LLM Fine-Tuning Engineer (Open-Weight Models / Secure Environments)

Trenchant; United States; Remote; USD 120,000 - 180,000; Full time

Job summary

Trenchant is hiring an LLM Fine-Tuning Engineer to adapt open-weight language models for secure, local environments. You will own the post-training lifecycle including dataset design, supervised fine-tuning, and deployment support.

You will collaborate with cybersecurity teams to build high‑quality datasets, generate synthetic data, compare tuning methods (SFT, LoRA, QLoRA), and evaluate models for reliability beyond headline metrics.

Qualifications

• Strong experience fine-tuning LLMs in production or serious research.
• Practical experience with multiple tuning approaches: SFT, LoRA, QLoRA.
• Experience building and curating datasets for post-training.
• Experience generating and validating synthetic data for model training.
• Strong Python skills and GPU/memory optimization knowledge.

Responsibilities

• Fine-tune open-weight LLMs for secure/local use.
• Design, run, and compare post-training approaches.
• Build and improve high-quality training/evaluation datasets.
• Generate synthetic data to expand coverage and robustness.
• Assess model quality beyond headline metrics.
• Collaborate with engineers to deploy tuned models.

Skills

Fine-tuning LLMs

Tuning approaches

Python

GPU optimization

Debugging mindset

Tools

SFT

LoRA

QLoRA

Job description

We are hiring an LLM Fine-Tuning Engineer to help us adapt and evaluate open-weight language models for use in secure environments.

This role is focused on the full post-training lifecycle: dataset design, supervised fine-tuning, parameter-efficient tuning, synthetic data generation, evaluation, and deployment support. You will work closely with technical teams operating in advanced cybersecurity contexts, helping turn foundation models into reliable task-specific systems.

IMPORTANT NOTE : We are not looking for prompt engineers or “ LLM whisperers.” We are looking for someone who understands post-training, data evaluation, and deployment deeply enough to make open-weight models reliable in real operational environments.

What You’ll Do

• Fine-tune and adapt open-weight LLMs for specialized use cases in secure/local environments

• Design, run, and compare different post-training approaches, including:

• Build and improve high-quality datasets for training and evaluation

• Generate synthetic data and use it responsibly to expand coverage, improve robustness, and accelerate iteration

• Assess model quality beyond headline metrics

• Work with engineering teams to operationalize tuned models for local or restricted deployments

• Collaborate with domain experts to translate real operational needs into measurable model requirements

What We’re Looking For

• Strong experience fine-tuning LLMs or adjacent foundation models in production or serious research environments

• Practical experience with multiple tuning approaches, ideally including SFT, LoRA, QLoRA

• Experience building and curating datasets for post-training

• Experience generating and validating synthetic data for model training

• Strong Python skills and solid experience with:

• Good understanding of GPU constraints, memory/performance tradeoffs, quantization-aware workflows and practical training optimization

• Strong debugging mindset and ability to investigate why a model improved, regressed, or failed

Nice to Have

• Experience with secure, air-gapped, or otherwise restricted deployment environments

• Experience deploying or serving tuned models

• Background in cybersecurity, especially offensive security, vulnerability research

• Experience with distributed training frameworks

How We Think About the Role

This is not a “prompt engineer” role, and it is not limited to running a few LoRA jobs. We want someone who can think deeply about:

• when fine-tuning is the right tool versus prompting or orchestration

• how to create data that improves behavior instead of just inflating metrics

• how to evaluate models in ways that reflect real operational value

• how to make open-weight models reliable in constrained environments