---
license: apache-2.0
base_model: Qwen/Qwen2.5-0.5B-Instruct
library_name: transformers
tags:
- orchestration
- routing
- qwen
- logits-head
pretty_name: mempool Qwen Logits Orchestrator Smoke
---

# mempool Qwen Logits Orchestrator Split Smoke

This repository contains a split-smoke checkpoint for the `mempool`
Qwen-small 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 smoke loss: `4.2856650959770635`

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: `0.4716981132075472`
- Workflow accuracy: `0.5660377358490566`
- Mean worker loss: `1.4386708387788736`
- Mean workflow loss: `0.7223093554790501`

Held-out evaluation:

- Worker accuracy: `0.3076923076923077`
- Workflow accuracy: `0.7692307692307693`
- Mean worker loss: `1.7572260361451368`
- Mean workflow loss: `0.5600809453485104`

This is a smoke artifact, not a promoted production policy.