Which evaluation infra was used for benchmarking?

[woffett]

A previous GH issue (here) mentions that a modified version of this script (here) was used to collect MMLU numbers. What about scripts for other benchmarks in the blogpost? As HuggingFace notes here, numbers can vary wildly if different evaluation codebases are used, would be useful to know if HELM vs Eleuther-AI's harness vs any internal benchmarking library was used for Hellaswag, Winogrande, etc. Same applies to the long sequence tasks (e.g. AMI, FD, SCROLLS) as well. Thanks!

[tianxie-9]

We used lm-evaluation-harness for the lm harness scores. For the long sequence tasks, we used our internal code.