A LoRA adapter fine-tuned on Llama 3.2 3B<\/strong>, built to analyse statements about how an organisation operates and return a structured compliance assessment.<\/p>\n Published openly on Hugging Face by AboutKnowledge (Hong Kong) Limited<\/strong>.<\/p>\n Give it a statement<\/strong> \u2014 “We do not document management review meetings” \u2014 and it returns:<\/p>\n A verdict.<\/strong> Compliant, partially compliant, or non-compliant.<\/p>\n The relevant clause.<\/strong> Which requirement this touches.<\/p>\n An explanation.<\/strong> Why that verdict.<\/p>\n Recommendations.<\/strong> What to do about it.<\/p>\n That structure is the point of the model<\/strong>, and it is what separates it from asking a general model the same question.<\/p>\n A general model asked this question gives you a paragraph.<\/strong> Useful to read, awkward to act on.<\/p>\n A structured answer can be used by a system.<\/strong><\/p>\n Three things become possible:<\/strong><\/p>\n Sorting.<\/strong> Show me everything non-compliant. Filter by clause.<\/p>\n Counting.<\/strong> How many findings, of which severity, against which standard.<\/p>\n Routing.<\/strong> Non-compliant findings go to one queue, partially compliant to another.<\/p>\n None of that works on prose.<\/strong> A downstream system needs fields, and a fine-tuned model produces them consistently \u2014 without a long instruction on every call.<\/p>\n FIGURE 1: PROSE VERSUS STRUCTURE<\/p>\n A general model<\/p>\n A fine-tuned analyzer<\/p>\n AboutKnowledge published two related models, and the split is deliberate.<\/strong><\/p>\n The 1B quality model<\/strong> answers questions and drafts documents. What does a clause require? Write me a checklist. Suggest some KPIs.<\/p>\n The 3B compliance analyzer<\/strong> makes judgements. Here is what we do \u2014 is it compliant, and against what?<\/p>\n Why not one model for both?<\/strong><\/p>\n Different jobs need different training.<\/strong> One is trained to explain; the other to assess and cite. Training a single model on both produces something that does neither cleanly.<\/p>\n And different jobs justify different sizes.<\/strong> Judgement with citation is a harder task than answering a question, which is why the analyzer is the larger of the two.<\/p>\n That is a design decision worth borrowing.<\/strong> Rather than one model that does everything adequately, two models each doing one thing properly.<\/p>\n From the published model card:<\/strong><\/p>\n | | |<\/p>\n |—|—|<\/p>\n | Base model | Llama 3.2 3B Instruct |<\/p>\n | Framework | Unsloth |<\/p>\n | LoRA rank \/ alpha | 16 \/ 16 |<\/p>\n | Dropout | 0.0 |<\/p>\n | Quantised training | 4-bit NF4 |<\/p>\n | Training records | 1,000,000 |<\/p>\n | Training steps | 3,000 |<\/p>\n | Final train loss | 0.1742 |<\/p>\n | Sequence length | 2048 |<\/p>\n A million training records<\/strong>, each an example of a statement and the structured assessment it should produce.<\/p>\n Which is the real work.<\/strong> Not the training run \u2014 assembling a million consistent examples of what a good compliance assessment looks like.<\/p>\n Standards<\/strong> \u2014 ISO 9001, 14001, 45001, 27001, 13485, 50001, plus IATF 16949 for automotive and AS9100D for aerospace.<\/p>\n Verdicts<\/strong> \u2014 compliant, partially compliant, non-compliant.<\/p>\n Output<\/strong> \u2014 verdict, cited clauses, explanation, recommendations.<\/p>\n The three-way verdict is worth noting.<\/strong> Compliance is rarely binary. “Partially compliant” is the honest answer for most real findings \u2014 the thing is being done, just not documented, or done inconsistently.<\/p>\n A model that only said yes or no would be less useful<\/strong>, because it would force a wrong answer on the most common case.<\/p>\n FIGURE 2: WHY THE DESIGN CHOICES MAKE SENSE<\/p>\n Three-way verdict<\/p>\n Cited clause<\/p>\n Recommendations included<\/p>\n 3B rather than 1B<\/p>\n Both routes are published<\/strong>, as with the sibling model.