Is this a complete dataset?

#1
by ifmain - opened

Did you use the full set of fineweb 150b or what percentage?

Hey @ifmain .

This is karpathy's build from https://github.com/karpathy/llm.c/discussions/580 converted to HF format to investigate bfloat16 performance - see https://github.com/karpathy/llm.c/pull/571. The run was 150B tokens, 1.5 epochs over the 100B FineWeb sample dataset.

There's active work underway at https://github.com/karpathy/llm.c so I'd suggest following the developments there as well!

ifmain changed discussion status to closed

Sign up or log in to comment