扒哥吃瓜

Throughout the process to create the UK Standard Skills Classification (SSC), we relied heavily on the use of Artificial Intelligence (AI), particularly Large Language Models (LLMs). These were mostly OpenAI models (and generally the best available model at the time) but during early development we also used open-source Llama models for some of the more data intensive tasks that would have been too costly using OpenAI. At each stage, we manually reviewed exceptions and inspected the outputs from the听AI听models, sometimes having to check large amounts of data. During the development of SSC Version 1.0, we investigated the performance of GPT-5.4. Evaluation of importance score differences between GPT-5.4 outputs and the GPT-4.1-derived prototype outputs for a sample of mappings showed a noticeably more accurate performance by GPT-5.4. This led to the decision to regenerate all primary and secondary mappings using GPT-5.4 for SSC Version 1.0.

A central part of our听AI听approach was the use of text embeddings. These are numerical representations (vectors) of text that allow computers to understand the semantic meaning of a piece of text. Text embeddings are used in clustering text and to compare meanings of text strings. To understand how related 2 text strings are to each other, we calculate the distance between the vectors using cosine similarity. The larger the score, the closer the two text strings are in semantic meaning.

This method of comparing text strings was used extensively throughout the project. Early in the project, we experimented with different embeddings models and decided on the use of OpenAI 3-Large (OAL3) embeddings in the main. For the development of SSC Version 1.0, we re-checked the performance of OAL3 against newer embedding models (e.g. Qwen3) but it remained the best performing for our specific needs.

AI听was also used in other ways in the project. We used听AI听prompts at many stages of the project, for example to quality assure skill and task statements and to detect inconsistencies and errors in mappings. Designing and refining prompts often involved several iterations as we learnt the best ways to interact with the听AI听to get the desired results.

The steps below outline the main stages in the creation of the听SSC.

Tasks

Figure 8: Development of UK SSC Occupational Tasks

Figure 8 outlines the development process of the听SSC听Occupational Task library, detailing the main input libraries used, the data cleaning steps, and the validation against other information sources.

This is displayed as a series of processing steps from T1 to T6 in a row across the top of the diagram with each step shown below in a flow diagram. On the left-hand side are the 4 input libraries which feed into the first processing step: T1.

The 鈥楾鈥� prefix for each step ID (such as T1) relates to 鈥楾ask鈥�. For similar processes shared in later sections an 鈥楽鈥� prefix relates to 鈥楽kills鈥� and a 鈥楰鈥� prefix to 鈥楰nowledge鈥�.

The input libraries from top to bottom are:

• the Graduate Futures Institute (GFI) responsibilities
• Skills England Occupational Standard duties
• the US Occupational Information Network (O*NET) tasks
• National Careers Service (NCS) day-to-day tasks

The processing steps show:

• T1 鈥榲alidate as Task Statements鈥�
• T2 鈥榗luster by听SOC听SUG
• T3 鈥榰se听AI听to sub-cluster by meaning鈥�
• T4 鈥榰se听AI听to merge and deduplicate鈥� which has arrows pointing to 2 steps under T5
• T5 鈥榲alidate via听SOC听SUG description鈥� and 鈥榲alidate against job ads鈥� which both have arrows to step T6
• T6 鈥極ccupational Tasks鈥�

T1: Process and validate inputs

Task statement libraries were obtained from GFI听(responsibilities), Skills England Occupational Standards听(duties), O*NET听(tasks), and National Careers Service (day-to-day tasks). These libraries were then cleaned and standardised using听AI听tools.听AI听tools were used to quality assure the tasks statements and correct tasks that were too generic, too specific, too wordy, incorrectly structured, compound tasks or not tasks. The quality assurance process also converted US spellings and phrasing to UK English.

T2 - T4: Refine, deduplicate and cluster

Text embeddings were generated using two models: OpenAI 3-Large and Bidirectional Encoder Representations from Transformers (BERT) MP-Net and then a variety of cluster models were tested and compared to remove duplicate and similar tasks. OpenAI 3-Large embeddings with a hierarchical clustering model was found to have the best results.

