About the role
Post-training turns a capable base model into one people can rely on. The team works on supervised fine-tuning, preference optimisation, reward modelling and the data that drives all three. The central problem is shaping behaviour without eroding capability: a model that is helpful, calibrated and well-behaved, and still sharp on hard tasks.
Post-training is where Clexerl's research becomes a product experience. Clexerl Go's agents depend on instruction following, tool-call reliability and sensible refusals, and those are properties we train and measure here. The team works closely with Alignment on honesty and steerability, and with Evaluations so that every improvement is verified against something independent.
Work is organised around hypotheses about what the model gets wrong. Scientists read transcripts, form a theory, design a data or training intervention and measure the result. Reward hacking and regressions are treated as findings, and written up as such.
What you’ll do
- Design post-training recipes that combine supervised fine-tuning, preference optimisation methods such as DPO and RLHF or RLAIF, and rejection sampling.
- Train and analyse reward models, and study their failure modes, including over-optimisation, length bias and sycophancy.
- Read model transcripts closely, categorise failures and turn each category into a targeted data or training intervention.
- Work with data and evaluation colleagues to build preference and demonstration datasets with measurable quality.
- Measure capability regressions after every training change and diagnose where the model lost ground.
- Write up findings with enough rigour that another team can reproduce the training recipe.
What you’ll bring
- Research experience in fine-tuning or aligning large language models, shown through papers, open-source work or shipped systems.
- Understanding of policy-gradient and preference-based objectives, including KL regularisation and why reference-model drift matters.
- Experience building or critiquing reward models and judging how well a proxy tracks what people actually want.
- Hands-on skill training models in PyTorch or JAX, with the debugging patience that long fine-tuning jobs demand.
- Strong evaluation instincts: you separate real improvements from artefacts of the judge, the prompt set or the sampling temperature.
- Willingness to read hundreds of model outputs by hand and trust what that tells you.
Nice to have
- Experience with constitutional or principle-based methods for generating training signal.
- Work on calibration, uncertainty expression or abstention in language models.
- Prior research on model distillation or on-policy data generation.
- Experience with multilingual or multimodal post-training.
Work you might take on
- 01
Build a reward model ensemble and measure how much it reduces over-optimisation compared with a single scorer.
- 02
Diagnose why a fine-tuned model became verbose and sycophantic, then fix it with a data change instead of a loss penalty.
- 03
Design a preference-data protocol for tool-using agents that rewards correct calls and penalises unnecessary ones.
- 04
Run an ablation on the mix of supervised and preference data and report its effect on capability retention.
Compensation
$340K – $560K
The US base salary range for this role. Every offer also includes equity in Clexerl and, when it spins out, in the venture you help build. We tell you the reasoning behind your number.
Benefits
Health, fully covered
Medical, dental and vision for you and your dependants, premiums paid in full.
Two kinds of equity
Equity in Clexerl, plus a stake in the venture you help build when it spins out.
22 weeks parental leave
Fully paid, for every parent, however your family arrives.
Rest that sticks
Flexible time off with a four-week minimum we actually check on, plus a company-wide week in December.
Learning and compute
A yearly budget for courses, conferences and books, and compute credits for side projects.
401(k) with a 4% match
Matched from your first paycheque and vested immediately.
How we hire
Apply 5 minutes
Send links to what you have built and a few lines on why this role. No cover letter needed.
Intro call 30 minutes
A conversation with the hiring manager about the work, the team and what you want next.
Work sample 2–4 hours
A paid take-home or a pairing session on a real-shaped problem. You pick the format.
Studio day One day
Meet the team, go deep on your work and ours. Remote or in person.
Offer Within 5 days
A written offer with the full range, the equity and the reasoning behind both.
Opens an email with a short template. Links beat résumés.