Fine-tuning needs examples. Thousands of them, in a consistent format.<\/p>\n
For most business subjects, those examples do not exist in any collectable form.<\/strong><\/p>\n Three reasons:<\/strong><\/p>\n Confidentiality.<\/strong> Real audit findings, client contracts, support tickets \u2014 you cannot train on those, and often you cannot legally keep them for that purpose.<\/p>\n Volume.<\/strong> Even where real examples exist, there are rarely enough. A company might have a few hundred nonconformity reports. Fine-tuning wants far more.<\/p>\n Consistency.<\/strong> Real data is written by different people in different ways. Training on it teaches the model that inconsistency.<\/p>\n Synthetic data solves all three<\/strong> \u2014 with a cost, covered below.<\/p>\n Start from a knowledge base.<\/strong> A structured set of source material \u2014 standards, rules, definitions, terminology.<\/p>\n Generate examples from it.<\/strong> Each one an input and the output you want.<\/p>\n At volume.<\/strong> Hundreds of thousands, covering every combination systematically.<\/p>\n AboutKnowledge’s published quality models used exactly this approach.<\/strong> Both model cards state that the data was generated synthetically from a compact ISO and quality-management knowledge base.<\/p>\n One million records<\/strong> for the compliance analyzer. Five hundred thousand<\/strong> for the quality Q&A model, drawn from a million-record set.<\/p>\n FIGURE 1: HOW SYNTHETIC DATA IS BUILT<\/p>\n A knowledge base<\/p>\n Generate<\/p>\n At volume<\/p>\n Train<\/p>\n Coverage.<\/strong> You can generate examples for every standard, every clause, every verdict type \u2014 rather than whatever happened to be available.<\/p>\n AboutKnowledge’s dataset covers<\/strong> ISO 9001, 14001, 45001, 27001, 13485 and 50001, plus IATF 16949 and AS9100D, along with PDCA, CAPA, 8D, 5 Whys, FMEA, SPC, MSA, 5S, Lean and Six Sigma.<\/p>\n No real collection would cover that evenly.<\/strong> Some standards would be over-represented and some barely present.<\/p>\n Consistency.<\/strong> Every example follows the same format, which is what teaches the model to produce that format reliably.<\/p>\n No confidentiality problem.<\/strong> Nothing in the training data belongs to a client.<\/p>\n Volume on demand.<\/strong> If you need more, generate more.<\/p>\n The limitation is real and it should be stated plainly.<\/strong><\/p>\n Both AboutKnowledge model cards say it:<\/strong> synthetic data can contain occasional grammar artifacts or mix standards in generic answers.<\/p>\n What that means in practice:<\/strong><\/p>\n Small errors get learned.<\/strong> If the generation produced a slightly wrong phrasing in some examples, the model learns that phrasing.<\/p>\n Generic answers can blur.<\/strong> Asked something that spans standards, the model may produce an answer mixing requirements that belong to different ones.<\/p>\n It is only as good as the knowledge base.<\/strong> An error in the source is reproduced across every example generated from it.<\/p>\n Which is why both model cards say the same thing:<\/strong> always verify against the official standard text for compliance decisions.<\/p>\n FIGURE 2: WHAT SYNTHETIC DATA GIVES AND COSTS<\/p>\n Gives<\/p>\n Costs<\/p>\n Five things that determine quality.<\/p>\n A correct knowledge base.<\/strong> Everything derives from it. This is where the effort belongs<\/strong> \u2014 an error here appears in a million examples.<\/p>\n Realistic inputs.<\/strong> If real users write short informal statements, the training inputs should look like that. Training on polished formal input produces a model that struggles with how people actually write.<\/p>\n Consistent output structure.<\/strong> Every example the same shape. This is what the model learns.<\/p>\n Deliberate variety.<\/strong> Different phrasings, different lengths, different edge cases. Generation that produces near-identical examples teaches the model a narrow pattern.<\/p>\n Review a sample by hand.<\/strong> Read a few hundred. You will find systematic errors<\/strong>, and finding them before training is much cheaper than after.<\/p>\n That last point is the one people skip<\/strong>, and it is the cheapest quality control available.<\/p>\n A useful thing to notice in these two models:<\/strong><\/p>\n The quality Q&A model used 500,000 records over 10,000 steps<\/strong>.<\/p>\n The compliance analyzer used 1,000,000 records over 3,000 steps<\/strong>.<\/p>\n Different balances of data and training.<\/strong> More data with fewer passes, or less data with more passes.<\/p>\n Both reached similar training loss<\/strong> \u2014 0.1618 and 0.1742.<\/p>\n Which suggests there is no single right answer<\/strong>, and that the balance depends on the task and the model size.<\/p>\n And a broader point:<\/strong> having a million records available meant the option existed to use half. Generating more than you need is cheap, and it gives you room to experiment.<\/p>\n When real data exists at volume and you can use it.<\/strong> Real examples beat generated ones, if you have enough and they are consistent.<\/p>\n When the subject has no clear source of truth.<\/strong> Synthetic generation needs something to generate from. A subject that is genuinely a matter of judgement rather than documented rules does not have that.<\/p>\n When the errors would be dangerous.<\/strong> For anything where a subtly wrong answer causes real harm, the artifacts synthetic data introduces are a serious concern rather than an acceptable trade.<\/p>\n Which is exactly why both of these model cards limit their models to drafting and triage<\/strong>, and say so.<\/p>\n FIGURE 3: WHERE THE EFFORT BELONGS<\/p>\n The knowledge base<\/p>\n Realistic input phrasing<\/p>\n Deliberate variety<\/p>\n Read a sample by hand<\/p>\n Synthetic data makes fine-tuning possible when real examples are confidential, too few, or too inconsistent.<\/strong><\/p>\n It buys coverage and consistency.<\/strong> It costs you the errors in your source, reproduced across everything.<\/p>\n Put the effort into the knowledge base<\/strong>, make the inputs look like real ones, and read a sample by hand before training.<\/p>\n And publish the limitation.<\/strong> Both of these models say plainly that synthetic data can produce artifacts and that compliance decisions need verifying against the official text. That is the honest position.<\/p>\n A subject where the training data is confidential or does not exist?<\/p>\n Get in touch. We build synthetic datasets from a verified knowledge base \u2014 and we tell you where the resulting model should not be trusted.<\/p>\n<\/div>","protected":false},"excerpt":{"rendered":" The data problem Fine-tuning needs examples. Thousands of them, in a consistent format. For most business subjects, those examples do not exist in any collectable form. Three reasons: Confidentiality. Real audit findings, client contracts, support tickets \u2014 you cannot train on those, and often you cannot legally keep them for that purpose. Volume. Even where […]<\/p>\n","protected":false},"author":3,"featured_media":9447,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[26],"tags":[],"class_list":["post-9446","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai"],"yoast_head":"\nHow it works<\/h2>\n
\n
\n
\n
\n
What it gives you<\/h2>\n
The cost<\/h2>\n
\n
\n
Making it good<\/h2>\n
Where the volume comes from<\/h2>\n
When synthetic data is the wrong answer<\/h2>\n
\n
\n
\n
\n
The short version<\/h2>\n