---
language:
- en
license: mit
base_model: Qwen/Qwen2.5-Coder-7B-Instruct
tags:
- qwen2.5
- qwen2.5-coder
- code
- cli
- debugging
- developer-tools
- lora
- mlx
- gguf
- ollama
model-index:
- name: YOLO-Coder-8B
  results: []
---
<div align="center">
<img src="yolo_coder_cropped.png" alt="YOLO-Coder" width="240" />

[Website](https://yolocoderai.com) &nbsp;|&nbsp; [GitHub](https://github.com/erdemozkan/YOLO-CODER) &nbsp;|&nbsp; [Twitter](https://twitter.com/erdemwrites) &nbsp;|&nbsp; [Dataset](https://github.com/erdemozkan/YOLO-CODER/tree/main/benchmark) &nbsp;|&nbsp; [YOLO-Coder-1.5B](https://huggingface.co/erdemozkan/YOLO-Coder-1.5B)

**License: [MIT](https://opensource.org/licenses/MIT)** &nbsp;|&nbsp; **Author: [@erdemwrites](https://twitter.com/erdemwrites)**
</div>

# YOLO-Coder-8B
**Fix broken CLI commands. One command output. Runs 100% locally.**
*Fine-tuned Qwen2.5-Coder-7B · MLX LoRA on Apple Silicon · No API key needed*

| | |
|---|---|
| 🎯 **Task** | CLI error → single bare bash fix command |
| 🏆 **Accuracy** | **77.1%** pipeline×3 · **59.2%** raw LLM (beats GPT-4o) |
| 💾 **Size** | ~4.4GB Q4_K_M GGUF · ~6GB RAM |
| ⚡ **Speed** | 1–3s on Apple Silicon |
| 🔒 **Privacy** | 100% local · no API key · no telemetry |

## Quickstart
```bash
ollama run hf.co/erdemozkan/YOLO-Coder-8B "ModuleNotFoundError: No module named 'flask'"
# → pip install flask
```
That's it. No account. No cloud. No cost per call.

## Benchmark — YOLO-Bench
218 verified CLI errors · structural match scoring (flag-order-independent)
```
YOLO-Coder-8B  pipeline×3  ████████████████████  77.1%  ★ best overall
YOLO-Coder-1.5B pipeline×3 ██████████████████    71.1%
Claude Sonnet  raw         ████████████████       60.1%
YOLO-Coder-8B  raw         ███████████████        59.2%  ★ best offline
GPT-4o         raw         ████████████           48.6%
YOLO-Coder-1.5B raw        ██████████             42.2%
```

| Mode | Structural Match |
|---|---|
| Raw LLM (no pipeline) | **59.2%** |
| Pipeline × 1 (interceptors + LLM) | **72.0%** |
| Pipeline × 3 (interceptors + memory + 3 LLM attempts) | **77.1%** |

> YOLO-Coder-8B pipeline×3 is the highest score of any model tested — including GPT-4o and Claude Sonnet — running entirely offline.

Scoring code and dataset: [github.com/erdemozkan/YOLO-CODER/tree/main/benchmark](https://github.com/erdemozkan/YOLO-CODER/tree/main/benchmark)

## How the pipeline works
```
Your error → [91 interceptors <1ms] → [fix memory <5ms] → [LLM 1-3s] → Fix
                ↑ ~50% of fixes stop here
```
Half of all fixes never reach the LLM. The model is the safety net, not the first guess.

## Usage with YOLO-CODER
```bash
pip install yolo-coder

yoco python3 myapp.py        # 8B is the default
yoco npm run dev
yoco --model hf.co/erdemozkan/YOLO-Coder-8B python3 myapp.py
```

## Prompt format (ChatML)
```
<|im_start|>system
You are a CLI repair tool. Output ONLY a single bare bash command to fix the error. No explanation. No markdown. No backticks.<|im_end|>
<|im_start|>user
[Linux] $ python3 myapp.py
Error:
ModuleNotFoundError: No module named 'requests'
FIX:<|im_end|>
<|im_start|>assistant
pip install requests<|im_end|>
```

## Training
> "Trained on a MacBook Air. No rented A100s."

| Property | Value |
|---|---|
| Base model | `Qwen/Qwen2.5-Coder-7B-Instruct` |
| Fine-tune method | LoRA via MLX on Apple Silicon |
| LoRA rank / scale | 8 / 20.0 |
| Layers trained | 28 |
| Training iterations | 500 |
| Learning rate | 1e-5 |
| Training examples | **6,719** error/fix pairs across 15 categories |
| Export | Merged weights → Q4_K_M GGUF for Ollama |

## Files
| File | Description |
|---|---|
| `YOLO-Coder-8B-Q4_K_M.gguf` | Q4_K_M quantized GGUF (~4.4GB) — use this with Ollama |
| `safetensors/` | fp16 safetensors — for further fine-tuning |

## 1.5B vs 8B
| | [YOLO-Coder-1.5B](https://huggingface.co/erdemozkan/YOLO-Coder-1.5B) | YOLO-Coder-8B |
|---|---|---|
| Size | ~941MB | ~4.4GB |
| RAM needed | ~2GB | ~6GB |
| Speed | <1s on Apple Silicon | 1–3s on Apple Silicon |
| Raw accuracy | 42.2% | 59.2% |
| Pipeline×3 accuracy | 71.1% | **77.1%** |
| Best for | Speed, low-RAM machines | Hard errors, best accuracy |

## Limitations
- Single-command output only — not designed for multi-step fixes without a wrapper
- Complex or highly novel errors may produce suboptimal output
- Not a general-purpose coding assistant

## License
MIT