Write your response in your own notes, then compare it with the sample answer. This practice is separate from the scored exam and does not affect your certificate.
Practice 1
Three invented releases: A has weights under Apache-2.0, no training code and a card saying only 'trained on public web data'; B has weights and code under OSI-approved licences and a data description that lists public datasets and describes a private one it cannot share; C has weights, code and full data information under a licence forbidding surveillance use. Say whether each could meet OSAID 1.0, and what you would record for A.
Compare with a sample answer
A cannot: the licence is permissive, but the complete code and the data information are missing, so I would record it as open weights under Apache-2.0, not as open source AI. B possibly can, because unshareable data may be described rather than published; I would check that the description is complete enough for a skilled person to build a substantially equivalent system. C cannot, because a use restriction conflicts with the freedom to use the system for any purpose without asking permission.
The sample applies the three OSAID elements and the freedom to use for any purpose to each case.
Practice 2
Your team wants to rely on an open model and a dataset for a public service. Say what status a dated, general course on licences has, and list when you would involve your legal team, a solicitor or your data protection officer.
Compare with a sample answer
A general course like this is education based on primary sources read on a stated date (26 September 2026); it is not legal advice, and terms and law can change after that date. I would involve our legal team or a solicitor before redistributing a model, before relying on a custom or use-restricted licence for a public service, before accepting another country's governing law, before using a dataset whose rights are unclear, and when a complaint arrives. I would involve our data protection officer whenever personal data may be present.
The sample keeps the not-legal-advice framing and lists each trigger for legal and data protection involvement.
Practice 3
Two invented datasets: civic-letters, under CC BY 4.0, holds residents' consultation letters with names and addresses and is wanted for a summarising model; clinic-notes, under CC BY-NC 4.0, is wanted to fine-tune a triage model that a supplier will sell. For each, name the main problem and the next step.
Compare with a sample answer
civic-letters: reuse is allowed with credit, but names and addresses are personal data, so data protection law applies whatever the licence says; the next step is a lawful basis and DPIA screening. I would also ask whether the publisher could license letters that their writers own. clinic-notes: selling the model is commercial use, which the NC condition rules out, so we must get other terms or not sell it; clinic notes may also hold health data, which needs care.
The sample separates the licence question from item rights and personal data, as the course teaches.
Practice 4
An invented card reads: 'kestrel-chat 3.0. License: other (Kestrel Community Licence 2.0). base_model: kestrel-base 3.0. Trained on a web crawl and forum-scrape. Intended use: general chat. Limitations: may make mistakes.' List the bill of materials fields it fills and what is missing.
Compare with a sample answer
Filled: the name and version; a declared licence of LicenseRef-Kestrel-Community-2.0, with the concluded licence pending review; the base model as pedigree; trainedOn the web crawl and forum-scrape; one use case; and one vague limitation. Missing: supplier, download location, release time, a hash, the base model's licence, the dataset licences and provenance, personal data, evaluation results and the user threshold.
The sample maps card fields to SPDX and CycloneDX terms and lists the gaps named in the course answer.
Practice 5
In the licence checklist practice, harbour-triage and kestrel-chat are corrected in a copy of the inventory and the tool is run with --write. Describe what the run reports, and explain which STOP no honest edit could fix.
Compare with a sample answer
The run prints ok for fern-summarise, only the DUTY line for harbour-triage, and STOP unknown dataset terms '' for kestrel-chat, with totals of one ok, one DUTY and one STOP, and it writes practice/NOTICE-inventory-fixed.txt. The remaining STOP comes from the forum-scrape dataset, which has no licence: no edit can honestly fix it until real terms are obtained, because edits are honest only if the facts have changed.
The sample reports the recorded output and explains why honest edits must reflect changed facts.
Practice 6
Explain how the harmless pickle demonstration shows that loading a pickle can run code, how you can inspect a pickle safely, and what safer formats do and do not solve.
Compare with a sample answer
In the disassembly, STACK_GLOBAL finds builtins.print and REDUCE calls it, so the last line is printed while the data is loading. Running python -m pickletools disassembles a file without loading it, so you can read the steps without running them. Safer formats such as safetensors hold data, not instructions: a header length, a JSON header and raw tensor bytes, designed so loading runs no arbitrary code. Loaders and shipped code still need checking, and I would still verify hashes against the publisher's value.
The sample names where the function is found and called, the safe inspection route and the limits of safer formats.