SVMRouter (training + inference)

SVMRouter is a classic classifier router that works well with high-dimensional embeddings.

Notebooks: - Training: https://github.com/ulab-uiuc/LLMRouter/blob/main/notebooks/svmrouter/01_svmrouter_training.ipynb - Inference: https://github.com/ulab-uiuc/LLMRouter/blob/main/notebooks/svmrouter/02_svmrouter_inference.ipynb

Router docs: https://github.com/ulab-uiuc/LLMRouter/blob/main/llmrouter/models/svmrouter/README.md

Configs

• Train: https://github.com/ulab-uiuc/LLMRouter/blob/main/configs/model_config_train/svmrouter.yaml
• Test: https://github.com/ulab-uiuc/LLMRouter/blob/main/configs/model_config_test/svmrouter.yaml

Note

The inference notebook writes a minimal svmrouter_inference.yaml for convenience. For the CLI, you can also use the repo test config above.

Run (CLI)

Train:

llmrouter train --router svmrouter --config configs/model_config_train/svmrouter.yaml

Route-only inference:

llmrouter infer --router svmrouter --config configs/model_config_test/svmrouter.yaml --query "Explain transformers." --route-only

Full inference:

llmrouter infer --router svmrouter --config configs/model_config_test/svmrouter.yaml --query "Explain transformers."

What to tweak

Tune the hparam section in the YAML (kernel/regularization) and the metric.weights trade-offs.