---
license: mit
---

This repository contains the trained **PersonalizedRouter** model weights saved as a `.pth` file.

In the project files, the suffix `v1` refers to the `Multi-cost-efficiency Simulation Strategy` described in the paper, while `v2` refers to the `LLM-as-a-Judge Simulation Strategy`.

For `best_model_v1.pth`, the model was trained on an interaction dataset generated by 10 LLMs, 240 queries, and 9 different performance and cost settings.

For `best_model_v2.pth`, the model was trained on an interaction dataset generated by 10 LLMs, 240 queries, and preferences from 9 different user groups.