Clustered tasks were sorted by meaning (based on embeddings) to identify overlapping and close clusters and merged through manual inspection. Orphan clusters (those containing only one task) were integrated with multi-task clusters using results from other clustering and embeddings models.

The centroid task statement within each cluster was identified and became the task label. These cluster labels then became the initial version of the听SSC听Task library.

T5: Validate against other sources

SOC SUGs

Tasks were extracted from all听SOC听SUG听descriptions (except n.e.c. groups ending /99) using the Llama3听LLM听and then embeddings were created to enable matching to the听SSC听Tasks. The similarity between the听SUG听description task embeddings and the听SSC听Task embeddings was calculated to provide a numerical score representing the degree of similarity. The best matching听SSC听Task for each听SUG听description task was identified so that听SSC听Tasks were assigned to all relevant听SOC听SUGs. Potential Task to听SUG听matches were also identified via an analysis of existing job profiles such as those within听O*NET听where associated task statements appear in clusters used to derive听SSC听Tasks.

Further听AI听prompts were used to check the combined mappings and estimate the relatedness of these tasks to the associated听SUGs. Significant discrepancies between a legacy mapping (such as a task match with a high level of importance within an听O*NET听profile but rejected by the听AI听analysis) were manually checked and reconciled.

Vacancy data

The听IER听holds a large vacancy database which is coded to听SOC听SUGs. A sample of distinct vacancy descriptions was created with a maximum size of 200 vacancies per听SUG. The sample was selected from vacancies with longer job descriptions and those that were well coded to each听SUG.

Llama3 was used to extract tasks from this database of vacancy descriptions and then the tasks were quality assured, clustered, and embeddings created using a similar process to the creation of the task library (T2-T4). These embeddings were then compared to the听SSC听Task embeddings. Vacancy tasks that were quality assured as being 鈥榞ood鈥� tasks but had a low similarity score to an existing听SSC听Task were manually inspected to identify any tasks that should be added to the听SSC听Task library.

The database of vacancy tasks was also used to identify additional tasks for听SUGs听with no or low numbers of associated tasks and similarly for听SSC听Skills with no linked听SSC听Tasks.

T6: Final SSC Occupational Tasks

The final list of听SSC听Tasks consists of 22,583 tasks. This is based primarily on the SSC prototype version but following analysis of tasks added to O*NET, Skills England Occupational Standards (i.e. duty statements) and GFI responsibilities since the original task library was created, an extra 692 tasks were added. This extended library was then mapped to听SSC听Skills and Knowledge concepts and to occupations.

Skills

Figure 9: Development of UK SSC Occupational Skills

Figure听9 shows the equivalent process for the construction of the hierarchical classification of听SSC听Occupational Skills together with a set of 13 Core Skills.

This is displayed as a series of processing steps from S1 to S7 in a row across the top of the diagram with each step shown below in a flow diagram. On the left-hand side are the 6 input libraries which feed into the first processing step: S1.

S1 to S7 refer to each processing step in the creation of the Occupational Skills library.

The input libraries from top to bottom are:

• European Skills, Competences, Qualifications and Occupations (ESCO) Level 4 skills
• the National Careers Service (NCS) skills
• O*NET听Detailed Work Activities (DWAs)
• Skills England听Occupational Standards skills
• GFI听skills
• the Workforce Foresighting Hub, Innovate UK (WFH) skills

The processing steps show:

• S1 鈥榲alidate as Skills鈥�
• S2 鈥榗luster by meaning鈥�
• S3 鈥榰se听AI听to merge and deduplicate鈥�
• S4 鈥榤ap against听SOC听SUGs鈥� which has arrows pointing to 2 steps under S5
• S5 鈥榲alidate against Tasks鈥� and 鈥榲alidate against job ads鈥� which both have arrows to step S6
• S6 鈥極ccupational Skills鈥�
• S7 鈥楥ore Skills鈥�

S1: Process and validate inputs

Skills statement libraries were obtained from听GFI听(skills),听ESCO听(Level 4 skills), Skills England听(skills), Innovate UK Workforce Foresighting Hub (skills), O*NET听(Detailed Work Activities) and the National Careers Service (skills). These libraries were cleaned and standardised using听AI听tools.听AI听tools were again used to quality assure the skill statements and correct any that were too generic, incorrectly structured, compound, invalid, elementary, ambiguous, traversal, or too specific.

