blazeofchi/mempool-qwen3-0p6b-logits-orchestrator-v1

mempool Qwen3 0.6B Logits Orchestrator v1

This repository contains a usable checkpoint for the mempool Qwen 0.6B logits-head orchestrator path with a deterministic held-out gate.

The checkpoint stores only the trained routing heads, not the Qwen base model weights. Load the base model separately and attach the heads.

• —Checkpoint: qwen_logits_heads.pt
• —Training rows: 53
• —Final training loss: 2.474162220954895

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: 0.8113207547169812
• —Workflow accuracy: 0.8867924528301887
• —Mean worker loss: 1.2073683018954295
• —Mean workflow loss: 0.23352967146432624

Held-out evaluation:

• —Worker accuracy: 0.6923076923076923
• —Workflow accuracy: 0.7692307692307693
• —Mean worker loss: 1.3018739131780772
• —Mean workflow loss: 0.6372018387684455

A sample router prediction is included at sample_prediction.json.

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