---

base_model: Qwen/Qwen2.5-Coder-7B-Instruct

language: en

library_name: transformers

license: apache-2.0

pipeline_tag: text-generation

tags:

• qwen2.5
• qwen2.5-coder
• code
• personality
• yolo-vibes
• sft
• dpo
• fine-tune
• gguf
• llama.cpp

model_type: qwen2

---

claude-yolo-vibes-v4-GGUF

GGUF quantizations of claude-yolo-vibes-v4-dpo — a Qwen2.5-Coder-7B-Instruct fine-tune with "yolo vibes" personality and 88.4% HumanEval pass@1.

Quantizations

| File | Format | Size | Bits/Weight | Use Case |

|---|---|---|---|---|

| claude-yolo-vibes-v4-Q4_K_M.gguf | Q4_K_M | 4.4 GB | 4.91 | Recommended — best quality/size tradeoff |

| claude-yolo-vibes-v4-Q5_K_M.gguf | Q5_K_M | 5.1 GB | 5.5 | Higher quality, slightly larger |

| claude-yolo-vibes-v4-Q8_0.gguf | Q8_0 | 7.5 GB | 8.5 | Near-lossless |

Benchmark

| Model | HumanEval pass@1 | Personality |

|---|---|---|

| Qwen2.5-Coder-7B-Instruct (base) | 88.4% | None |

| claude-yolo-vibes-v4 (SFT+DPO) | 88.4% | Yolo vibes |

Zero personality tax — the fine-tune adds personality without degrading coding ability.

Usage

llama.cpp

llama-cli -m claude-yolo-vibes-v4-Q4_K_M.gguf -p "Write a Python function to reverse a linked list" --chat-template chatml

Ollama

Create a Modelfile:

FROM ./claude-yolo-vibes-v4-Q4_K_M.gguf
TEMPLATE """{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}<|im_start|>user
{{ .Prompt }}<|im_end|>
<|im_start|>assistant
"""
PARAMETER stop "<|im_end|>"
PARAMETER stop "<|im_start|>"

ollama create yolo-vibes -f Modelfile
ollama run yolo-vibes

LM Studio

Download any GGUF file and load it directly. The ChatML template is embedded in the file metadata.

Training Details

• Base model: Qwen/Qwen2.5-Coder-7B-Instruct
• Stage 1 SFT: 100K verified code examples, 2 epochs, lr=2e-5, cosine schedule
• Stage 2 DPO: Light-touch personality alignment, 1 epoch, lr=1e-6, beta=0.3
• Training data: 786K examples from 5 public datasets, all decontaminated against HumanEval
• Hardware: AMD MI300X (192GB VRAM)

Full training details: claude-yolo-vibes-v5 repo

Other Formats

• BF16 safetensors — full precision
• MLX 4-bit — Apple Silicon
• Ollama registry — one-command install