Scale AI New Grad ML Research Engineer LLM Onsite

The earlier phone screen used NumPy for data processing and image classification.

During the onsite, the LLM practical was Training Data And Model Error Debugging. It used a question-answering model. The first step was to read JSON files from a directory and turn them into CSV training data. The next step was to find three bugs that caused runtime errors. The code used TensorFlow, and the normal debugging plugin wasn’t available in VS Code.

The Python coding round asked them to build a graph from a dictionary and solve Shortest Distance Between Two Nodes. They ran out of time after hitting errors while building the graph.

The ML discussion covered overfitting, CNNs, vision transformers, different vision tasks, projection methods, and a masked-transformer paper. They were rejected after the onsite.