Fine-Tuning and Open-Weights Bootcamp

Live time each week on Zoom: a 90-minute class where the instructor reviews real submissions, a 60-minute squad lab, 60 minutes of office hours, and your own 30-minute 1-1. Nothing is lectured live; the recorded lessons do that.

1. 1

   Open weights

   Weeks 1 to 2

   Week 1

   Open weights

   • The open model landscape
   • Running and serving open weights
   • Benchmarks and baselines
   • Licences

   You build: Baseline eval

   Week 2

   Project 1: the baseline

   • Reproducibility
   • Choosing the task

   You build: Reproducible baseline of three models

   Project 1, the gate at week 2

   Baseline

   Pick the right open model for a structured-extraction task.

   You hand in

   • Eval harness
   • Three baselines
   • Report

   The gate

   Reproducible from the repo.

2. 2

   Data

   Weeks 3 to 4

   Week 3

   Data

   • Building a training set
   • Synthetic data
   • Filtering and deduplication
   • Provenance and licensing

   You build: Curated dataset

   Week 4

   Project 2: the data

   • Provenance per example
   • Documentation

   You build: Documented training set

   Project 2, the gate at week 4

   Training data

   Build the data that will make the model better.

   You hand in

   • Curated set
   • Synthetic augmentation
   • Provenance and licence per example

   The gate

   Every example has a source and a licence.

3. 3

   Supervised fine-tuning

   Weeks 5 to 6

   Week 5

   Supervised fine-tuning

   • LoRA and full fine-tuning
   • Training loops and checkpoints
   • Evals during training
   • Overfitting

   You build: SFT run; squads form

   Week 6

   Project 3: SFT

   • Held-out evaluation
   • Deploying a checkpoint

   You build: A fine-tuned model that beats the baseline

   Project 3, the gate at week 6

   SFT

   Beat the baseline.

   You hand in

   • LoRA fine-tune
   • Checkpoints
   • Training evals
   • Deployed endpoint

   The gate

   The gain measured on a held-out set.

4. 4

   Preference tuning

   Weeks 7 to 8

   Week 7

   Preference tuning

   • DPO and its relatives
   • Reward signals
   • Safety tuning

   You build: DPO run

   Week 8

   Project 4: preference

   • Preference sets
   • Win-rate evals

   You build: Preference-tuned model with a win-rate eval

   Project 4, the gate at week 8

   Preference tuning

   Make it prefer the right answers.

   You hand in

   • DPO run
   • Preference set
   • Win-rate eval

   The gate

   Win rate against the SFT model measured.

5. 5

   Deployment

   Weeks 9 to 10

   Week 9

   Deployment

   • Quantisation and serving
   • Cost per token versus a hosted API
   • Monitoring a custom model

   You build: Serving plan

   Week 10

   Project 5 and the viva

   • The model in production
   • Defending the data and the cost

   You build: Your model in production with monitoring and a cost comparison; the viva

   Project 5, the gate at week 10

   Production

   Serve it cheaper than the API.

   You hand in

   • Quantised model
   • Serving
   • Monitoring
   • Cost comparison

   The gate

   The viva: defend the model, its data and its cost.

6. 6

   Employer brief and hiring sprint

   Weeks 11 to 12

   Week 11

   Employer brief

   • A real problem from a hiring partner, in squads
   • A partner's model problem
   • Working to someone else's definition of done

   You build: The employer brief in progress

   Week 12

   Hiring sprint

   • CV and portfolio built from graded work
   • Applications and follow-ups in the hiring plan
   • Mock interviews scored against the role
   • Demo day

   You build: Project 6 delivered; demo day

   Project 6, the gate at week 12

   Employer brief

   A partner's model problem.

   You hand in

   • The brief delivered
   • Demo day

   The gate

   The brief's owner accepts the result.