English:Features, labels and data dictionaries - AI Year 5
Introduction

A table can look tidy while hiding a missing value, a copied source or the answer itself. Investigate twelve invented component-card records, build a clear data dictionary and distinguish features from the target. Keep the source, units and uncertainties visible before anyone claims to train a useful model.
Define features labels and a data dictionary
Separate what you observe from what you want to predict
A feature describes an example. A label is the target category for the task you are trying to learn or predict. A data dictionary explains what each field means, which values are allowed, what units are used and how unknown values are recorded. Without those definitions, a neat-looking table can be misleading.
Worked leaf example: a table might describe leaf colour and edge type as features, species as the target label, and Width_cm as a measured width in centimetres. A row's green colour is not itself a verified species name. Edge type and width may help describe a leaf, but these few fields do not guarantee a correct botanical identification. A real species label needs an appropriate reliable source; this lesson does not identify or collect actual plants.
For a new leaf whose species is unknown, you cannot supply its correct species label as an input and then claim to predict that same answer. During supervised training, labels are available to guide fitting; during a new prediction, the target answer is the thing you do not yet know. A field can have a different role in another task, so state the task before deciding which columns are features.
Use an explicit fictional card task
Our main dataset has twelve invented rows about imagined printed component cards. Each card illustrates a mouse, keyboard or monitor. We want to distinguish the component category using permitted descriptive features. No physical cards were measured and no model has been trained. The rows were created for this teaching worksheet, version V1; they are not observations of pupils or items in the Commons diagram.
The diagram shows real kinds of computer components, including a monitor, keyboard and mouse. It can help you understand the category words. It does not supply our invented card widths, row IDs or labels. Measuring the picture on a phone screen would produce a display-size measurement, not the recorded width of our imagined printed card.
The main features are colour, shape and Width_cm. Row_id identifies a record; card_id connects repeated records to an imagined source card. Component_label is the target. Keeping source metadata is useful for checking duplicates, but a row identifier is not automatically a sensible predictive feature. In particular, repeated copies of the same card do not create new independent examples.
Read the complete dictionary before the table
| Field | Role in this task | Allowed values or form | Meaning and units | Missing/conflict rule |
|---|---|---|---|---|
| row_id | Record metadata | Text; L01 to L12 | No measurement unit; unique table-row identifier | Required; not a prediction feature |
| card_id | Source metadata | Text identifying the same imagined source card | No measurement unit; repeat means revisit the same source | Required; not a new independent example just because row_id differs |
| colour | Feature | black, white or grey | Dominant colour of the imagined printed illustration | MISSING means not known; do not guess |
| shape | Feature | rounded or rectangular | Outline category of the imagined printed illustration | MISSING means not known; do not infer function from outline alone |
| Width_cm | Feature | Number from 0.5 to 20 inclusive | Centimetres across the printed illustration, not the real device | MISSING is unknown, not zero; companion JSON uses null |
| component_label | Target label | mouse, keyboard or monitor | The category this fictional prediction task seeks | A conflict needs source review; it is not an input for predicting itself |
| answer_tag | Proposed leaking field | A copy of component_label | Not part of the original six-column table | Exclude from prediction inputs; the answer would be unavailable on a new unlabelled card |
The teaching width range is about imagined printed illustrations, not a rule that all real keyboards or monitors are between half a centimetre and twenty centimetres wide. A value outside that range is flagged against this particular worksheet definition; a different legitimate dataset might have a different range. Do not invent a universal natural law from one schema.
The actual V1 table has six columns. Answer_tag is shown in the dictionary as a proposed extra field so you can detect why adding it would leak the answer. It is not secretly a seventh measured feature in the original table.
Audit twelve rows and preserve uncertainty
Inspect the twelve-row source, including its deliberate faults
The row IDs are unique, but source-card IDs can recur. MISSING is the visible marker for an unknown width; it corresponds to null in the retained companion data. Keep the original version while recording findings rather than silently rewriting it to look perfect.
| row_id | card_id | colour | shape | Width_cm | component_label |
|---|---|---|---|---|---|
| L01 | C01 | black | rounded | 4 | mouse |
| L02 | C02 | grey | rectangular | 3 | keyboard |
| L03 | C03 | white | rectangular | 5 | monitor |
| L04 | C04 | white | rounded | 4 | mouse |
| L05 | C05 | black | rectangular | MISSING | keyboard |
| L06 | C06 | black | rectangular | 6 | monitor |
| L07 | C07 | white | rectangular | 2 | keyboard |
| L08 | C08 | grey | rounded | 0.5 | mouse |
| L09 | C02 | grey | rectangular | 3 | mouse |
| L10 | C10 | grey | rectangular | 20 | monitor |
| L11 | C11 | black | rounded | 7 | mouse |
| L12 | C03 | white | rectangular | 5 | monitor |
L05 has no width. L03 and L12 describe the same C03 source card with identical features and label: in our worksheet history, L12 is an accidental copy. L02 and L09 both refer to C02 with the same features but disagree on the target label. This is a source-label conflict, not proof that one label is correct because it appears more often elsewhere in the table.
