Week 8: Did the Behaviour Change, and What Regressed?

Multi-objective post-training evaluation, blind A/B review, paraphrase robustness, and checkpoint selection against a pre-declared priority

19 April 2027

Stage: Shape model behaviour · Project 2 due

Summary#

Post-training creates a multi-objective evaluation problem. A checkpoint may match target vocabulary while becoming repetitive, follow an exact instruction while failing a paraphrase, or improve style ratings while losing general narrative coherence. This lecture shows how to define target-style, instruction-compliance, narrative-quality, memorisation, and retention measures without collapsing them into one opaque score. Students practise blind paired comparisons, recognise evaluator and prompt-order effects, inspect source overlap, and select checkpoints against a declared priority rather than whichever metric looks best. The resulting report should say what changed, what did not, and which evidence cannot distinguish competing explanations.

Learning outcomes#

After this lecture, you should be able to:

1. operationalise a behavioural target using observable anchors and held-out evidence;
2. design a blind A/B review that separates model identity from candidate quality;
3. evaluate paraphrase robustness, copying, and retained narrative ability;
4. select a checkpoint across competing objectives using a pre-declared rule; and
5. write a calibrated claim that distinguishes measurement, interpretation, and uncertainty.

1. Separate target behaviour from surface mimicry#

Target-word frequency can rise without a convincing voice. Human anchors therefore cover coherence, target behaviour, and major defects, while target-domain PPL provides a complementary distributional measure. Selected examples must include failures and cannot be silently edited.

2. Blind comparison and rater variability#

Candidate order is randomised and model identities are hidden. Three raters score official story outputs independently under the published anchors; medians reduce sensitivity to one extreme score, and adjudication addresses large disagreement. Student-run formative reviews should record prompt, sampler, rater count, and order policy.

3. Robustness, retention, and copying#

The same intent is tested through paraphrased prompts or changed story openings. General narrative prompts reveal catastrophic forgetting. Exact and near-duplicate checks plus phrase-level inspection distinguish plausible genre convention from suspicious reproduction of source text.

4. Multi-objective checkpoint selection#

Students identify one primary target and minimum acceptable floors for retained behaviour and technical validity. A checkpoint that wins one metric but violates a floor is not selected. The report presents the trade-off openly instead of manufacturing a single composite score after seeing results.

Required reading#

• Jeffrey Zhou et al., Instruction-Following Evaluation for Large Language Models, abstract and Sections 1–2.
• Course page, Human review protocol and rating anchors.

Optional reading#

• Percy Liang et al., Holistic Evaluation of Language Models, Sections 1–2.

This week’s action#

Deadline: Project 2, Sunday 25 April 2027 at 23:59 AET. Preserve the unchanged-checkpoint baseline outputs alongside the final outputs.