ML researcher in LLM post-training

грейд
middle
формат
удалённо
язык
Английский B1
опубликована
вчера

описание

We’re the ML team at Toloka. We post-train open-weight models with fine-tuning and RL, so that small models match or beat frontier ones on a specific task at a fraction of the cost. We do it for enterprise clients, for the datasets Toloka sells, and for our own production systems. You’ll run short, self-contained research projects on the platform and publish each one as a blog post under your name. A project typically takes 20–40 hours from idea to draft.

задачи

  • Reproduce results that are making noise. When a paper or post takes off (weight-space programming is a recent example), rerun it on our platform, add an ablation of your own, and report what holds up.
  • Run autoresearch studies. Point a coding agent at a training problem, let it run dozens of experiments, and write up what it tried, what worked, and where it got stuck.
  • Do small head-to-head case studies. Take one well-defined task, train on our platform and on alternatives, and compare quality, cost, and effort.
  • Write it up. Each project ends in a blog post that shows the full path: the experiments you ran, the ones that failed, and how you got to the final number. We help with editing and charts.
  • Tell us what broke. You’ll be one of the first outside users of the platform, and your feedback goes straight to the people building it.

требования

  • Hands-on experience training or fine-tuning LLMs (PyTorch, the Hugging Face stack, LoRA/SFT).
  • You can take a project from a vague idea to a result on your own: design the experiment, run it, and tell a real effect from noise.
  • You already use coding agents (Claude Code, Codex) in your work.
  • Clear technical writing in English. Please include a link to something you’ve written or built: a blog post, a paper, or a repo.
  • Будет плюсом: RL experience is a plus

условия

  • Freelance (B2B) contract via Deel, $55–60/hour.
  • Part-time, about 40 hours a month, fully remote. The schedule is flexible, and when a topic is hot, we’d rather you spend the hours in a one- or two-week sprint.
  • An initial 3-month engagement, with the option to extend.
  • Access to our post-training platform and GPU compute for your experiments.
  • Your research published on the Toloka blog under your name.
  • Direct work with the researchers and engineers who build the platform.

навыки

компания

похожие вакансии

статьи для разработчиков