ShadowLM Trainer
A fine-tuning SDK that captures an AI agent's real traffic, trains a small open model to match it, and gradually shifts the agent from a costly frontier model to the model you now own.
ShadowLM moves one task an agent performs on a costly, hosted frontier model onto a small open model the developer owns, without changing the agent itself: the agent keeps calling the same endpoint, only the model behind it changes. A drop-in capture proxy records the agent's real traffic into trajectories, an LLM judge scores whole episodes, and the model is then fine-tuned on that data, so the resulting shadowLM runs in parallel with the frontier model until it performs as well and can take over.
It exposes twelve or more training methods, including LoRA, QLoRA, DoRA, full fine-tuning, DPO and GRPO, behind one method argument, and can also build a training set directly from existing OpenTelemetry GenAI trace data without a capture proxy. It installs with pip and is maintained by Lyzr Research Labs; a separate paid product, ShadowLM Studio, wraps the same engine into a one-click migration workflow.