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 — ISO 9001, 14001, 45001, 27001, 13485, 50001, plus IATF 16949 for automotive and AS9100D for aerospace.
Methods — PDCA, risk-based thinking, the process approach, CAPA, 8D, 5 Whys, FMEA, SPC, MSA, 5S, Lean and Six Sigma.
Practical outputs — audit checklists, nonconformity reports, KPI suggestions, interview questions.
What it is for: answering questions and drafting. Somebody preparing for an audit, writing a checklist, or checking what a clause requires.
Why a 1B model
The obvious question, given that far larger models exist.
Three reasons, and they compound.
All its capacity is on one subject. 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.
It runs anywhere. Roughly 1.3 GB as a Q8_0 GGUF file — a normal laptop, no GPU required.
It costs nothing per question. No API bill that grows with use.
The trade-off is real: ask it about anything outside quality management and the answer is unremarkable.
Which is fine, because that is not what it is for.
FIGURE 1: WHY SMALL WORKED HERE
One subject, all the capacity
- Nothing spent on general knowledge it will never need.
Runs on a laptop
- 1.3 GB. No GPU, no server.
No per-call cost
- Ask it a thousand questions. Same cost.
Data stays local
- Nothing sent anywhere.
The training
Numbers from the published model card:
| | |
|—|—|
| Base model | Llama 3.2 1B Instruct |
| Framework | Unsloth |
| LoRA rank / alpha | 16 / 16 |
| Dropout | 0.0 |
| Quantised training | 4-bit NF4 |
| Training records | 500,000 |
| Training steps | 10,000 |
| Final train loss | 0.1618 |
| Sequence length | 2048 |
Two things worth noting.
500,000 records, drawn from a larger million-record dataset. More data is available than was used — a deliberate choice rather than a limitation.
Dropout at zero. 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.
Where the training data came from
Synthetic, generated from a compact ISO and quality-management knowledge base.
Why that approach makes sense here:
Real audit data is confidential. You cannot train on client nonconformity reports.
Volume. Half a million examples of real quality-management Q&A do not exist in any collectable form.
Coverage. Generating lets you cover every standard, every method, every combination — rather than whatever happened to be available.
And the honest limitation, which the model card states plainly: synthetic data can contain grammar artifacts or mix standards in generic answers.
Always verify against the official standard text for compliance decisions.
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.
How it is used
Two routes, both published.
In code, loading the adapter onto the base model and addressing it with a system prompt establishing it as a quality management and ISO expert.
Locally, using the published GGUF file and a Modelfile with Ollama — one command to create, one to run.
The published example matters. 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.
FIGURE 2: HOW IT GETS USED
A question
- What does a clause require?
Local model
- Runs on a laptop, no network
Answer
- Grounded in the training subject
A person verifies
- Against the official standard text
What it is good for
Preparation and drafting.
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.
All of it a first draft that a person reviews.
What it saves: the blank page and the lookup. Both real, both repetitive, and neither requiring the judgement of a qualified auditor.
What it is not for
The model card is explicit, and this is the section to read most carefully.
It should not replace human auditors or regulatory review.
Compliance decisions must be verified against the official standard text.
Why that matters: 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.
Useful for drafting and triage. Not for the decision.
FIGURE 3: WHERE IT BELONGS
Good use
- Drafting a checklist before an audit
- Understanding what a clause covers
- First version of a nonconformity report
- KPI and question suggestions
Not this
- Deciding whether you are compliant
- Anything a certification body will review
- Quoting clause wording as authoritative
- Replacing a qualified auditor
What this shows
Three things worth taking from it.
A narrow subject can support a small model. You do not need a large one to be useful at one thing.
Synthetic data can work, where real data is confidential or does not exist at volume — with the caveat stated openly.
Local matters for this subject. Quality and compliance work involves confidential material. A model on your own machine sends nothing anywhere.
The short version
A 1B model, fine-tuned on quality management and ISO standards, running on a laptop with no per-question cost.
It answers questions and drafts documents. Checklists, reports, KPIs, audit questions.
It does not decide compliance, and the model’s own documentation says so — which is the mark of a model published honestly.
Small worked because the subject is narrow. All the capacity, one job.
Preparing for audits and starting from blank pages?
Get in touch. We build fine-tuned models for narrow business subjects — running locally, so confidential material stays where it is.
