Here’s the translation to American English:
—
In the recent Build 2025 conference, Microsoft unveiled an innovative fine-tuning feature for its small-scale language model, Phi Silica, which incorporates the low-rank adaptation (LoRA) technique. This advanced method allows for the optimization of the model’s parameters using a limited set of custom data while maintaining its overall performance.
The development of this functionality has focused on a specific practical use: creating high-quality quizzes for the Kahoot! platform. Thanks to this advancement, rejection rates have been reduced by 75% and the subjective quality of the generated quizzes has improved by 4.6 times.
Microsoft Learning Zone, the first educational application designed specifically for PCs with Copilot+, has partnered with Kahoot! to facilitate the creation of interactive classroom games using the power of Phi Silica. This tool supports a variety of generation tasks, from dynamic presentations to multiple-choice formats. The integration of LoRA allows the base Phi Silica model to be specialized to meet diverse educational needs without the complication of developing multiple tailored models.
Regarding the quality of the generated quizzes, Microsoft has defined two axes: verifiable quality, which encompasses the formatting requirements set by Kahoot!, and subjective quality, which examines aspects such as clarity and educational value. Specific metrics have been established, and a new evaluation framework has been implemented, including AI agents that simulate a review team to measure the effectiveness of the quizzes.
To achieve effective fine-tuning with LoRA, a high-quality dataset was developed that combines educational materials with question and answer generation, using a leading language model as a reference. This allowed for the creation of a more diverse and enriched initial dataset, crucial for training the model.
Additionally, during the parameter optimization in the LoRA training, it was validated that the default values of the AI toolkit were suitable for improving the quality of the generated responses. These adjustments resulted in a noticeable enhancement in the user experience, with a strong focus on the efficiency and relevance of the answers.
Quality testing results have shown that the customized Phi Silica system with LoRA significantly outperforms the base model across all metrics, increasing satisfaction in both automatic and human evaluations. In total, approximately 13,000 synthetic examples were generated for the model’s training and evaluation.
As a step towards public implementation, Kahoot! game generation through Microsoft Learning Zone will be available for educators to test later this summer. This advancement underscores how smaller models, when suitably adapted, can provide effective and personalized AI experiences in educational contexts.
—
If you need further modifications or adjustments, feel free to ask!
Referrer: MiMub in Spanish
