How specialist knowledge changes a model.
Exploring domain-adaptive pretraining and fine-tuning: how a model responds to specialist material, what improves and what still needs evaluation. Project OWL is our first active investigation.
Pulseee / AI LabsHuman-led
Human-led
research.
Openly shared.
A passion-led space for model adaptation, practical benchmarking and human-led analysis. We share what we learn so others can examine it, question it and build on it.
Research / Beyond the surfaceBetter questions.
Better questions.
Clearer evidence.
Our research interests centre on how AI learns, how it behaves and what makes it useful. We aim to turn practical experiments into human-led analysis that others can examine and build on.
That means looking closely at the data, the model and the surrounding tools, then explaining what changed, what helped and what remains uncertain.
How we plan to share the work ↗What happens outside the ideal example.
Exploring how models handle unfamiliar questions, incomplete information, longer context and multi-step tasks. The difficult cases can be as informative as the successful ones.
Where the technology earns its place.
Hands-on analysis of AI tools, platforms and services: their capabilities, constraints and integration trade-offs, and the work required to make them useful in practice.
Benchmarks / A closer testCapability,
Capability,
put to the test.
A model can look impressive in a demonstration. A useful evaluation asks how it performs on a defined task, under stated conditions, across more than one attempt.
Our benchmarking focus is on the strengths, limitations and practical trade-offs of current models and tools. These are the dimensions we intend to examine and share.
Task quality
Did the output satisfy the requirements, follow the instructions and produce a useful result?
Groundedness
Are claims supported by the supplied evidence? Does the model recognise when information is missing?
Consistency
Does performance hold across repeated runs, varied wording and more difficult examples?
Speed
How long does a useful result take, under the stated model, hardware and workflow conditions?
Cost
What does the tested task cost, including retries and supporting tools where relevant?
Tool use
Can the system choose appropriate tools, use them correctly and complete the intended steps?
Context belongs beside the score.
Comparisons will identify the model versions, task setup and evaluation criteria. We’ll explain where a result applies and where it should not be generalised.
How we plan to share the workThe method belongs
The method belongs
beside the result.
Research becomes more useful when people can understand how it was done. Our publications will aim to make the reasoning, conditions and limits visible.
What we set out to learn.
The question, its practical relevance and the assumptions behind the work.
How the work was tested.
Model and tool versions, task selection, data sources and evaluation criteria.
What the evidence showed.
Observed results, meaningful differences and failure cases worth examining.
What remains unresolved.
Limitations, open questions and supporting materials where they can be shared.
Human-led means accountable. People frame the questions, review the evidence and take responsibility for the conclusions. AI can assist the work; careful human judgement shapes what we publish.
Active research / Project OWLFine-tuning in progress
Open Workplace Law.
Exploring how a language model can be adapted to Australian workplace law.
Following domain-adaptive pretraining on the Fair Work Act, our Qwen3.5-27B model is currently being fine-tuned. Evaluation is planned before a future public release on Hugging Face.
- Research domain
- Australian workplace law
- Training focus
- Fair Work Act
Base modelQwenQwen3.5-27B
Adapted by Pulseee
Planned public releaseHugging FaceAfter evaluation is complete
- Domain pretrainingComplete
- Fine-tuningIn progress
- EvaluationPlanned
- Public releasePlanned
Our planned evaluation approach
We plan to combine technical evaluation with review by lawyers and subject matter experts familiar with the Fair Work Act. The aim is to understand where the model performs well, where it falls short and what needs further work. Methods and limitations will accompany the findings.
From the lab.
Research papers, project notes and benchmarks. Freely available when published.
Research in progress
The work is underway.
Our first research papers and benchmarks will appear here as they are ready to share. In the meantime, explore Project OWL and the questions behind it.
Explore the active project ↑Human-led analysisMethods alongside findingsLimitations made clear
Curious about the same questions?
Have an idea, a useful test case or a research perspective to share? We’d like to hear from you.