diff options
| author | Jonas Knobloch <jonas.knobloch@t-online.de> | 2026-06-27 20:45:08 +0200 |
|---|---|---|
| committer | Jonas Knobloch <jonas.knobloch@t-online.de> | 2026-06-27 20:45:08 +0200 |
| commit | 48cac2575c9ba5377ddeba2f17ddb68277e410d9 (patch) | |
| tree | fbf07971a15bb6677182b5aa8cd95d92338a0816 | |
| parent | c87625c71737f41f42f0188ff87b8d9313c8548a (diff) | |
Clamp perplexity evaluation
| -rw-r--r-- | llm/cmd/eval/main.go | 7 |
1 files 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 { |
