pocketcoder-dpo

DPO fine-tune of Ananda100/100m-sftd-python on jondurbin/py-dpo-v0.1, a Python-only preference dataset (chosen = tested/validated code, rejected = lower-quality generations).

• Beta: 0.1
• Learning rate: 5e-06
• Held-out DPO accuracy (prefers chosen over rejected): 84.8%
• Block size: 512
• Vocab size: 32022

Loading

This is a custom architecture, not a native transformers model:

import json, torch
from safetensors.torch import load_file

with open("config.json") as f:
    cfg = json.load(f)

config = GPTConfig(**cfg)
model = GPT(config)
state_dict = load_file("model.safetensors")
model.load_state_dict(state_dict)
model.eval()