Large Language Model Algorithm Engineer Midtrain Agentic Data

Skills

About the Role

You will research and develop data for the Midtrain stage of large language models. You will identify model failure patterns from user signals, create correction methods, scale user-data-driven trajectories, and optimize agentic-data mixes. You will build measurable data-quality, evaluation, and error-analysis systems that connect data design, training, and real-world performance.

Requirements

• Have research or practical experience in large-model data, LLM agents, data mixing, quality evaluation, or related fields
• Have published high-quality papers at international conferences or journals
• Use Python and large-scale data-processing frameworks such as Spark and Ray
• Have experience leading or contributing to impactful large-model machine-learning projects
• Design rigorous experiments and derive reliable conclusions from data

Responsibilities

• Identify model failure patterns from user interaction signals and build detection, attribution, and correction workflows
• Create filtering, loss masking, rewriting, and synthetic correction trajectories and validate them through training experiments
• Research methods to clean, select, rewrite, and expand user-data-driven training data
• Study how data scale, diversity, and quality affect model capability scaling
• Analyze and optimize the distribution and mixing of code, tool-use, search, and long-horizon task data
• Define agentic-data quality standards for correctness, process quality, diversity, safety, and compliance
• Implement automated quality inspection using rules, model scoring, and execution validation
• Maintain data deduplication, decontamination, provenance, and version management
• Build agentic-capability evaluation and error-analysis systems and drive iterative improvements across data, training, and evaluation