About the role
Matilda is Australia's LLM. What ends up in the corpus is what the model learns, so the quality of the data sets the ceiling on the quality of the model.
We're hiring a Signal Engineer to own that ceiling. You will build the pipelines that turn massive, messy, raw data into the dataset Matilda trains on. The work is part engineering, part editorial judgment, done in code.
A lot of the real gains in frontier models come from the data, and most of that work is underinvested in across the field. It is one of the highest-leverage places you can spend your time as an engineer.
What you'll work on
- Pipelines that ingest, clean, dedupe, filter, and score training data at TB to PB scale
- Quality classifiers and heuristics that separate useful data from the rest
- Dataset mixture design and experiments on what actually improves the model
- Tools to explore, sample, and audit what's in the corpus
- Close work with researchers and training engineers so data choices connect to model behaviour
What we're looking for
- Strong engineer. Python, data tooling, distributed processing, clean pipelines.
- High attention to detail. Small errors compound fast at this scale.
- Taste and judgment about what good training data looks like.
- Comfort working with very large, very messy datasets.
- Curiosity about how data shapes model behaviour.
- High learning velocity. You don't need a PhD or prior LLM experience.
Nice to have
- Experience with web-scale corpora or pretraining data pipelines
- Experience working with unstructured text data
- Familiarity with distributed data frameworks (Spark, Ray, or similar)
- Exposure to deduplication, quality classification, or tokenisation
Note
Full-time role based in Melbourne, working closely with our in-person team. At this time we are not able to offer visa sponsorship, so applicants must have existing and unrestricted work rights in Australia.
