You can see the process with instructions for creating the model in the notebook: Aligning_DPO_phi3.ipynb
To create it, we started with the Phi-3-Mini-4K-Instruct model and applied DPO alignment using the distilabel-capybara-dpo-7k-binarized dataset.
Phi-3 is a state-of-the-art model with 3.8 billion parameters that has outperformed other models with 7 billion parameters. The DPO alignment process has produced good results, modifying the model's responses and making them more similar to those in the capybara dataset.
Row in the Dataset:
{'prompt': '<|user|>\nAssist me in calculating 9319357631 plus 595. Numbers and symbols only, please.<|end|>\n<|assistant|>\n', 'chosen': 'The sum of 9319357631 and 595 is 9319358226.<|end|>\n', 'rejected': 'The result of adding 9319357631 and 595 is 9319363626.<|end|>\n'}
Prompt: 3713841893836/4? Limit your response to mathematical expressions and symbols.
Response from the Base model:
To find the result of the division, we can simply divide the given number by 4: $$ \frac{3713841893836}{4} = 928460473459
Response from the fine-tuned model:
3713841893836 ÷ 4 = 928460473459
If you want to see how the model was created, you can check out the repository where the book's notebooks are kept up-to-date.