Multilingual structural evaluation of code language models

This project extends A Critical Study of What Code-LLMs (Do Not) Learn to Java, Go, and JavaScript using 3,000-program cohorts. It retains the original Python results as references, alongside additional t-SNE results computed on the original Python dataset.

• Analysis runbook: commands to generate the multilingual results.
• Result locations: paths to the retained results and comparison figures.

Upstream sources and licenses

• Original project
• DirectProbe
• Repository license and DirectProbe license