🧑‍🔬 Hany Hassan Awadalla

Hany Hassan Awadalla's papers here focus on small language models. They ask how far a compact model can go when the effort goes into the quality of its training data rather than into more parameters, and how such a model can be extended to handle images and speech. The work falls under Language models & pretraining and Vision & language. It takes the form of technical reports that cover data curation, staged training, alignment and evaluation.

The most cited of these is the Phi-3 Technical Report (2024). It presents phi-3-mini, a 3.8B-parameter model pretrained in two stages. The first stage used broad filtered web text. The second used a stricter slice of that text, combined with synthetic material produced by larger models. After pretraining, the model was tuned for conversation, robustness and safety through supervised training and DPO. The authors describe this as a "data optimal" approach, in which web content is kept according to what it teaches and how much reasoning it supports.

The more recent Phi-4-Mini Technical Report (2025) stays at the same 3.8B scale, with training data weighted toward math and code. It also introduces Phi-4-Multimodal, which keeps the language backbone frozen. Each modality gets its own encoder, projection layer and LoRA module. This "Mixture-of-LoRAs" design lets one checkpoint accept text, images, audio or combinations of them without degrading the model's text abilities.

Hany Hassan Awadalla is currently at Meta in London. Bylines from 2015 to 2025 list Microsoft in Redmond. Co-authors on these reports include Alon Benhaim, Jianmin Bao and Nguyen Bach, along with Marah Abdin and others.

Hany Hassan Awadalla has 2 papers in our library, in 2024–2025, most of them on Language models & pretraining, Vision & language; often with Alon Benhaim, Jianmin Bao, Nguyen Bach.