LLM Router Public Artifacts

Files

• checkpoints/prefill_router_combined.pt (default, 11.74 MiB): Active prefill cost-router checkpoint trained from public verified trace labels.
• checkpoints/session_health_public.pkl (default, 1.87 MiB): Active v2 session-health checkpoint trained from public HF trajectory windows.
• configs/combined-pool.yaml (default, 0.00 MiB): Active router pool config.
• configs/litellm-combined.yaml (default, 0.00 MiB): Active LiteLLM model config.
• docs/SESSION_HEALTH_ROUTER.md (default, 0.00 MiB): Session-health training and runtime notes.
• embedding/complexity_model/config.json (embedding, 0.00 MiB): Optional embedding-router bundle trained from public WildChat traces.
• embedding/complexity_model/embedder.onnx (embedding, 224.16 MiB): Optional embedding-router bundle file.
• embedding/complexity_model/embedder_hf_config.json (embedding, 0.00 MiB): Optional embedding-router bundle file.
• embedding/complexity_model/eval_report.md (embedding, 0.00 MiB): Optional embedding-router bundle file.
• embedding/complexity_model/parity_fixture.jsonl (embedding, 0.07 MiB): Optional embedding-router bundle file.
• embedding/complexity_model/special_tokens_map.json (embedding, 0.00 MiB): Optional embedding-router bundle file.
• embedding/complexity_model/tokenizer.json (embedding, 16.29 MiB): Optional embedding-router bundle file.
• embedding/complexity_model/tokenizer_config.json (embedding, 0.00 MiB): Optional embedding-router bundle file.
• embedding/complexity_model/weights.safetensors (embedding, 0.20 MiB): Optional embedding-router bundle file.

Provenance

• prefill_router_combined.pt is the active prefill router checkpoint used by configs/combined-pool.yaml. It was built from public verified trace labels and the callable model pool in that config.
• session_health_public.pkl is the active v2 session-health checkpoint. Its embedded report points at data_public/session-health-windows-v2.csv, built from public SWE-bench and TerminalBench Hugging Face traces.
• embedding/complexity_model/ is the optional embedding-router classifier bundle trained from public WildChat conversations labeled by an OpenAI judge. It is used only when routing through the embedding-routed alias.

Use