There are twelve rows but ten distinct source-card IDs. Eleven rows have a non-missing width, including repeated source cards. The raw target counts are five mouse, three keyboard and four monitor. Those are counts in a deliberately flawed invented table, not reliable population proportions or a representative sample of real equipment.
Handle a missing value without inventing a measurement
MISSING does not mean zero. Replacing L05's unknown width with zero would claim a number that was never supplied and would fail our positive-width definition. Mark it unknown and ask for an authorised source check. If a simple width-only calculation excludes that row, state explicitly that eleven non-missing row values were used and that repeated cards still affect the result.
Do not delete the original source record merely to hide the problem. In a different task, an analyst may document a justified estimation method or use a method designed for missing values. Neither is what this worksheet has done. Here we flag the gap and preserve uncertainty; an automatically filled value would need a separately documented decision and evaluation.
Distinguish a duplicate from a genuine second example
Compare row_id and card_id. L03 and L12 have different row IDs but the same source C03. Their repeated colour, shape, width and target are a duplicate copy in the supplied history. Keep the source/version record; in a reviewed analysis copy, mark or remove the accidental extra record according to the stated policy. Do not count it as another independently observed card.
Repeated observations can be legitimate in another study, but that intention must be documented. Equal feature values alone do not prove that two different real objects are duplicates. Conversely, changing only a row ID does not make a copied source independent. If train and test sets contain copies of the same source, a score can give a misleading impression of success on genuinely new examples.
Review conflicting labels rather than voting for an answer
C02 is named in L02 and L09. One says keyboard and the other mouse. Compare the authorised original source or ask the responsible reviewer; our invented table alone does not tell you which version should be accepted. Do not change the minority label just to agree with the majority of all rows.
Two different cards can share colour, shape and width while depicting different components. That alone is not necessarily a labelling error: these features may be insufficient to distinguish the categories. The conflict we flag is stronger evidence because the same supplied card_id is said to refer to one source card. Source identity and task definitions matter, not just matching feature strings.
CodeAI: AI: Training Data & Bias. With teacher support, connect the explanation to data choice, labels and coverage. The older-grade source supplies context, not a requirement to complete its whole unit. Our twelve-card example is a separate original exercise, not the video's dataset or proof that its small sample can train a fair real system.
Find leakage test boundaries and transfer
Find a field that gives the answer away
Suppose someone adds answer_tag by copying component_label into it on every V1 row. A pretend predictor could copy answer_tag back into its answer and match twelve out of twelve stored labels. That includes the conflicted source record. It has not learned useful visual recognition or established the correct label for a new card.
This is label leakage for our task: the target answer has entered the proposed inputs even though it would not be known on a genuinely unlabelled card. Keep the target column for training/review, but exclude it and its answer-copy field from prediction inputs. Check when and how every field becomes available. A perfect worksheet score obtained from the answer column is not evidence of generalisation.
Test unit and range boundaries explicitly
Width_cm means centimetres across the imagined printed illustration. Forty millimetres is four centimetres, not forty centimetres. Record the original unit and an authorised conversion before comparing a value with a centimetre range. A number without its meaning and unit is incomplete evidence.
Our inclusive boundary accepts 0.5 and 20; it flags 0.49 and 20.01. A missing value is handled separately from a small value. These checks enforce the declared worksheet definition, not an automatically correct scientific range. Do not silently clip a value to the nearest boundary or replace unknown with zero so that validation passes.
Transfer the checks to a different domain
Return to the leaf example. Colour and edge type are descriptive features; species is the new target; Width_cm now measures a leaf, not an illustration of a computer part. The same-looking field name and unit do not make the meaning identical. Before combining tables, check what was measured, the permitted categories, source method and missing-value convention.
The useful checks transfer: define fields, retain provenance, flag unknown values, investigate duplicates and contradictory source labels, and keep the target separate from prediction-time inputs. The allowed values and measurement meaning do not transfer unchanged. A width range suitable for a printed card need not fit a leaf study. Correctly reading our table does not certify botanical identification.
