15m · tier 1 machine-verified

M2 · Supervised fine-tuning, and the baseline that must beat it

0/6 · 12h

M3 · Parameter-efficient methods

0/6 · 12h

  • reading · 2h · tier 0 self-marked

    LoRA: Low-Rank Adaptation of Large Language Models

  • reading · 2h · tier 0 self-marked

    huggingface/peft

  • assignment · 3h · tier 1 machine-verified

  • assignment · 2h 45m · tier 2 panel-assessed

  • reading · 2h · tier 0 self-marked

    huggingface/transformers

  • retention · 15m · tier 1 machine-verified

M4 · Preference optimisation

0/6 · 12h

M5 · Did it help, and what did it cost

0/6 · 11h

  • assignment · 3h · tier 2 panel-assessed

  • project · 3h · tier 3 artifact

  • assignment · 2h · tier 2 panel-assessed

  • defense · 1h · tier 4 defended

  • lecture · 1h 45m · tier 0 self-marked

    Intro to Large Language Models — Andrej Karpathy

  • retention · 15m · tier 1 machine-verified