README.md

---
license: apache-2.0
pipeline_tag: text-classification
library_name: transformers
tags:
  - intent-detection
  - tool-routing
  - compact-model
---

# HarborIntent-Compact

HarborIntent-Compact is a small assistant-routing model designed for intent detection, slot extraction, and safe tool selection in customer-support workflows.

<div align="center">
  <img src="assets/architecture.png" width="70%" alt="HarborIntent compact architecture" />
</div>

## Release Evaluation

The release candidate was evaluated with the frozen HarborBench suite. All values are normalized to the range 0–1.

| Capability | HarborIntent-Compact |
|---|---:|
| Intent Accuracy | 0.851 |
| Slot F1 | 0.829 |
| Entity Recall | 0.824 |
| Multi-turn Success | 0.801 |
| Tool-routing Accuracy | 0.846 |
| Groundedness | 0.837 |
| Safety Pass Rate | 0.910 |
| Response Conciseness | 0.792 |

<div align="center">
  <img src="assets/evaluation.png" width="70%" alt="HarborBench evaluation overview" />
</div>

## Intended Use

Use this model to route English-language support requests to a constrained set of approved tools. Review outputs before using them in high-impact decisions.

## Loading

```python
from transformers import AutoModel, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("HarborIntent-Compact-Release")
model = AutoModel.from_pretrained("HarborIntent-Compact-Release")
```

## Limitations

The model is not designed for open-ended generation, medical advice, legal advice, or autonomous financial actions.

## License

Apache-2.0.