---
license: apache-2.0
base_model: Qwen/Qwen3-8B
tags:
  - dbt
  - tool-calling
  - gguf
  - qwen3
  - fine-tuned
  - unsloth
model_type: qwen3
quantized_by: llama.cpp
pipeline_tag: text-generation
---

# ecu-pilot-qwen3-8b-GGUF

A fine-tuned Qwen3-8B model for **dbt project assistance via tool calling** against [dbt-index](https://github.com/dbt-labs/dbt-index).

## Overview

ecu-pilot is trained to autonomously select and call the correct dbt-index tool (out of 9 available tools) based on natural language questions about a dbt project, then summarize the results concisely.

### Tools

| Tool | Description |
|------|-------------|
| `status` | Project overview — model/test/source counts |
| `schema` | Database schema exploration |
| `search` | Full-text search across models, columns, descriptions |
| `node` | Detailed info on a specific model/source/test |
| `lineage` | Upstream/downstream dependency graph |
| `query` | Run SQL against the dbt-index DuckDB |
| `report` | Coverage reports (tests, docs, etc.) |
| `impact` | Blast radius analysis for model changes |
| `diff` | Compare project state across branches |

## Training

- **Base model**: Qwen3-8B (via `unsloth/Qwen3-8B-bnb-4bit`)
- **Method**: QLoRA (r=32, alpha=32) with Unsloth + TRL SFTTrainer
- **Data**: 667 tool-calling conversation examples from 8 dbt projects (BIRD benchmark databases)
- **Epochs**: 1
- **Hardware**: AWS EC2 g6 (NVIDIA L40S 48GB)

## Quantization

- **Format**: GGUF Q4_K_M (4-bit quantized via llama.cpp)
- **Size**: ~4.7 GB

## Chat Template

This model uses a **custom chat template** for tool calling. The template is included as [`chat_template.jinja`](chat_template.jinja) in this repo.

Key differences from stock Qwen3:
- Tool definitions are injected into the system message within `<tools>` XML tags
- Tool calls use `<tool_call>` / `</tool_call>` XML delimiters
- Tool responses are wrapped in `<tool_response>` / `</tool_response>` and sent as user messages

When using with inference servers that support Jinja templates (vLLM, TGI, llama.cpp server), point to this template file.

## Usage

### LM Studio

1. Download the `.gguf` file
2. Place in `~/.lmstudio/models/json-l/ecu-pilot-qwen3-8b-GGUF/`
3. Load in LM Studio, set the chat template to the contents of `chat_template.jinja`, and start the local server

### Ollama

Create a `Modelfile`:

```
FROM ./ecu-pilot-qwen3-8b-q4km.gguf

TEMPLATE """{{- if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{- if .Tools }}<|im_start|>system
# Tools
You are provided with function signatures within <tools></tools> XML tags:
<tools>
{{- range .Tools }}
{{ json . }}
{{- end }}
</tools>
<|im_end|>
{{ end }}{{- range .Messages }}<|im_start|>{{ .Role }}
{{- if .Content }}
{{ .Content }}
{{- end }}{{- if .ToolCalls }}
<tool_call>
{{- range .ToolCalls }}
{"name": "{{ .Function.Name }}", "arguments": {{ json .Function.Arguments }}}
{{- end }}
</tool_call>
{{- end }}<|im_end|>
{{ end }}<|im_start|>assistant
"""

SYSTEM """You are an expert dbt project assistant with access to a dbt-index metadata server."""

PARAMETER stop "<|im_end|>"
PARAMETER stop "<tool_call>"
PARAMETER temperature 0.7
PARAMETER num_ctx 4096
```

Then:

```bash
ollama create ecu-pilot -f Modelfile
ollama run ecu-pilot
```

## Evaluation

On 8 test prompts (2 runs each):

| Metric | Score |
|--------|-------|
| Tool called | 88% |
| Correct tool | 75% |
| Answer generated | 100% |

## License

Apache 2.0