King's College London · FACULTY OF ARTS & HUMANITIES

7AAVCD42 Chap.4 Data Lifecycles, Post-Training and Safety

- one subject, every graph, every model, every mark
5 Chapters2-page Bible
Our own words - no uploaded lecturer files
Updated for this semester
Chapter 4 of 10 · 7AAVCD42

Data Lifecycles, Post-Training and Safety

Failures can begin before modelling

The module follows data through collection, cleaning, labelling, crowdsourcing and review, then examines human feedback, fine-tuning and safety rules after initial training. This chapter maps that lifecycle and locates the first unresolved compromise. It treats data labour, copyright, consent and model steering as connected decisions rather than separate ethical add-ons.

Use the method to distinguish what post-training can contain from missing coverage or weak evidence that requires new data and evaluation.

Map the complete lifecycle

Begin with collection purpose and population. Record what was gathered, whose activity produced it and which contexts are missing. Move to cleaning and labelling, where categories, instructions and disagreement shape the usable dataset.

Follow the handover into training and evaluation. Limits must travel with the data; a polished file without documentation can make an early compromise invisible to the next team.

Deployment is not the endpoint. Complaints, overrides and observed failures can provide evidence for repair. Ask who records that feedback and who has authority to change data, model behaviour or use rules.

Monitoring without a return path produces information while the same problem continues.

Locate a data cascade

A cascade occurs when an unresolved compromise propagates. Missing representation can distort evaluation. Ambiguous labels can become apparent model inconsistency. A deployment team may interpret a benchmark as proof of fitness for a different setting.

Identify the earliest repairable stage and explain why later adjustments cannot fully compensate.

Do not assume that more data is always the solution. New material can reproduce the same categories or consent problems.

Specify which population, condition or disagreement must be represented, how quality will be checked and which outcome will show improvement.

Keep labour visible

Data is produced through human work. Labellers, moderators, reviewers and domain experts receive instructions, face time pressure and resolve ambiguity.

Crowdsourcing can distribute tasks widely while preventing any worker from understanding the complete purpose. Explain how labour organisation affects consistency, documentation and responsibility.

Copyright and consent questions enter at collection. State what material is used, the claimed authority and whether people have a practical choice.

Keep legal analysis distinct from ethical evaluation: an objection about fairness does not by itself establish the legal status of every use.

Evaluate post-training and safety

Human feedback, fine-tuning and behavioural rules can steer a trained model. They may reduce known harmful outputs or define refusal behaviour. They cannot manufacture evidence about a population missing from training.

Decide whether the problem calls for containment, new data, redesigned evaluation or delayed deployment.

Safety choices encode values. Identify who decides which requests should be refused, what harm is prioritised and how users challenge an error. Finish a lifecycle analysis by assigning the repair to a data supplier, model team, deploying institution or regulator.

A named actor and decision make the recommendation operational.

In this chapter

What this chapter covers

  • 01

    Data Collection

  • 02

    Data Labelling

  • 03

    Data Cascades

  • 04

    Post-Training

  • 05

    AI Safety

Worked example · free

Trace a data shortcut

Q [7 marks]. A team reuses a cheap dataset for a high-stakes service. Explain the main risks. The marks shown are an AskSia study allocation, not an official King's marking scheme.
Data cascade.Data cascade. The narrowing layers show how an unresolved collection or labelling compromise can survive training and reach deployment.collection and labellingtraining and post-trainingdeployment outcome
Data cascade. The narrowing layers show how an unresolved collection or labelling compromise can survive training and reach deployment.
  • 2Test fitness for purpose.
  • 3Inspect labels and handover.
  • 2Evaluate post-training limits.
The audit should compare represented groups and contexts, examine how labels were produced and whether limits travelled with the data. Post-training cannot create coverage that is absent from the dataset.
Sia tip — Locate the first compromise before examining the final score.
Glossary

Key terms

Data Cascade
Data Cascade is an early data compromise that propagates through later stages and produces downstream failure.
Data Labour
Data Labour comprises the human collection, labelling, review and moderation work supporting an AI system.
Post-Training
Post-Training adjusts model behaviour through feedback, fine-tuning and rules after initial training.
FAQ

Data Lifecycles, Post-Training and Safety FAQ

What is a data cascade?

It is the propagation of an unresolved collection, labelling or handover problem into training, evaluation and deployment. The concept directs attention to the earliest repairable stage.

Can post-training fix missing representation?

Post-training can steer or contain known behaviours, but it cannot supply absent populations or contexts. Missing coverage requires new evidence, revised evaluation and possibly delayed deployment.

Why should data labour enter technical analysis?

Time, instructions, disagreement and working conditions affect label consistency and documentation. The organisation of labour can therefore change the technical meaning and reliability of the dataset.

Study strategy

Assessment move

Draw the lifecycle, place one actor and one failure at every stage, then explain which repair requires new data and which could be addressed through post-training. Add a return arrow from deployment complaints to the team able to revise the system. Choose a data-cascade case and identify the first unresolved compromise, the handover that concealed it and the downstream symptom.

Next, write separate notes on representation, label quality, worker conditions and consent so one broad bias claim does not replace the mechanism. Test a proposed safety rule by asking what behaviour it can contain and what missing evidence it cannot create. End the exercise by assigning each repair to a supplier, model team, procuring institution or regulator.

Compare a technical patch with a governance intervention for the same failure. State which symptom each addresses, what remains unresolved and how the organisation would know whether the repair worked. Redraw the handover points using the documentation each receiving team actually needs.

A+Everything unlocked
Unlocks this Bible + all 18 of your subjects at King's College London, plus 1,000+ more Bibles across the AskSia library.
Sia - your 7AAVCD42 tutor, unlimited, worked the way the exam marks it
The full 2-page Bible + practice bank with worked solutions
Chrome extension - sync your LMS so Sia knows your deadlines
Bilingual EN / Chinese on every Bible and every Sia answer
$0.99 Trial
30-day money-back · cancel in one tap · how it works