We are looking for a Senior AI Data & Distillation Engineer to lead the mathematical transfer of advanced programming logic from frontier teacher models into our independent 26B student models. You will own the objective functions that teach the model "how to think" rather than just memorising code. Additionally, you will build data-streaming pipelines that process hundreds of billions of code tokens spanning Java, .NET, Unix, and legacy COBOL.
Core Responsibilities:
* Implement multi-stage distillation objective functions utilizing Kullback-Leibler (KL) Divergence, logit extraction, and temperature scaling to align student-teacher outputs.
* Build high-throughput synthetic dataset generation pipelines powered by frontier LLM APIs to create multi-turn, token-dense structural reasoning paths.
* Audit and customize tokenizer vocabularies to ensure critical programming syntax (e.g., indents, braces, tabs,
rare legacy mainframe structures) is preserved without fragmenting into inefficient byte chunks.
* Use tools like Ray or Apache Arrow to deduplicate, clean, and stream massive linguistic datasets into the training loop, keeping the GPUs fed without creating data bottlenecks.
Required Technical Skills:
* ML Mathematics: Robust foundational understanding of probability distributions, cross-entropy minimization, and logit-level distillation patterns.
* Data Scaffolding: Advanced experience with Ray, Apache Arrow, and the Hugging Face ecosystem (Tokenizers, Datasets).