README.md

---
license: llama3.2
base_model: meta-llama/Llama-3.2-1B-Instruct
tags:
  - function-calling
  - tool-use
  - git
  - cli
  - code-assistant
  - distillation
  - llama
datasets:
  - distil-labs/gitara-synthetic
language:
  - en
library_name: transformers
pipeline_tag: text-generation
---

# Llama-3.2-Gitara-1B

*Gitara = **git** + **ara** (the parrot genus): your local stochastic parrot for git commands.*

A 1B parameter function-calling model fine-tuned by [Distil Labs](https://distillabs.ai) to translate plain English into `git` commands. The smallest Gitara variant, optimized for resource-constrained environments while maintaining strong accuracy.

**[GitHub Demo and Code](https://github.com/distil-labs/distil-gitara)**

## Model Details

| | |
|---|---|
| **Developed by** | [Distil Labs GmbH](https://distillabs.ai) |
| **Model type** | Causal language model, fine-tuned for function calling |
| **Language** | English |
| **License** | [Llama 3.2 Community License](https://github.com/meta-llama/llama-models/blob/main/models/llama3_2/LICENSE) |
| **Fine-tuned from** | [meta-llama/Llama-3.2-1B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct) |

## Use Case

Given a natural language description of a git operation, the model outputs a structured JSON tool call that can be converted to an executable git command.

**Supported commands:** `status` · `add` · `commit` · `push` · `pull` · `branch` · `switch` · `restore` · `merge` · `stash` · `rebase` · `reset` · `log`

### Example

**Input:**
```
push feature-x to origin, override any changes there and track it
```

**Output:**
```json
{"name": "git_push", "parameters": {"remote": "origin", "branch": "feature-x", "force": true, "set_upstream": true}}
```

**Resulting command:**
```bash
git push origin feature-x --force --set-upstream
```

### More Examples

| Query | Tool Call Output |
|-------|------------------|
| `what's in the latest stash, show diff` | `{"name": "git_stash", "parameters": {"action": "show", "patch": true}}` |
| `undo last commit but keep the changes` | `{"name": "git_reset", "parameters": {"mode": "soft", "ref": "HEAD~1"}}` |
| `show 8 commits for current branch with graph` | `{"name": "git_log", "parameters": {"limit": 8, "graph": true}}` |
| `merge vendor branch preferring ours` | `{"name": "git_merge", "parameters": {"branch": "vendor", "strategy": "ours"}}` |

## Training

The model was trained using knowledge distillation from the teacher model GPT-OSS-120B.

| | |
|---|---|
| **Training method** | LoRA fine-tuning |
| **Seed examples** | ~100 (manually validated) |
| **Synthetic examples** | 10,000 |
| **Teacher model** | GPT-OSS-120B |

### Training Process

1. Created ~100 seed examples covering all 13 git commands with realistic query phrasings
2. Expanded seed data to 10,000 synthetic training examples using the Distil Labs platform
3. Fine-tuned Llama 3.2 1B Instruct using LoRA
4. Validated on held-out test set

Training data and configuration available in the [GitHub repository](https://github.com/distil-labs/distil-gitara/tree/main/finetuning).

## Evaluation

Evaluated on 50 held-out test examples. Accuracy is measured by parsing outputs into normalized Python dicts and comparing for structural equality.

| Model | Parameters | Accuracy |
|-------|------------|----------|
| GPT-OSS-120B (teacher) | 120B | 0.92 ± 0.02 |
| Llama 3.2 3B Instruct (tuned) | 3B | 0.92 ± 0.01 |
| **Llama 3.2 1B Instruct (tuned)** | **1B** | **0.90 ± 0.01** |
| Llama 3.2 3B Instruct (base) | 3B | 0.12 ± 0.05 |
| Llama 3.2 1B Instruct (base) | 1B | 0.00 ± 0.01 |

The tuned 1B model achieves **0.90 accuracy** while being **120x smaller** than the teacher. The base 1B model completely fails (0.00 accuracy), confirming that fine-tuning is essential.

### When to Use 1B vs 3B

| Choose 1B | Choose 3B |
|-----------|-----------|
| Memory-constrained devices | Maximum accuracy needed |
| Faster inference required | Complex or ambiguous queries |
| 0.90 accuracy is acceptable | Edge cases matter |

## How to Use

### With Ollama (Recommended)

```bash
# Download model
huggingface-cli download distil-labs/Distil-gitara-v2-Llama-3.2-1B-Instruct --local-dir distil-model

# Build with Ollama
cd distil-model
ollama create gitara-1b -f Modelfile

# Run
ollama run gitara-1b "show staged changes with diffs"
```

### With Transformers

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "distil-labs/Distil-gitara-v2-Llama-3.2-1B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

# See GitHub repo for full tool-calling implementation
```

For complete usage instructions, see the [GitHub repository](https://github.com/distil-labs/distil-gitara).

## Inference Speed

Faster than the 3B variant due to smaller size. Suitable for on-device deployment on mobile or embedded systems.

## Limitations

- Accuracy is 0.90, meaning approximately 1 in 10 queries may produce incorrect output
- Limited to the 13 supported git commands and their common options
- Does not support `git checkout` (use `switch` and `restore` instead)
- Single-turn only; does not support multi-step workflows
- May struggle more with ambiguous or complex queries compared to 3B variant

## Model Sources

| | |
|---|---|
| **Homepage** | [https://distillabs.ai](https://distillabs.ai) |
| **Repository** | [https://github.com/distil-labs/distil-gitara](https://github.com/distil-labs/distil-gitara) |
| **Blog post** | [https://distillabs.ai/blog/gitara](https://distillabs.ai/blog/gitara) |
| **Contact** | [[email protected]](mailto:[email protected]) |

## Related Models

- [Llama-3.2-Gitara-3B](https://huggingface.co/distil-labs/Llama-3_2-gitara-3B) — Larger variant (0.92 accuracy)

## Citation

```bibtex
@misc{gitara2025,
  author = {Distil Labs},
  title = {Gitara: A Function-Calling Git Agent},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/distil-labs/Distil-gitara-v2-Llama-3.2-1B-Instruct}
}
```