The text below shows an example of a prompt used to quality assure skill statements:

Please note that this prompt was developed in May 2024 and used with the听LLM听model OpenAI GPT-4o. Current听LLMs听are significantly more capable and the prompt could be improved (quite possibly by an听LLM) to produce better results. Use of this exact prompt is therefore not recommended.

S2 - S3: Refine, deduplicate and cluster

OpenAI 3-Large embeddings were created and a hierarchical clustering model was used to deduplicate and refine the library of skills.

Skill clusters were then sorted by meaning to identify overlapping clusters and these were manually inspected for inclusion or deletion.听AI听prompts were used to analyse the consistency of the skill clusters and generate a new skill label to best describe the cluster of skills (rather than using the centroid skill as the label).

The verbs in the skill labels were standardised and became the听SSC听Skills.

AI听tools were used to write a description of the听SSC听Skill label and then a further prompt identified any ambiguous skills labels and descriptions which were rewritten.

Create Skill Groups, Areas and Domains

The听SSC听Skills were clustered to create Skill Groups and parent or child overlaps were manually checked. An听AI听prompt was used to check the听SSC听Skills within each Skill Group and identify any overlapping Skill Groups.

The Skill Groups were then clustered to create Skill Areas and the language of the Skill Groups and Skill Areas was standardised. An听AI听prompt was used to check the skills in each Skill Area and return a skill relatedness score.

The Skill Areas were then mapped to Skill Domains and an听AI听prompt used to check听SSC听skills within Skill Domains.

S4: Map against SOC SUGs

The original prototype mapping from听SOC听SUGs听to听SSC听Skills was based primarily on the occupational mappings in the input skill libraries.

The final Version 1.0 mapping was however entirely regenerated by first identifying potential matches from a text embedding comparison of the new Version 6 SOC SUG titles and descriptions against SSC skills and descriptions. The potential match lists were then extended by the addition of any of the top 30 skill matches from the original mapping not already included.

The original SUG to skill mapping contained only a single weighted importance score but, even with the latest LLMs (e.g. GPT-5.4), prompts to generate similar importance scores were inconsistent (i.e. running the same prompt against the same dataset would generate significantly different scores). This aligned with broader concerns about prototype mapping score accuracy. SUGs are quite broad occupational concepts and skills within some related roles may be very important while unrelated to others. For example, different application developer roles will involve using different programming languages and libraries which, in turn, will need different skills. A prompt evaluating the requirement 鈥榩robability鈥� for a skill within an SUG, and then the percentage 鈥榠mportance to competence鈥� separately led to the generation of significantly more consistent scores. Moreover, these were significantly more correlated with a sample of independent importance evaluations of the existing SSC skill matches than the original mapping.

The augmented potential match lists were therefore evaluated using the AI prompt format below to generate these two distinct estimates of relatedness. Matches with a frequency score below 10 and an average weighted score of below 25 were typically excluded, although some were retained to improve overall coverage (see 鈥淭he UK Standard Skills Classification鈥� for details).

Example prompt:

Weighted frequencies were then calculated to show how SUGs relate to SSC Skill Groups and SSC Skill Areas.

S5: Validate against other sources

SSC Tasks

The听SSC听Skills embeddings were compared to听SSC听Tasks embeddings to identify links between them. This mapping was then checked using an听AI听prompt and a further prompt defined the importance score of the听SSC听Skill to the听SSC听Task.

