Computing Internship - Comparative Neural Architecture Research: Supervised Fine-Tuning (SFT) vs. Continued Pre-training on Vertex AI

24 Apr 2026

This position is offered through the ANU Computing Internship courses (COMP4820 / COMP8830).

Semester 2, 2026 applications open on Monday 18th May 2026 and close on Sunday 31st May 2026. 

Company

Tailored Accounts

Tailored Accounts is an award-winning management accounting and business intelligence company. Our team combines a deep passion for numbers with a personalised service approach to ensure the best financial outcomes for every client.

Project

This project is an independent Research & Development initiative designed to benchmark the efficiency of various model adaptation strategies within the Google Cloud Vertex AI ecosystem. The intern will explore the “Knowledge vs. Behaviour” trade-off: specifically, whether it is more effective to teach a model new domain knowledge through Continued Pre-training (unsupervised) or to refine its task-specific performance through Fine-Tuning (supervised).

The intern will build an end-to-end experimental framework using open-weights models and public datasets. This is a greenfield project where the intern develops the methodology, the integration code, and the final performance analysis, retaining full ownership of the resulting IP.

Key Objectives

  1. Experimental Design: Define a hypothesis comparing the accuracy gains of Parameter-Efficient Fine-Tuning (PEFT/LoRA) against Full-Parameter Continued Training on specialised public-domain corpora.

  2. Vertex AI Pipeline Orchestration: Use Python to script and automate training jobs in Vertex AI, managing hyperparameters like learning rates, rank (for LoRA), and batch sizes.

  3. Cross-Language Middleware: Develop a standalone C# / .NET “Experiment Controller.” This application will interface with the Vertex AI API to trigger training runs, monitor job status, and fetch evaluation metrics.

  4. Benchmarking & Distillation: Evaluate if “distilling” a large fine-tuned model into a smaller, cheaper-to-run model (e.g., Gemini Pro to Gemini Flash) maintains sufficient accuracy for specific categorisation tasks.

Required technical skills 

AI/ML Platforms: Deep dive into Vertex AI Model Garden, Generative AI Studio, and Vertex AI Pipelines.

Languages: * C# / .NET (For building the experiment management system and interacting with Google Cloud Client Libraries) & Python (For data wrangling, loss curve visualisation, and executing ML scripts.)

Methodologies: Hands-on experience with LoRA (Low-Rank Adaptation) and Supervised Fine-Tuning (SFT).

Data: Handling large-scale structured and unstructured data using JSONL and Google Cloud Storage.

Required/preferred professional and other skills 

Research Autonomy: Ability to lead a project from hypothesis to conclusion without relying on proprietary internal systems.

Computational Economics: Learning to balance model performance against the cost of GPU/TPU hours on cloud infrastructure.

Technical Communication: Producing a comprehensive “Optimisation Report” that documents which tuning methods yielded the best results (a valuable asset for the intern’s portfolio).

Architectural Cleanliness: Designing a system that is decoupled from any specific business logic, ensuring the code is portable and reusable.

Delivery Mode 

In-Person Internship

Project’s Special Requirements/ Conditions 

None

Type of internship 

Educational internship (unpaid)

How to apply 

Applications are invited from eligible students to apply for the Computing Internship courses COMP4820 or COMP8830. Eligibility details of COMP4820 / COMP8830 and further information about the Computing Internship can be found on the Computing Internship page

Eligible students can apply through the Computing Internship application form which will be available via the Computing Internship page between Monday 18th May 2026 and close on Sunday 31st May 2026. 

You can nominate multiple preferred Internship projects and host organisations through the one online application form. 

Eligibility and room available in degree to undertake COMP4820/COMP8830 will be assessed at the time of application. If you do not meet the eligibility criteria or do not have room in your degree to fit COMP4820/COMP8830, your application will not be progressed. 

Your application will require you to upload the following documents: 

  • an updated copy of your resume, and 
  • an expression of interest (limit 350 words) for each organisation you wish to apply to (organisations with multiple projects may only submit one expression of interest, so state clearly which project/s you wish to be considered for).
arrow-left bars magnifying-glass xmark