# Site

## Pages

- [Page not found](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/404/): That page doesn't exist, or it has moved since you last looked
- [Assessment](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/assessments/): Every piece of graded work in the course, what it is worth, and when it is due.
- [Evaluation Protocol](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/assessments/evaluation-protocol/): The shared evaluation and marking mechanics used by Project 1, Project 2, and Project 3 — public evaluation layers, prohibited uses, separation validation, perplexity calculation, human evaluation, and technical gates.
- [Project 1 — Build a Narrative Base Model](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/assessments/project-1/): Train a decoder-only language model from scratch that can continue an incomplete story with recognisably narrative language.
- [Project 2 — Give the Model a Voice](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/assessments/project-2/): Post-train the frozen Project 1 checkpoint to produce stories with a recognisable target narrative style, and prove the change without hiding a regression.
- [Project 3 — Build a Specialist](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/assessments/project-3/): Choose Track A (narrative specialisation) or Track B (arithmetic reasoning), fine-tune an appropriate starting checkpoint, and demonstrate the acquired behaviour through controlled evaluation.
- [Lectures](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/lectures/): What is covered, week by week, and which topics each lecture draws on.
- [What a Language Model Learns](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/lectures/week-01/): Probability over text, next-token prediction, maximum likelihood and cross-entropy, n-gram intuition, and perplexity as a first, limited piece of evidence
- [Inside a Decoder-Only Transformer](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/lectures/week-02/): Encoder, encoder-decoder, and causal decoder-only attention; tracing tensor shapes through a decoder block; architecture as a 32M parameter allocation
- [Scale, Data, and Optimisation Under Fixed Compute](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/lectures/week-03/): Power-law intuition, Kaplan- and Chinchilla-style compute allocation, data quality versus data quantity, and reading a training pilot honestly
- [Evaluate, Decode, and Package a Base Model](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/lectures/week-04/): Hold-out and contamination, checkpoint selection, decoding as an experimental variable, and reproducible submission packaging
- [From a Base Model to a Target Behaviour](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/lectures/week-05/): Capability versus elicited behaviour, continued pre-training versus SFT versus preference learning, and a measurable narrative-voice target
- [Supervised Fine-Tuning and Data That Teaches Behaviour](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/lectures/week-06/): Response-only loss masking, instruction schemas, synthetic demonstration data, and learning without catastrophic overwriting
- [Learning from Preferences Without Hiding the Cost](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/lectures/week-07/): Bradley–Terry preference data, reward models and RLHF, RLAIF and DPO, and proportionate preference learning at 32M parameters
- [Did the Behaviour Change, and What Regressed?](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/lectures/week-08/): Multi-objective post-training evaluation, blind A/B review, paraphrase robustness, and checkpoint selection against a pre-declared priority
- [Reasoning as Generated and Verifiable Behaviour](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/lectures/week-09/): Chain-of-thought as generated intermediate tokens, self-consistency, exact-answer verification, and rationale faithfulness
- [Train a Specialist That Can Be Tested](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/lectures/week-10/): Task contracts, frozen starting checkpoints, split integrity, and full versus parameter-efficient adaptation for Project 3
- [Generalisation, Supervision, and One Honest Ablation](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/lectures/week-11/): Template and paraphrase shift, outcome versus process supervision, verifiable rewards as frontier context, and a matched ablation design
- [Audit the Training System](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/lectures/week-12/): Reconstructing the artifact chain, reproducing the core metric, what transfers across scale, and the final engineering retrospective
- [People](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/people/): Who teaches the course, and how to reach them.
- [Eli Morgan](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/people/eli-morgan/): Eli protects the fairness and operability of the constrained-training environment, from compute accounting to platform incidents.
- [Maya Rao](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/people/maya-rao/): Maya runs the practical labs and model clinics, with a focus on PyTorch, dataset pipelines, training diagnostics, checkpoint loading, and reproducible evaluation.
- [Dr Yiwei Easton](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/people/yiwei-easton/): Yiwei teaches language-model training at the intersection of machine learning, empirical evaluation, and resource-aware systems.
- [Policies and support](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/policies/): The course-wide rules for compute, data, integrity, submission, extensions, accessibility, safety, and privacy that apply to all three projects.
- [Project 1 Launch and Data Validation](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/sessions/01-project-1-launch/): Verify the supported environment and data checksums, and write a bounded Project 1 target before selecting an architecture
- [Lab 1 — Spend a 32M Parameter Budget](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/sessions/02-lab-1-parameter-budget/): Trace shapes through a causal decoder block and compare two legal architectures under the 32M/33.6M parameter boundary
- [Project 1 Pilot Clinic](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/sessions/03-project-1-pilot-clinic/): An evidence-led drop-in clinic for diagnosing a training pilot before committing to the principal Project 1 run
- [Project 1 Evaluation and Packaging](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/sessions/04-project-1-evaluation-and-packaging/): Run the released evaluation pipeline, verify a fresh-process load, and reconcile the report, ledger, and model package before submission
- [Project 2 Target and Baseline Planning](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/sessions/05-project-2-target-and-baseline/): Define a target voice in observable dimensions, freeze the unchanged baseline, and cost one viable post-training route
- [Lab 2 — Build and Inspect a Post-Training Batch](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/sessions/06-lab-2-post-training-batch/): Render a response-only loss mask, run a bounded SFT update, and test whether an apparent behaviour survives prompt paraphrase
- [Project 2 Behaviour Clinic](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/sessions/07-project-2-behaviour-clinic/): An evidence-led drop-in clinic for interpreting target-versus-retention evidence and difficult preference labels
- [Project 2 Blind Evaluation and Packaging](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/sessions/08-project-2-blind-evaluation/): Compare the unchanged and tuned checkpoints under blinded, matched conditions and package a reproducible Project 2 submission
- [Project 3 Task Contract and Baseline](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/sessions/09-project-3-task-contract/): Choose a permitted track and bounded target behaviour, write a valid task contract, and establish the unchanged starting-model baseline
- [Lab 3 — Train and Verify a Specialist](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/sessions/10-lab-3-specialist-verifier/): Turn a task contract into valid training and evaluation records, build a verifier before training, and test one shifted slice and one retained capability
- [Project 3 Generalisation Clinic](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/sessions/11-project-3-generalisation-clinic/): An evidence-led drop-in clinic for locating the generalisation boundary and separating supported conclusions from alternative explanations
- [Project 3 Fresh-Load Audit and Submission](https://comp4020-agentic-coding-studio.github.io/comp4020-ass2-Easton-Yi/sessions/12-project-3-fresh-load-audit/): Reproduce the final specialist in a clean environment and submit one internally consistent evidence chain