Vacancy data

Following a similar process to the validation of tasks using vacancy data, skills were extracted from a sample of vacancy descriptions using Llama3. These were quality assured using听AI听and then embeddings were created and the vacancy skills were clustered within each听SUG听and then across all听SUGs. The centroid embedding within each cluster became the vacancy skill label. These were then compared to the听SSC听Skills embeddings to check coverage and any vacancy skills quality assured as being of good quality with a low similarity score to the听SSC听Skills were inspected for inclusion. This resulted in eight new concepts (e.g. S.1388 - Install EV charging points) being added to the prototype classification.

S6: Final SSC Occupational Skills

The set of听SSC听Skills consists of a hierarchy of 3,350 Occupational Skills, 607 Skill Groups, 106 Skill Areas and 22 Skill Domains. This is based primarily on the prototype classification but, following user feedback and evaluation of pilot outputs, 10 new occupational skills were added, 27 skill labels modified (e.g. 鈥楽.0271 - Build axed arches and haunch brickwork鈥� changed to 鈥楤uild arches and angled brickwork鈥� to improve clarity) and two redundant concepts removed. All occupational skills descriptions were also revised using GPT-5.4 to improve consistency and readability. The datafile changelog contains full details.

S7: Core skills

The Skills Builder Partnership essential skill concepts were considered and then a list of 13听SSC听Core Skills and definitions were drawn up.

AI听prompts were used to help create definitions for each of the 5 skill levels of each听SSC听Core Skill and then to evaluate the level of Core Skill proficiency in each听SSC听Skill and each听SOC听SUG. Several听AI听models were used in this step to try to attain the best and most consistent results.

Knowledge

Figure 10: Development of UK SSC Knowledge concepts

Figure 10 illustrates the process to develop the听SSC听library of Knowledge concepts.

K1 to K6 refer to each processing step in the creation听of the Occupational Knowledge library.

The main input libraries from top to bottom are:

• ESCO听(European Skills, Competences, Qualifications and Occupations) Knowledge concepts
• Higher Education Coding of Subjects (HECoS)
• Learn Direct Classification of Subject Codes (LDCSC)
• O*NET听(knowledge, tools used and technology skills)
• Stack Exchange (topic tags)
• Wikipedia (article titles)

The processing steps show:

• K1 鈥榲alidate as Knowledge concepts鈥�
• K2 鈥榗luster by meaning鈥�
• K3 鈥榰se听AI听to merge and deduplicate鈥� which has arrows pointing to 5 steps under K4
• K4 鈥榲alidate versus听Ofqual鈥�, 鈥榲alidate versus听Skills England鈥�, 鈥榲alidate versus tasks鈥�, 鈥榲alidate versus job ads鈥�, and 鈥榲alidate versus prototype鈥� which each have arrows to step K5
• K5 鈥榠dentify primary concepts鈥�
• K6 鈥極ccupational Knowledge鈥�

The Knowledge concept, subject and topic names were collected from the input libraries.

K1: Process and validate inputs

Knowledge libraries were obtained from听ESCO听(Knowledge),听HECoS听(Higher Education Coding of Subjects), LDCSC听(Learn Direct Classification of Subject Codes), O*NET听(Knowledge, Tools Used and Technology Skills), Stack Exchange (Topic Tags) and Wikipedia (Article Titles). These were cleaned and standardised using听AI听tools. The list of Knowledge concepts was checked for any matching or equivalent terms and then filtered to only include concepts that were evident within a UK context.

K2 - K3: Refine, deduplicate and cluster

Knowledge concepts were clustered by meaning using embeddings and further deduplicated using clustering methods.

K4: Validate against other sources

Ofqual

Up to 50 potential matches per qualification were identified by comparing a text embedding vector of a concatenated text string of each qualification title and its associated qualification units against a text embedding for each听SSC听Knowledge concept label.

Text embedding vectors were generated using the OpenAI 3-Large Model with a cosine-similarity match threshold of 0.3 being applied. Matches above this threshold were then evaluated by prompting an听LLM听(GPT-5.4) with a simplified text string for each qualification (its simplified title and up to 5 example qualification units) to validate each match and also, where appropriate, assign a percentage probability that 鈥渁 significant amount of knowledge in that area would be learnt by achieving that qualification鈥�. Following a sample inspection, matches assigned a match probability score below 50% were rejected.

