From 48cac2575c9ba5377ddeba2f17ddb68277e410d9 Mon Sep 17 00:00:00 2001 From: Jonas Knobloch Date: Sat, 27 Jun 2026 20:45:08 +0200 Subject: Clamp perplexity evaluation --- llm/cmd/eval/main.go | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-) diff --git a/llm/cmd/eval/main.go b/llm/cmd/eval/main.go index 3a710ac..5e3390f 100644 --- a/llm/cmd/eval/main.go +++ b/llm/cmd/eval/main.go @@ -16,12 +16,15 @@ func main() { perplexity() + // pad right: 11.930400581007826 / 22544 tokens + // unpadded: 11.246722646921588 / 19451 tokens + if err := gpt2.DestroyEnvironment(); err != nil { log.Fatal(err) } } -func data() *dataset.ParquetReader { +func data() dataset.Reader { var miniPile *dataset.ParquetReader if r, err := dataset.NewParquetReader(shelf.Abs("data/minipile/validation")); err != nil { @@ -30,7 +33,7 @@ func data() *dataset.ParquetReader { miniPile = r } - return miniPile + return dataset.NewClampedReader(miniPile, 10) } func model() *gpt2.Model { -- cgit v1.2.3