Explain why more rows are not automatically better
Copying V1 ten times creates one hundred and twenty row records but still only ten distinct original card IDs, with the same source conflict and missing measurement repeated. It has not collected a hundred and twenty independent observations. Quantity alone does not establish accuracy, coverage, suitability or fairness.
Ask which examples are represented, which contexts are absent and whether the intended inputs would actually be available at prediction time. The made-up colours and outlines are for data literacy; they do not describe every real device or every leaf. No trained model or real-world performance estimate is produced here.
Keep a short quality and evidence note
Identify three features and the target, retain a complete dictionary with units/allowed values, and record the twelve-row source version. Flag the L05 missing value, the L03/L12 duplicate and the L02/L09 conflict. Explain the answer_tag leak and demonstrate an inclusive boundary check. Add a limitation, an untested future case and a proportionate human-review route.
Try the activities before reading the worked checks. Correct mistakes and explain what evidence would settle an unresolved source question. Use only invented or approved non-personal data, not pupil names, photos or private histories. This AI Year 5 lesson is editorial enrichment linked to England's KS2 data-evaluation expectations; a completed quiz does not prove model training, practical data expertise or a real deployment's safety.
Interactive Tasks
Quiz: Test Your Knowledge
Which field is the target of the component-card prediction task? (component_label) (!colour) (!shape) (!Width_cm)
What does Width_cm measure in the twelve-row task? (The width of an imagined printed illustration in centimetres) (!A real monitor size automatically measured from the browser) (!A pupil age in years) (!The number of training sessions)
Which row has a deliberately missing width? (L05) (!L01) (!L03) (!L10)
Which pair is the supplied accidental copy of source C03? (L03 and L12) (!L01 and L04) (!L07 and L10) (!L05 and L08)
What should happen to the conflicting labels for C02? (Check the original source or responsible reviewer.) (!Automatically choose whichever label is most common in all rows.) (!Assume equal features always force one real category.) (!Delete the conflict without retaining a record.)
Why is copying component_label into answer_tag a problem for prediction inputs? (It gives away the target answer unavailable on a new unlabelled card.) (!It proves the data are representative.) (!It collects new independent examples.) (!It measures leaf width automatically.)
Which pair is accepted by the declared inclusive width bounds? (0.5 and 20) (!0.49 and 20.01) (!Zero and a missing value) (!Forty and one hundred)
How many distinct source-card IDs are in V1? (Ten) (!Twelve) (!Eleven) (!Twenty)
How many centimetres are forty millimetres? (Four) (!Forty) (!Four hundred) (!Zero)
What changes when V1 is copied ten times? (Row count increases but original source diversity does not.) (!There are automatically one hundred and twenty independent observations.) (!The unknown width becomes a real measurement.) (!The C02 conflict is automatically resolved.)
Memory Game
| Feature | A permitted description of an example |
| Target label | The category this task seeks to predict |
| Data dictionary | Definitions of fields units allowed values and missing conventions |
| Provenance | Where data came from and how they were created |
| Missing value | Information not supplied rather than an invented zero |
| Label leakage | The target answer improperly enters prediction inputs |
Drag and Drop
| A width with no unit | Check the field definition before comparing numbers |
| A source card copied into another row | Do not count it as a new independent observation |
| Two labels for the same source card | Retain the conflict and seek authorised source review |
| A field copied from the target answer | Exclude the leak from prediction inputs |
| A new domain with the same column name | Check the measurement meaning and allowed values again |
Crossword Puzzle
| FEATURE | What describes an example for the stated prediction task? |
| LABEL | What target category is sought in the stated task? |
| MISSING | What term describes an unsupplied value rather than zero? |
| SOURCE | What should you check to investigate a record’s provenance? |
| UNITS | What must accompany a measured number? |
| LEAKAGE | What occurs when the answer improperly enters prediction inputs? |
LearningApps
Optional extra practice. Choose an appropriate activity with your teacher. If the external collection is unavailable, use the native activities and independent tasks in this course.
Cloze Text
Open Tasks
Easy
- Name features and label: Identify colour, shape, Width_cm and component_label with their different roles.
- Read the missing marker: Explain what MISSING means for L05 without inventing a number.
- Check a unit: Convert forty millimetres to centimetres and state what was measured.
- Find a repeated source: Compare row IDs with C03 source-card IDs.
Standard
- Write a complete dictionary: Include field roles, units, allowed values, source identity and missing/conflict conventions.
- Audit twelve rows: Flag the gap, accidental copy and contradictory source labels with row evidence.
- Identify the leaking field: Explain why answer_tag can fake a perfect stored-label score.
