This is the intermediate SFT checkpoint, before DPO alignment. It has the personality but not the preference optimization. For the final production model, use davidnichols-ops/claude-yolo-vibes-v4-mlx-4bit (quantized) or davidnichols-ops/claude-yolo-vibes-v4-dpo (full BF16).

Training Details

Parameter	Value
Base model	Qwen2.5-Coder-7B-Instruct
Training data	1,031 verified agent sessions
Epochs	3
Learning rate	2e-5
Batch size	4 (with gradient accumulation)
Hardware	AMD MI300X (ROCm)
Training time	~45 min
Final loss	0.56
Token accuracy	91.2%

HumanEval

Stage	Pass@1
Base	88.4%
SFT (this model)	88.4%
DPO (final)	88.4%

Personality tax: 0.0%. SFT alone preserved full coding capability.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("davidnichols-ops/claude-yolo-vibes-v4-sft", torch_dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("davidnichols-ops/claude-yolo-vibes-v4-sft")

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)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

License

Apache 2.0.