keonkim/coreb-task-type-router-f2llmv2-330m-c2llm-7b

CoREB task-type router: F2LLM-v2-330M + C2LLM-7B

This is a deterministic, training-free MTEB router for code retrieval:

• —MTEB Retrieval tasks route to `codefuse-ai/F2LLM-v2-330M` at revision e8ef9a8eb907a9dffdd9442424a967ba73e70d31.
• —MTEB Reranking tasks route to `codefuse-ai/C2LLM-7B` at revision c1dc16d6d64eb962c783bfb36a6d9c2f24a86dca.

The router reads only the coarse public TaskMetadata.type field. It does not inspect the benchmark name, examples, labels, or scores. Evaluation keeps only one child model resident at a time.

Important disclosure

The composition was selected after comparing public CoREB results and is therefore leaderboard-adapted. CoREB exposes only a test split. The submitted result should be interpreted as the performance of this transparent routing policy, not as evidence of zero-shot model selection.

Embedding dimensions

The output dimension depends on the route:

Each individual evaluation task has a fixed output dimension. Current MTEB ModelMeta has no representation for task-dependent dimensions, so the model implementation leaves embed_dim unset.

CoREB v1 results

All six results were newly evaluated through the router with MTEB 2.18.6 on an NVIDIA L4. Retrieval used bfloat16, batch size 4, and F2LLM's configured FlashAttention 2 implementation. Reranking used bfloat16, batch size 2, and C2LLM's official attention configuration.

Reproduction

The MTEB model implementation pins both child model revisions and delegates to each child's official MTEB loader configuration. In particular, the F2LLM Retrieval route uses its configured bfloat16/FlashAttention 2 implementation, while the C2LLM Reranking route keeps C2LLM's official adapter defaults. Evaluate all six CoREB v1 tasks with MTEB 2.18.6 and CUDA. The submission includes per-task MTEB JSON results, an environment record, and the complete run log.