初始化项目,由ModelHub XC社区提供模型

Model: smashingtags/nova-router-1.5b
Source: Original Platform
This commit is contained in:
ModelHub XC
2026-07-17 07:49:11 +08:00
commit 4630ee8e10
19 changed files with 152285 additions and 0 deletions

56
README.md Normal file
View File

@@ -0,0 +1,56 @@
---
license: apache-2.0
base_model: Qwen/Qwen2.5-Coder-1.5B-Instruct
library_name: transformers
pipeline_tag: text-generation
tags:
- function-calling
- tool-use
- agent
- qwen
- gguf
- eightly-os
- nova
language:
- en
---
# Nova Router 1.5B
**Nova Router 1.5B** is the tool-routing model for **Nova**, the local AI assistant built into [Eight.ly OS](https://eight.ly). Given a user request, it decides **which** NAS management tool to call (and with what arguments) across the full Eight.ly tool catalog — Docker, storage, VMs, LXC, file sharing, and system administration.
It is a fine-tune of **Qwen2.5-Coder-1.5B** and scores **99%** tool-selection accuracy on the Eight.ly evaluation suite.
## Role in Nova
Nova is **router-in-front**: this 1.5B model is the single tool router for every conversation. It runs as a **hidden dependency** — it is installed automatically alongside whichever *talker* (conversational model) the user chooses, and is never selected directly. The router picks the tool; the talker writes the reply.
```
user query → intent classifier → retrieval (nomic-embed-text) → Nova Router 1.5B (picks the tool) → talker (writes the reply)
```
## Files
| File | Quant | Size | Use |
|------|-------|------|-----|
| `gguf/eightly-agent-router-q16-Q8_0.gguf` | Q8_0 | ~1.6 GB | Shipped quant (Ollama) |
| `gguf/eightly-agent-router-q16-Q4_K_M.gguf` | Q4_K_M | ~1.0 GB | Smaller alternative |
| `*.safetensors` | fp16 | — | Merged weights + LoRA adapter |
## Usage
```bash
ollama pull hf.co/smashingtags/nova-router-1.5b:Q8_0
```
Within Eight.ly OS this is handled automatically — installing any Nova talker pulls this router and the retrieval embedder with it.
## Training
- **Base:** Qwen2.5-Coder-1.5B
- **Data:** the Eight.ly tool-calling dataset (41 tools across 6 domains)
- **Eval:** 99% tool-selection accuracy, 0 false positives on no-tool queries
## License
Apache-2.0.