A LoRA adapter fine-tuned on Llama 3.2 1B<\/strong>, trained on quality management systems and ISO standards.<\/p>\n Published openly on Hugging Face, and built by AboutKnowledge (Hong Kong) Limited<\/strong>.<\/p>\n What it covers:<\/strong><\/p>\n Standards<\/strong> \u2014 ISO 9001, 14001, 45001, 27001, 13485, 50001, plus IATF 16949 for automotive and AS9100D for aerospace.<\/p>\n Methods<\/strong> \u2014 PDCA, risk-based thinking, the process approach, CAPA, 8D, 5 Whys, FMEA, SPC, MSA, 5S, Lean and Six Sigma.<\/p>\n Practical outputs<\/strong> \u2014 audit checklists, nonconformity reports, KPI suggestions, interview questions.<\/p>\n What it is for:<\/strong> answering questions and drafting. Somebody preparing for an audit, writing a checklist, or checking what a clause requires.<\/p>\n The obvious question<\/strong>, given that far larger models exist.<\/p>\n Three reasons, and they compound.<\/strong><\/p>\n All its capacity is on one subject.<\/strong> A 1B model has limited room. Spent entirely on quality management, that room goes further than the same subject occupying a fraction of a much larger general model.<\/p>\n It runs anywhere.<\/strong> Roughly 1.3 GB as a Q8_0 GGUF file \u2014 a normal laptop, no GPU required.<\/p>\n It costs nothing per question.<\/strong> No API bill that grows with use.<\/p>\n The trade-off is real:<\/strong> ask it about anything outside quality management and the answer is unremarkable.<\/p>\n Which is fine, because that is not what it is for.<\/strong><\/p>\n FIGURE 1: WHY SMALL WORKED HERE<\/p>\n One subject, all the capacity<\/p>\n Runs on a laptop<\/p>\n No per-call cost<\/p>\n Data stays local<\/p>\n Numbers from the published model card:<\/strong><\/p>\n | | |<\/p>\n |—|—|<\/p>\n | Base model | Llama 3.2 1B 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 | 500,000 |<\/p>\n | Training steps | 10,000 |<\/p>\n | Final train loss | 0.1618 |<\/p>\n | Sequence length | 2048 |<\/p>\n Two things worth noting.<\/strong><\/p>\n 500,000 records, drawn from a larger million-record dataset.<\/strong> More data is available than was used \u2014 a deliberate choice rather than a limitation.<\/p>\n Dropout at zero.<\/strong> Reasonable with a training set this size. Dropout guards against a model memorising its training data; with half a million examples, that is a smaller risk.<\/p>\n Synthetic<\/strong>, generated from a compact ISO and quality-management knowledge base.<\/p>\n Why that approach makes sense here:<\/strong><\/p>\n Real audit data is confidential.<\/strong> You cannot train on client nonconformity reports.<\/p>\n Volume.<\/strong> Half a million examples of real quality-management Q&A do not exist in any collectable form.<\/p>\n Coverage.<\/strong> Generating lets you cover every standard, every method, every combination \u2014 rather than whatever happened to be available.<\/p>\n And the honest limitation<\/strong>, which the model card states plainly: synthetic data can contain grammar artifacts or mix standards in generic answers.<\/strong><\/p>\n Always verify against the official standard text for compliance decisions.<\/strong><\/p>\n That caveat being in the model’s own documentation is the right way to publish a model. Its absence elsewhere is a reason for caution.<\/p>\n Two routes, both published.<\/strong><\/p>\n In code<\/strong>, loading the adapter onto the base model and addressing it with a system prompt establishing it as a quality management and ISO expert.<\/p>\n Locally<\/strong>, using the published GGUF file and a Modelfile with Ollama \u2014 one command to create, one to run.<\/p>\n The published example matters.<\/strong> A fine-tuned model expects to be addressed a certain way, and using the intended system prompt is the difference between good output and disappointing output.<\/p>\n FIGURE 2: HOW IT GETS USED<\/p>\n A question<\/p>\n Local model<\/p>\n Answer<\/p>\n A person verifies<\/p>\n Preparation and drafting.<\/strong><\/p>\n Working out what a clause requires before an audit. Drafting a checklist. Producing a first version of a nonconformity report. Suggesting KPIs. Generating interview questions for an internal audit.<\/p>\n All of it a first draft that a person reviews.<\/strong><\/p>\n What it saves:<\/strong> the blank page and the lookup. Both real, both repetitive, and neither requiring the judgement of a qualified auditor.<\/p>\n The model card is explicit<\/strong>, and this is the section to read most carefully.<\/p>\n It should not replace human auditors or regulatory review.<\/strong><\/p>\n Compliance decisions must be verified against the official standard text.<\/strong><\/p>\n Why that matters:<\/strong> an ISO clause has a precise wording, and certification depends on it. A model trained on synthetic data can produce a plausible summary of a clause that is subtly wrong.<\/p>\n Useful for drafting and triage. Not for the decision.<\/strong><\/p>\n FIGURE 3: WHERE IT BELONGS<\/p>\n Good use<\/p>\n Not this<\/p>\n Three things worth taking from it.<\/p>\n A narrow subject can support a small model.<\/strong> You do not need a large one to be useful at one thing.<\/p>\n Synthetic data can work<\/strong>, where real data is confidential or does not exist at volume \u2014 with the caveat stated openly.<\/p>\n Local matters for this subject.<\/strong> Quality and compliance work involves confidential material. A model on your own machine sends nothing anywhere.<\/p>\n A 1B model, fine-tuned on quality management and ISO standards, running on a laptop with no per-question cost.<\/p>\n It answers questions and drafts documents.<\/strong> Checklists, reports, KPIs, audit questions.<\/p>\n It does not decide compliance<\/strong>, and the model’s own documentation says so \u2014 which is the mark of a model published honestly.<\/p>\n Small worked because the subject is narrow.<\/strong> All the capacity, one job.<\/p>\n Preparing for audits and starting from blank pages?<\/p>\n Get in touch. We build fine-tuned models for narrow business subjects \u2014 running locally, so confidential material stays where it is.<\/p>\n<\/div>","protected":false},"excerpt":{"rendered":" What it is A LoRA adapter fine-tuned on Llama 3.2 1B, trained on quality management systems and ISO standards. Published openly on Hugging Face, and built by AboutKnowledge (Hong Kong) Limited. What it covers: Standards \u2014 ISO 9001, 14001, 45001, 27001, 13485, 50001, plus IATF 16949 for automotive and AS9100D for aerospace. Methods \u2014 PDCA, […]<\/p>\n","protected":false},"author":3,"featured_media":9441,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[26,29],"tags":[],"class_list":["post-9440","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai","category-quality_management"],"yoast_head":"\nWhy a 1B model<\/h2>\n
\n
\n
\n
\n
The training<\/h2>\n
Where the training data came from<\/h2>\n
How it is used<\/h2>\n
\n
\n
\n
\n
What it is good for<\/h2>\n
What it is not for<\/h2>\n
\n
\n
What this shows<\/h2>\n
The short version<\/h2>\n