whitecircle - ML Research Engineer

• Train and post-train LLMs for safety and moderation tasks: SFT, RLHF, DPO, and related alignment methods • Build and train reward models from human and synthetic preference data • Design and run high-throughput data pipelines: collection, synthetic generation, filtering, deduplication, and quality control at very large scale • Run distributed training on multi-GPU clusters and debug what goes wrong when it does • Build evaluation systems and benchmarks that actually measure model behavior, and use them to drive training decisions • Optimize models for production inference: quantization, speculative decoding, serving with vLLM/TensorRT or similar • Move fast from experiment to production – your models ship, and you see their effect on real traffic • You'll fit right in if you: • Have hands-on experience with modern LLM post-training – SFT, RLHF, DPO, or related methods – on models you trained yourself • Have worked with data at genuinely large scale: building pipelines for training corpora, preference data, or synthetic data generation • Have trained models on distributed multi-GPU setups and are comfortable in PyTorch or JAX • Have built or worked with reward models and preference data • Understand evaluation deeply: you know why benchmarks lie, and how to build ones that don't • Have experience optimizing inference: quantization, speculative decoding, vLLM, TensorRT, Triton, or similar • Are strong in Python and comfortable with SQL-like data tooling for large-scale data work • Have a strong ownership mindset: you can take an ambiguous modeling problem, make it concrete, ship a working model, and improve it from real feedback • A public builder footprint: open-source models, datasets, or training frameworks on HuggingFace/GitHub, benchmarks, papers (workshop or main conference), or technical posts with real usage • Experience training models at a frontier or near-frontier lab, or leading open-source model releases with documented adoption • Experience with RL methods for LLMs beyond standard RLHF: online RL, GRPO-style methods, or novel alignment approaches • Experience with moderation, safety, or classification models at scale • Multilingual model training experience