blazeofchi/mempool-qwen-logits-orchestrator-v0

mempool Qwen Logits Orchestrator

The latest intended coordinator backbone for this repository is Qwen/Qwen3-0.6B, the current official Qwen 0.6B text-generation model.

The repository stores routing-head artifacts for the mempool Qwen-small logits-head orchestrator path with a deterministic held-out gate. The checkpoint stores only trained routing heads, not Qwen base model weights. A refreshed checkpoint should be trained and uploaded before using these heads with the latest Qwen 0.6B backbone.

• —Checkpoint: qwen_logits_heads.pt
• —Training rows: 53
• —Final training loss: 2.1550024236951555

Worker labels:

• —ollama-cloud-deepseek-v4-pro
• —ollama-cloud-glm-5.2
• —ollama-cloud-kimi-k2.7-code
• —ollama-cloud-qwen3-coder-480b

Train-row evaluation:

• —Worker accuracy: 0.7358490566037735
• —Workflow accuracy: 0.9433962264150944
• —Mean worker loss: 1.124755915605797
• —Mean workflow loss: 0.142171498247475

Held-out evaluation:

• —Worker accuracy: 0.5384615384615384
• —Workflow accuracy: 0.7692307692307693
• —Mean worker loss: 1.3366980185875525
• —Mean workflow loss: 0.672384113073349

A sample router prediction is included at sample_prediction.json.

This is a v0 research artifact. It is usable for routing experiments, but it is not yet a promoted production policy.