- Keep an evidence note: Retain the original version, quality findings, boundary test, limitation, untested case and human fallback.
Advanced
- Transfer to the leaf task: Change target, allowed feature categories and measurement meaning rather than just renaming columns.
- Explain ambiguous feature sets: Show why two different objects can have equal listed features without necessarily being mislabelled.
- Evaluate dataset size: Compare one copy with ten copies, distinguishing records from source diversity.
- Design a prediction-time check: State which inputs are available before knowing the target and what remains to verify.
Learning Check
- Required leaf worked case and field roles: Separate features, target, units and metadata.
- Twelve-row audit: Identify missing, duplicate and conflicting-source cases.
- Dictionary and unit boundary: Explain the declared width meaning and inclusive range.
- Leakage and the false perfect score: Show why the proposed answer_tag is not a useful input.
- Duplicate size and feature ambiguity: Explain two different ways a table can mislead.
- Transfer and evidence limits: Apply the quality process to a new task.
Worked checks and assessment guidance
- In a leaf table, colour and edge type can be features, species the target, and Width_cm a leaf width in centimetres. In our component-card task, colour, shape and printed-illustration width are features, component_label the target, and row/card IDs record provenance. These are different tasks; the same column name alone does not establish the same measurement meaning.
- L05 lacks width. L03 and L12 are the supplied accidental copy of C03. L02 and L09 both refer to C02 but disagree between keyboard and mouse; seek the original source rather than voting from other rows. There are 12 rows,10 distinct card IDs and 11 non-missing row widths. Raw targets are 5 mouse,3 keyboard,4 monitor, not reliable population frequencies.
- Width_cm describes imagined printed illustrations, not real device dimensions measured from a screen. Forty millimetres converts to 4 centimetres. The declared bounds accept 0.5 and 20 and flag 0.49 and 20.01. MISSING/null is not zero; preserve the gap and source version rather than clipping or guessing a value.
- Answer_tag is a copy of the target on each stored row. Copying it back matches all 12 stored labels, including the unresolved source conflict, without recognising new cards. The answer would not be known on a new unlabelled card. Keep labels for training/review while excluding the target and its copied field from prediction inputs.
- Ten copies of V1have 120 rows but the same 10 original card IDs and repeated faults. Equal feature values on different source cards do not automatically prove duplication or a bad label; the description may lack distinguishing information. Our C02case is flagged because its source identity is the same and labels conflict.
- Keep field definitions, provenance, missing-value handling, duplicate/conflict investigation and leakage checks. Redefine the actual target, allowed values and measurement units/meaning. Retain a source version, boundary result, limitation, genuinely untested case and human-review route. Synthetic worksheet completion is not a trained model, botanical identification or a real-world performance guarantee.
Learning evidence
Retain the 12-row source version, a dictionary with feature/target/metadata roles, units, allowed values and missing/conflict rules. Explain L05,L03/L12and L02/L09with evidence; identify answer_tag leakage; demonstrate a unit or inclusive-boundary check; and transfer definitions to the leaf task. State a limitation, an untested future case and a human-review route. No real pupil data, physical measurements, trained model or biological identification is claimed.
OER and Wikipedia
DfE: KS2 Computing supplies the data-evaluation context; this separate AI Year 5 route is editorial. Google ML Crash Course: data scrubbing and Google: monitoring and label leakage were consulted for data-quality and target-leakage concepts. Our child-appropriate cases, worksheet and dictionary are original; adult source exercises are not copied. Reviewed 14 November 2026.
Personal computer, exploded.svg, Gustavb, CC BY 2.5. Unchanged component illustration; it supports category vocabulary, not the invented widths or row labels. The proposed leaf-chart candidate was not resolved natively; the leaf worked example remains as text and this course uses the explicitly defined component-card dataset.
https://www.youtube.com/watch?v=x2mRoFNm22g — CodeAI: AI: Training Data & Bias. Original publisher recording embedded unchanged, used with teacher-supported vocabulary.
Linked learning areas
aiMOOC Projects
Schulfach+


aiMOOCs


aiMOOC Projekte


MOOCwiki · Deutsch
Nach dem Lernen ist vor dem Lernen
Entdecke direkt den nächsten Lernkurs. Weitere Inhalte erscheinen, wenn Du weiter nach unten scrollst.
Zur MOOCwiki-HauptseiteMediathek
Mediathek wird aus dem Wiki geladen ...
Keine passenden Inhalte gefunden. Bitte ändere Suche oder Filter.
NEWSLernweltNOAH fragen