<\/p>\n In code<\/strong>, loading the adapter with the intended system prompt \u2014 establishing it as a quality-management auditor and ISO consultant, and asking for a verdict, clauses, reasoning and recommendations.<\/p>\n Locally<\/strong>, using the published Q8_0 GGUF file \u2014 about 3.4 GB \u2014 with Ollama.<\/p>\n The system prompt matters more here than usual.<\/strong> The model was trained to produce a specific structure in response to a specific framing. Address it differently and the structure degrades.<\/strong><\/p>\n Triage.<\/strong> Run a set of statements about your operations through it and see what comes back flagged. That gives an audit team a starting list rather than a blank page.<\/p>\n Preparation.<\/strong> Before an internal audit, check what is likely to come up.<\/p>\n Drafting findings.<\/strong> A first version of a nonconformity write-up, with the clause already identified.<\/p>\n Consistency.<\/strong> Different auditors assess the same thing differently. A model applies the same standard to everything, which is useful as a cross-check even when a person makes the final call.<\/p>\n The model card is explicit<\/strong>, and this is the important part.<\/p>\n It should not replace human auditors or regulatory review.<\/strong><\/p>\n Always verify against the official standard text for compliance decisions.<\/strong><\/p>\n Why:<\/strong> certification depends on precise clause wording. A model trained on synthetic data can produce a plausible clause reference that is subtly wrong \u2014 right standard, wrong sub-clause, or right idea in slightly wrong words.<\/p>\n For triage and drafting, that is acceptable.<\/strong> For a decision a certification body will examine, it is not.<\/p>\n FIGURE 3: HOW TO USE IT PROPERLY<\/p>\n Appropriate<\/p>\n Not appropriate<\/p>\n Three transferable points.<\/p>\n Structured output is a training objective, not a prompting trick.<\/strong> If a downstream system needs fields, train for fields.<\/p>\n Split tasks across models rather than overloading one.<\/strong> Two models doing one job each beat one doing both adequately.<\/p>\n Publish the limitations.<\/strong> Both these model cards say plainly where the output should not be trusted. That is what a professionally published model looks like<\/strong>, and its absence elsewhere is a warning.<\/p>\n A 3B model fine-tuned to return a verdict, a clause, an explanation and recommendations<\/strong> \u2014 consistently, in a shape a system can use.<\/p>\n Trained on a million examples<\/strong>, which is the real work behind it.<\/p>\n A three-way verdict<\/strong>, because most real compliance findings are partial rather than pass or fail.<\/p>\n For triage and drafting, not for the decision<\/strong> \u2014 and the model’s own documentation says so, which is how it should be.<\/p>\n Audit preparation starting from a blank page every time?<\/p>\n Get in touch. We build models that produce structured, consistent output your systems can act on \u2014 running locally, so nothing confidential leaves your infrastructure.<\/p>\n<\/div>","protected":false},"excerpt":{"rendered":" What it is A LoRA adapter fine-tuned on Llama 3.2 3B, built to analyse statements about how an organisation operates and return a structured compliance assessment. Published openly on Hugging Face by AboutKnowledge (Hong Kong) Limited. Give it a statement \u2014 “We do not document management review meetings” \u2014 and it returns: A verdict. Compliant, […]<\/p>\n","protected":false},"author":3,"featured_media":9444,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[26,30],"tags":[],"class_list":["post-9443","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai","category-compliance"],"yoast_head":"\nWhy structured output matters<\/h2>\n
\n
\n
Two models, two jobs<\/h2>\n
The training<\/h2>\n
What it covers<\/h2>\n
\n
\n
\n
\n
How it is used<\/h2>\n
Where it fits<\/h2>\n
Where it does not fit<\/h2>\n
\n
\n
What this shows<\/h2>\n
The short version<\/h2>\n