DCRouter (training + inference)

DCRouter is a stronger, heavier router that uses a transformer backbone. It is usually more compute-intensive than embedding-only baselines.

Notebooks: - Training: https://github.com/ulab-uiuc/LLMRouter/blob/main/notebooks/dcrouter/01_dcrouter_training.ipynb - Inference: https://github.com/ulab-uiuc/LLMRouter/blob/main/notebooks/dcrouter/02_dcrouter_inference.ipynb

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

Configs

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

Checkpoint loading

DCRouter loads checkpoints from the directory containing model_path.save_model_path (it looks for best_model.pth or best_training_model.pth). Make sure you ran training at least once before expecting good results.

Run (CLI)

Train:

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

Route-only inference:

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

Full inference:

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

Tips

• model_path.backbone_model controls the transformer backbone.
• For full inference, set API_KEYS and ensure llm_data has api_endpoint and model.