claude-yolo-vibes-v4 (DPO — final)
The final DPO-aligned model of claude-yolo-vibes-v4 — a Qwen2.5-Coder-7B fine-tune with a personality layer and zero capability tax.
What is this?
This is the full BF16 DPO checkpoint — the production model. It has been through SFT (personality injection) and DPO (preference alignment). HumanEval Pass@1 is identical to the base model: 88.4%.
For local deployment on Apple Silicon, use the MLX 4-bit version: davidnichols-ops/claude-yolo-vibes-v4-mlx-4bit.
Training Pipeline
| Stage | Data | Hardware | Time | Loss |
|---|---|---|---|---|
| SFT | 1,031 verified agent sessions | MI300X (ROCm) | ~45 min | 0.56 |
| DPO | 1,031 preference pairs (beta=0.3) | MI300X (ROCm) | 3 min | 0.52 |
Benchmarks
| Metric | Base | SFT | DPO (this) | Tax |
|---|---|---|---|---|
| HumanEval Pass@1 | 88.4% | 88.4% | 88.4% | 0.0% |
| Reward accuracy | — | — | 100% | — |
| Model size (BF16) | 15 GB | 15 GB | 15 GB | — |
| Model size (4-bit MLX) | 4.0 GB | — | 4.0 GB | — |
System Prompt
Two-mode design:
- Vibes mode: Witty, dark humor, raunchy. Still ships working code.
- Code mode: Silence, then code. No filler.
The model snaps between modes based on the user's request.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("davidnichols-ops/claude-yolo-vibes-v4-dpo", torch_dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("davidnichols-ops/claude-yolo-vibes-v4-dpo")
messages = [
{"role": "system", "content": "You are a helpful coding assistant."},
{"role": "user", "content": "Write a Python function to reverse a linked list"},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Also Available
- MLX 4-bit:
davidnichols-ops/claude-yolo-vibes-v4-mlx-4bit(4.0 GB, Apple Silicon) - Ollama:
ollama pull davidnicholsops/claude-yolo-vibes-v4(4.7 GB, Q4_K_M GGUF) - SFT checkpoint:
davidnichols-ops/claude-yolo-vibes-v4-sft(intermediate)
License
Apache 2.0.