Project 2 — Give the Model a Voice

Post-train the frozen Project 1 checkpoint to produce stories with a recognisable target narrative style, and prove the change without hiding a regression.

Due: 25 April 2027 · Weight: 50%

Page summary#

Starting from either your frozen Project 1 submission or the course narrative fallback checkpoint, post-train the model to produce stories with a recognisable target narrative style. The supplied target corpora are processed, verified public-domain editions of Grimm’s Fairy Tales and One Thousand and One Nights. You may use one or both, or justify a permitted supplement or replacement.

“Post-training” is used broadly here: any training performed after Project 1 to change the model’s domain, style, or conditioned behaviour. You may choose target-domain continued pre-training, supervised instruction tuning, preference-based optimisation, or a justified combination. The method is open; the target behaviour and evaluation obligation are not.

Key facts#

What you must demonstrate#

Define the target style in operational terms, use the supplied corpora or justify a permitted replacement, compare the post-trained model with the unchanged starting checkpoint, and evaluate target adaptation alongside narrative quality and retention. The project rewards the engineering fit between method, data, budget, and evidence — not the complexity or novelty of the post-training algorithm.

If you use instruction tuning, document the schema, special tokens, and loss masking. If you use preference optimisation, first establish an SFT or otherwise appropriate baseline and demonstrate why the extra pipeline and compute were justified. Work without genuine human preference labels must not be described as full RLHF.

Required investigation#

• define the target style in operational rather than purely impressionistic terms
• construct or curate suitable post-training data
• compare the post-trained model with its unchanged starting checkpoint
• separate held-out target-style texts at story or source level
• evaluate both target-style adaptation and narrative quality
• test for regressions in the original narrative capability
• analyse the effect of important training and sampling choices
• explain why the final method was selected over feasible alternatives

Marking breakdown#

Engineering report — 35 marks#

1. Target behaviour and method rationale — 5 marks. A testable definition of the intended voice or style, success criteria beyond a vague resemblance claim, and a reasoned choice among continued pre-training, SFT, preference optimisation, or a combination.

2. Data design and provenance — 5 marks. Permitted and well-documented source editions, story/source-level splits, appropriate cleaning and formatting, token and source statistics, deduplication or overlap checks, and a defensible balance between target-style data and retained narrative ability.

3. Implementation and reproducibility — 5 marks. A correct transition from the starting checkpoint, a reproducible post-training pipeline, traceable tokenizer/chat-template/loss-mask behaviour, declared hyperparameters and seeds, and a final package matching the reported method.

4. Experimental design and budget judgement — 7 marks. Meaningful baseline comparisons, controlled ablations or pilots, justified training duration and checkpoint selection, and evidence of prioritising experiments under the compute deadline.

5. Multi-objective evaluation and regression analysis — 8 marks. Held-out target-style language modelling, blind or clearly structured behavioural comparison, narrative coherence/completeness assessment, repetition and source-overlap checks, and evaluation of the original Project 1 capability.

6. Communication, limitations, and engineering reflection — 5 marks. A coherent report making the decision trail visible: initial expectation, evidence, revision, final choice, limitations, and what the project reveals about post-training under constraints.

Submitted model — 15 marks#

Submission validity is a technical gate. A valid model receives up to 10 marks from the shared tutor-evaluation PPL rule and up to 5 marks from blinded task-specific review. See the Evaluation Protocol for the complete public evaluation layers, prohibited uses, separation validator, perplexity formula, human-evaluation process, and technical gates that apply to this component.

7. Tutor-evaluation perplexity — 10 marks. The shared formula applied to continuation-only reference-token-normalised PPL across all ten tutor prompts: P ≤ 25 earns all 10 marks; 25 < P < 50 receives an exponential decay of exp(-0.1 × (P − 25)); P ≥ 50 earns zero for this criterion.

8. Target behaviour, narrative quality, and retention — 5 marks. Whether readers can consistently recognise the operationally defined target behaviour across varied openings while the output remains fluent, connected, and narratively usable. Task-specific dimension weights: recognisable target behaviour 35%; narrative coherence and usability 30%; robustness across varied openings 20%; retention and integrity, including unacceptable regression or copying 15%. A single showcase, memorised passage, or strong style signal attached to an incoherent story cannot earn high marks.

Individual work#

This project is submitted individually. AI assistance is permitted when declared in the required AI Assistance Statement; responsibility for the submitted work remains with you. See Policies for the full individual-work, collaboration, and AI-assistance rules.