AI Model Register
Last updated August 2026
This page lists every artificial intelligence model that Mark My Words Australia Pty Ltd (ABN 93 662 681 450) uses to deliver its platform, what each model does, what data it receives, where it runs, and what its known limitations are. It also records the version of each model, a history of changes, and how we notify schools when a model changes.
We publish this register so that schools, school authorities and assessment bodies can see exactly what is processing student writing. It supports our assessment under the Safer Technologies 4 Schools (ST4S) framework and its Artificial Intelligence module, and should be read alongside our Privacy Policy and Sub-Processor Register.
How our AI is built and hosted
Mark My Words builds and runs its own AI models. We do not send student writing, assessment data or any other platform data to a third-party AI service such as OpenAI, Google, Anthropic or Alibaba Cloud. There is no external AI provider in the path between a student’s work and an assessment result.
Our models are built on open-weight base models, principally the Qwen family published by Alibaba Cloud. An open-weight model is one whose trained parameters are published for anyone to download and run. We download those parameters and run them on infrastructure we control, inside Microsoft Azure’s Australian regions. Using a Qwen base model does not mean that any data is sent to Alibaba, that Alibaba can access our data, or that we hold an account with Alibaba — the relationship is the same as using any other openly published software component.
On top of each base model we train our own adaptations using de-identified data, as described in the Privacy Policy. This is what makes the models specific to Australian curriculum writing assessment and to the particular rubrics a school uses. The base model provides general language and vision capability; our training provides the assessment capability.
- Where models run: Microsoft Azure, Australia East and Australia Southeast regions
- Who operates them: Mark My Words, on infrastructure we control
- Third-party AI services in the processing path: none
- Data used to train our adaptations: de-identified data only, as set out in our Privacy Policy
Models currently in service
The following models are in service today. Each entry shows the version currently deployed to production.
Marking and rubric scoring
In productionVersion MMW-Marking 2026.07
- What it does
- Scores a piece of student writing against the criteria and descriptors of a marking rubric, and produces the reasoning and evidence supporting each score. Separate adapters are trained for each assessment framework we support, including NAPLAN, VCOP and Seven Steps.
- Base model
- Qwen2.5-14B-Instruct — an open-weight model published by Alibaba Cloud. We download the model weights and run them on our own infrastructure. We do not use Alibaba Cloud services and no data is sent to Alibaba.
- Model type
- Generative language model with supervised fine-tuning and transfer learning (low-rank adapters trained by Mark My Words).
- Data it receives
- The transcribed text of a student writing sample, the rubric criteria and descriptors for the selected assessment framework, and the assessment task or prompt where one is provided.
- What it produces
- A score against each rubric criterion, the descriptor selected, and the reasoning and textual evidence supporting that score.
- Where it runs
- Microsoft Azure, Australia East and Australia Southeast regions.
Known limitations: Scores are an estimate produced by a statistical model and are intended to support a teacher’s professional judgement, not replace it. Accuracy varies by rubric, year level and text type. Teachers can review and override every score.
Written feedback generation
In productionVersion MMW-Feedback 2026.07
- What it does
- Generates written feedback for a student based on the rubric outcomes produced by the marking model, aligned to the descriptors the student has and has not yet met.
- Base model
- Qwen2.5-14B-Instruct — the same open-weight base model as the marking model, with a separate adapter and prompt trained for feedback.
- Model type
- Generative language model with supervised fine-tuning and transfer learning.
- Data it receives
- Rubric outcomes from the marking model, the transcribed writing sample, and the student’s year level.
- What it produces
- Written formative feedback and suggested next steps for the student.
- Where it runs
- Microsoft Azure, Australia East and Australia Southeast regions.
Known limitations: Feedback is generated text and may occasionally be generic, repetitive, or misaligned with a teacher’s intent. All feedback is visible to the teacher and can be edited or withheld before a student sees it.
Handwriting recognition (OCR)
In productionVersion MMW-OCR 2026.03
- What it does
- Transcribes handwritten and printed student writing from uploaded images and documents into text so that it can be assessed. Also detects page orientation and identifies the type of content on each page.
- Base model
- A Mark My Words model built on Qwen VL Instruct, an open-weight vision-language model published by Alibaba Cloud. The weights are downloaded and run on our own infrastructure. No data is sent to Alibaba.
- Model type
- Generative vision-language model, fine-tuned by Mark My Words using supervised fine-tuning followed by reinforcement learning to improve the accuracy of text positioning on the page.
- Data it receives
- Images and document pages of student writing uploaded by a teacher.
- What it produces
- The transcribed text of the page, the position of that text on the page, page orientation, and a classification of the page content.
- Where it runs
- Microsoft Azure, Australia East and Australia Southeast regions.
