Machine Learning Researcher, LLM Post-Training

Thurn Partners; New York (NY); On-site; USD 250,000 - 380,000; Full time

Job summary

Thurn Partners is seeking a Machine Learning Researcher in New York to adapt frontier-scale language models to finance data. You will post-train models on petabyte-scale proprietary text and tabular data, and own alignment and fine-tuning end-to-end, from SFT to RLHF, leveraging integral training infra.

The role demands deep transformer knowledge, strong Python with PyTorch or JAX, and a track record of rigorous evaluation. PhD preferred; shipping results counts as much as publications.

Qualifications

• Hands-on experience post-training large language models at a research lab, big tech group or applied AI team shipping real systems.
• Deep understanding of transformer architectures and modern alignment techniques.
• Strong Python with PyTorch or JAX, and the low-level GPU fluency to exploit large clusters efficiently.
• A track record of rigorous evaluation: you can prove a model got better, not just feel it.
• PhD or equivalent research depth preferred; publications welcome, shipped results count just as much.
• No finance background required. Genuine curiosity about applying frontier ML to a new domain is.

Responsibilities

• Post-train and domain-adapt state-of-the-art large language models on proprietary financial text and tabular data at petabyte scale.
• Own alignment and fine-tuning end to end: SFT, DPO, RLHF and parameter-efficient methods (LoRA/PEFT), plus the training infrastructure behind them.
• Build rigorous evaluation pipelines for reasoning, quantitative accuracy and strict factuality, in a setting where model outputs inform real decisions within days.
• Work at the metal when it matters: GPU memory management, mixed precision (FP16/BF16), quantisation and parallelisation strategies across a very large cluster.
• Collaborate closely with a small group of researchers and a dedicated research platform team, with a compute-to-researcher ratio few laboratories anywhere can match.

Skills

Python

PyTorch

JAX

Transformer models

GPU computing

Education

PhD or equivalent research depth

Tools

CUDA

RLHF training

LoRA/PEFT

Job description

A leading global quantitative trading firm is hiring a Machine Learning Researcher into its central AI group in New York, adapting frontier-scale language models to one of the richest proprietary datasets in finance. The group's remit is simple to state and hard to do: take the best open-weight models in the world and make them genuinely expert at quantitative reasoning over petabyte-scale text and tabular data, on one of the largest private accelerator estates in the industry.

What you'll do:

• Post-train and domain-adapt state-of-the-art large language models on proprietary financial text and tabular data at petabyte scale.
• Own alignment and fine-tuning end to end: SFT, DPO, RLHF and parameter-efficient methods (LoRA/PEFT), plus the training infrastructure behind them.
• Build rigorous evaluation pipelines for reasoning, quantitative accuracy and strict factuality, in a setting where model outputs inform real decisions within days.
• Work at the metal when it matters: GPU memory management, mixed precision (FP16/BF16), quantisation and parallelisation strategies across a very large cluster.
• Collaborate closely with a small group of researchers and a dedicated research platform team, with a compute-to-researcher ratio few laboratories anywhere can match.

Your profile:

• Hands-on experience post-training large language models at a research lab, big tech group or applied AI team shipping real systems.
• Deep understanding of transformer architectures and modern alignment techniques.
• Strong Python with PyTorch or JAX, and the low-level GPU fluency to exploit large clusters efficiently.
• A track record of rigorous evaluation: you can prove a model got better, not just feel it.
• PhD or equivalent research depth preferred; publications welcome, shipped results count just as much.
• No finance background required. Genuine curiosity about applying frontier ML to a new domain is.

Why this role:

Frontier-lab-scale compute without frontier-lab queue politics, private data no model has ever been trained on, and a feedback loop measured by the real world in days rather than by benchmark leaderboards. The work is pure modern ML - post-training, alignment, evaluation - applied where it compounds fastest, with compensation at the very top of the quantitative industry.

Pre-Application:

• This is a full-time, on-site role based in New York; fully remote candidates will not be considered.
• Applicants must have the right to live and work in the US, or be eligible for sponsorship (confirmed case by case).