Open Vectorrent opened 10 months ago
There's a basic shell version of the Pythia harness currently working, but we really need to convert to a Python-native version of it. The reason is for models like ModuleFormer; we need to load a custom model architecture, which hooks into the Huggingface Transformers API, before ever calling the harness. I don't think we can do that from the shell (nor are we able to re-use any of our other code).
Probably via the Pythia test harness: https://github.com/EleutherAI/lm-evaluation-harness