Known limitations: Transcription accuracy depends heavily on handwriting legibility, image quality and page layout. Very faint, heavily corrected or unusually formatted work may transcribe poorly. Teachers can view the transcription alongside the original image.
Writing year-level estimate
In productionVersion MMW-YearLevel 2026.05
- What it does
- Estimates the writing year level demonstrated by a piece of student writing, based on vocabulary range, sentence structure, idea development and text structure.
- Base model
- Qwen2.5-14B-Instruct — open weights, run on our own infrastructure. No data is sent to Alibaba.
- Model type
- Predictive model built on a generative language model using transfer learning.
- Data it receives
- The transcribed text of a student writing sample.
- What it produces
- An estimated writing year level between Foundation and Year 10, which may be a fractional value.
- Where it runs
- Microsoft Azure, Australia East and Australia Southeast regions.
Known limitations: This is a normative estimate of the writing itself, not an assessment of the student. It should not be used to make placement, streaming or reporting decisions on its own.
How we version models
Every model in this register carries a version number in the form Name YYYY.MM, where the date is the month that version was released to production. A version identifies a specific, fixed combination of:
- the base model and its published revision;
- the adaptation we have trained on top of it;
- the instructions given to the model at the time of assessment; and
- the rubric configuration applied to the model’s output.
A version is fixed once released. If any part of that combination changes, we issue a new version number rather than altering an existing one. This means a school can always identify which version assessed a particular piece of work, and we can reproduce or review any past assessment.
We distinguish between two kinds of change. A material change is one that can reasonably be expected to alter the scores, year-level estimates or feedback a school receives for the same piece of work. A non-material change improves speed, reliability or resilience without changing assessment output. Both are recorded in the change log below; only material changes trigger advance notice.
How we notify schools about new model releases
Whenever we release a new or updated AI model:
- We tell you before it happens. Material changes are announced in the platform in advance of reaching production, so you can raise any concerns with us first.
- We publish the version. The version number of each model in service is listed above and updated on the day a new version is released.
- We keep a record. Every release, material or not, is added to the change log below, with the date, the model affected and a plain-English description of what changed.
- We show what is coming. Planned releases we expect to be material are listed before they happen.
Our models run on our own infrastructure in Australia, and a model release does not change that.
Planned releases
The following changes are in development. Dates are indicative and each will be notified in accordance with the commitments above before it reaches production.
| Date | Model | Change |
|---|---|---|
| Expected late 2026 | Marking and rubric scoring | Material changeA new marking model that takes account of a school or partner’s rubric configuration at the time of marking, so that criteria and descriptors can be adjusted without retraining the model. This will change marking output and will be notified as a material change before release. |
| Expected late 2026 | Handwriting recognition (OCR) | Material changeA new transcription model built on an updated vision-language base model, targeting improved accuracy on less legible handwriting. |
Model change log
This log records changes to the AI models used in our platform. Earlier records, and details of any version not listed here, are available on request.
| Date | Model | Change |
|---|---|---|
| July 2026 | Marking and rubric scoring | Material changeReleased the NAPLAN marking adapter to production, alongside the existing VCOP and Seven Steps adapters. Schools using these frameworks are now marked against a rubric-specific model rather than a general one. |
| May 2026 | Writing year-level estimate | Material changeReleased the normative writing year-level model, replacing an earlier approach that produced year-level estimates as part of the marking output. |
| March 2026 | Handwriting recognition (OCR) | Released improved page rotation and page classification, reducing transcription errors on scanned and photographed work that is not upright on the page. |
| September 2025 | Handwriting recognition (OCR) | Replaced the transcription model with a smaller model of equivalent measured accuracy, improving processing speed and reliability at volume. |
Human oversight and limitations
Every model listed here supports a teacher’s judgement rather than replacing it. Scores and feedback are presented to a teacher for review, and a teacher can change or disregard any output before it is used or shared with a student.
All of these models are statistical systems. They produce estimates, and those estimates are sometimes wrong. Accuracy varies with the legibility of handwriting, the rubric being applied, the year level, and the type of writing being assessed. We measure the accuracy of each model before release against writing marked by teachers, and we do not release a model that performs worse than the version it replaces on those measures.
Output generated by our models is identified as AI-generated within the platform. Schools and school administrators can control which AI features are available to their users, and teachers can report inaccurate or inappropriate output to us from within the platform.
Questions and updates
We review this register at least annually and update it whenever a model is added, removed, replaced or re-versioned. Each published version of this register is dated above, and a record of previous versions is available on request.
If you have questions about the models we use, want detail on how a particular assessment was produced, or need supporting documentation for a procurement or assessment process, please contact us:
- Email: hello@markmywords.au
- Phone: (+61) 403 651 221
- Address: 700 Connect, University of Melbourne, Victoria