The closest Sector Subject Areas were identified using embedding matches and validated using an听LLM听prompt and manual inspection.

IfATE and Skills England

Up to 10 potential matches per Occupational Standard Knowledge statement were identified by comparing a text embedding vector of a concatenated text string of each statement and its associated occupational standard against a text embedding for each听SSC听Knowledge concept label. Text embedding vectors were generated using the OpenAI 3-Large Model with a cosine-similarity match threshold of 0.3 being applied. Matches above this threshold were then evaluated by prompting an听LLM听(GPT-5.4) and, where appropriate, assign a percentage importance of the knowledge to that statement. Following a sample inspection, matches assigned a probability score below 50% were rejected.

SSC Tasks

Embeddings matches were also used to assign听SSC听Knowledge concepts to听SSC听Tasks and then an听AI听prompt checked whether the Knowledge concepts had been correctly assigned to Tasks. An听AI听prompt was then used to define the importance score of the Knowledge to the Task.

Vacancy data

The sample of vacancy descriptions was searched for the听SSC听Knowledge concepts to check that they are all terms in common usage.

K5: Primary concepts

The primary concept type and potentially related concepts were identified using embeddings matches and checked using听LLM听prompts.

K6: Final SSC Occupational Knowledge concepts

The final set of听SSC听Knowledge concepts consists of 5,056 concepts linked to SSC听Tasks, SSC听Skills and subjects. This is based primarily on the prototype classification but, following user feedback, evaluation of pilot outputs and a re-analysis of previously excluded terms, 145 new concepts were added, 10 concept labels were modified to improve clarity (e.g. 鈥楰.0663 鈥� Casting鈥� changed to 鈥楥asting (Manufacturing)鈥�) and 15 redundant concepts were removed. All concept descriptions were also revised using GPT-5.4 to improve consistency and readability. The datafile changelog contains full details.

Secondary mappings

Secondary mappings to existing classifications of Skills, Tasks and Knowledge concepts were created using embeddings matches. The full list of secondary mappings available can be found in Appendix A.

Skill categorisations

1. Numeracy skills and Digital skills

These classifications were created using the听SSC听Skills that were rated as requiring an expert level of proficiency in the听SSC听Core Skills of Numeracy and Digital Literacy.

2. Green skills

An听AI听prompt was used to score each听SSC听Skill in how related (directly or indirectly) it is to the UK鈥檚 net zero emissions target and other environmental goals. Using previous work to define the Green听SOC听(see听, the skills mapped to green听SUGs听were also identified. A manual inspection of the skills with high听AI听green scores and those mapped to green听SUGs听was then carried out to identify a list of Green and Green enabling skills.

3. STEM-M&H (Science, Technology and Engineering, Mathematics, Medicine and Health) skills

The definitions of STEM-M&H used to define SUGs as STEM-M&H in were used in an AI prompt to score the SSC skills against each of the four categories. After a manual inspection and comparison to the STEM-M&H SUGs linked to each skill, a threshold score was applied to define the听STEM-M&H听category.

4. Artificial Intelligence (AI) skills

A model was developed to define 4 different categories of AI skills, as listed in Table 5 below.

Table 5: AI Skills Categories within the UK SSC

AI Skill Category Name	AI Skill Category Description
AI Development Skills	Technical skills that help develop, implement and maintain Artificial Intelligence (AI) tools and capabilities.
AI Operation Skills	Skills that directly relate to the use of Artificial Intelligence (AI) tools and capabilities.
AI-Augmented Skills	Skills that can be performed without AI tools and capabilities but can be materially simplified, accelerated, improved or scaled through their use.
AI Oversight Skills	Skills that help plan, govern, monitor, audit, assure, validate, regulate, approve, or oversee the safe, lawful, ethical, effective, or responsible use of Artificial Intelligence (AI) tools and capabilities.

An听AI听prompt was used to assign a percentage score for each of the categories for each听SSC听Skill. For scores above 50% a rationale was also generated for validation purposes.