README.md

---
license: apache-2.0
language:
- en
base_model: mistralai/Mistral-7B-Instruct-v0.3
datasets:
- hvss/dispatch-7b-data
pipeline_tag: text-generation
tags:
- agent
- router
- orchestration
- tool-use
- lora
- unsloth
- mistral
---

# Dispatch-7B — LoRA adapter 🚦

QLoRA adapter (r=16, attention + MLP projections) for [**Dispatch-7B**](https://huggingface.co/hvss/Dispatch-7B), an agent orchestrator built on Mistral 7B Instruct v0.3 that routes agentic work: request + tool catalog in → JSON execution plan out. **97.4% valid-plan rate** on held-out tasks — full evaluation, usage, and prompt format on the [main model card](https://huggingface.co/hvss/Dispatch-7B).

Use this repo to:
- **Continue fine-tuning** on your own domain's tools (load with Unsloth or PEFT and resume training on your data — the training pipeline is documented on the main card)
- Serve base + adapter separately (e.g., vLLM LoRA serving, or hot-swapping adapters over one shared base model)

For plain inference, prefer the [merged weights](https://huggingface.co/hvss/Dispatch-7B) or the [GGUF quants](https://huggingface.co/hvss/Dispatch-7B-GGUF).

```python
from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    "hvss/Dispatch-7B-LoRA", max_seq_length=3072, load_in_4bit=True,
)
```

License: Apache 2.0.