---
title: pm-skills-router
canonical_url: "https://www.modelscope.ai/models/mohitagw15856/pm-skills-router"
md_url: "https://www.modelscope.ai/models/mohitagw15856/pm-skills-router.md"
repository: mohitagw15856/pm-skills-router
last_updated: 2026-10-05
license: mit
pipeline_tag: text-classification
tasks:
  - text-classification
library_name:
  - other
frameworks:
  - other
language:
  - en
  - zh
downloads: 3
stars: 0
tags:
  - agent-skills
  - routing
  - retrieval
  - chinese
---

# pm-skills-router

> pm-skills-router - An open-source model by mohitagw15856 on ModelScope. A small, dependency-free router for the PM Skills library: give it a request in English or Chinese and it returns the best-matching skills. TF-IDF nearest-centroid over 1,234 skills,…

mohitagw15856/pm-skills-router is a text-classification model on ModelScope. licensed under mit.

- **Repository**: mohitagw15856/pm-skills-router
- **License**: mit
- **Tasks**: text-classification
- **Tags**: agent-skills, routing, retrieval, chinese
- **Downloads**: 3
- **Stars**: 0
- **Last updated**: 2026-10-05

Source: https://www.modelscope.ai/models/mohitagw15856/pm-skills-router

---

# pm-skills-router

A small, dependency-free router for the [PM Skills](https://github.com/mohitagw15856/pm-claude-skills) library. Give it a request in English or Chinese and it returns the best-matching skills from 1,297 skills. It runs anywhere Python runs: no GPU, no network, standard library only.

一个小巧、零依赖的技能路由模型：输入中文或英文的请求，返回 PM Skills 技能库中最匹配的技能。不需要 GPU，不联网，只用 Python 标准库。

## Use it 使用

```python
from modelscope import snapshot_download
import sys
path = snapshot_download('mohitagw15856/pm-skills-router')
sys.path.insert(0, path)
from pm_router import Router

router = Router.load(f"{path}/router.json")
router.route("帮我把这些笔记整理成周报", k=3)
# [('cn-weekly-report', 0.34...), ('dingtalk-work-log', ...), ...]
router.route("my landlord is keeping my deposit", k=3)
```

Then load the matching skill's `SKILL.md` from the library (npm `pm-claude-skills`, PyPI `pm-skills`, or the [Gitee mirror](https://gitee.com/mohitagw/pm-claude-skills)) and give it to your model as instructions.

## How it works 原理

TF-IDF nearest centroid. Latin text becomes word unigrams and bigrams; Chinese becomes character bigrams. Each skill's centroid is built from its trigger phrases (the [pm-skills-instruct](https://www.modelscope.ai/datasets/mohitagw15856/pm-skills-instruct) routing set, weighted double), its description, and its Simplified Chinese description where one exists, then pruned to the 80 strongest features. The training script is in the model files (`pm_router.py train`).

## Evaluation 评测

Measured when this version was built:

| Test | Top-1 | Top-3 | Cases |
|---|---|---|---|
| Held-out trigger phrases (one per skill with two or more, not used in training) | 72% | 92% | 1275 |
| Chinese eval inputs from the library's eval set | 96% | 100% | 69 |

The Chinese test is small, so treat it as a sanity check rather than a benchmark. Lexical routing misses paraphrases that share no words with a skill; when the top score is low (below about 0.1), ask the user to rephrase or show the top few choices.

## Updates 更新

Rebuilt and republished on every library release by `.github/workflows/publish-modelscope-model.yml` in the source repository. Version: v82.0.0.