LLM Fine-Tuning Engineer — San Francisco, CA

About the Role

We are seeking an LLM Training Data Engineer who specializes in teaching large language models new domain expertise through carefully designed training examples. The focus is on deciding what the model must learn, developing effective instruction/response pairs, and evaluating whether the intended knowledge and behaviors were learned. We especially value candidates who have successfully designed training datasets across multiple companies, products, or substantially different domains. This is not an AI infrastructure or model research position.

Key Responsibilities

• Learn unfamiliar technical and business domains rapidly; work with subject-matter experts to identify the concepts, relationships, rules, and behaviors a model must learn.

• Design high-quality instruction/response and input/output training pairs, including examples for edge cases and structured outputs.

• Build domain coverage taxonomies and reusable dataset-design methods rather than relying on example volume alone.

• Create, curate, validate, and refine training and synthetic examples; identify gaps, contradictions, ambiguity, and duplication.

• Design evaluation sets, analyze model errors, and iteratively improve training examples based on observed results.

• Use Python or comparable tools to generate, transform, validate, and manage large structured datasets.

Required Qualifications

• Demonstrated hands-on experience designing training datasets used to teach or fine-tune LLMs.

• Strong understanding of training-pair quality, diversity, coverage, and their effects on model behavior.

• Ability to independently determine which training examples are needed—not merely label or format examples specified by others.

• Experience with instruction/response pairs, supervised training datasets, synthetic examples, and model-output error analysis.

• Strong Python or equivalent data-processing skills, analytical judgment, and attention to detail.

Strongly Preferred

• Successful LLM training-data design across two or more distinct companies, domains, products, or problem areas.

• Concrete examples of how the candidate adapted their training-data strategy across different domains.

• Experience teaching models specialized schemas, metadata, XML, JSON, code, configuration, or other structured representations.

• Experience designing dataset coverage frameworks and evaluations that demonstrate actual learning and generalization.

Not Required

RAG, vector databases, retrieval systems, RLHF, DPO, reinforcement learning, LoRA/QLoRA implementation, distributed GPU training, model serving, inference infrastructure, or foundation-model research are not requirements. Familiarity with supervised fine-tuning is useful, but this position owns training-data design—not training infrastructure.

What Success Looks Like

The model accurately learns and applies the intended domain expertise. The work follows an iterative cycle: understand the domain → define learning objectives → design examples → evaluate results → identify gaps → improve the dataset. Success is measured by the quality and transferability of what the model learns, not the sophistication of the infrastructure.

Compensation

Expected base salary: $150,000–$190,000 annually, depending on relevant experience and demonstrated success in multi-domain LLM training-data design.