From bc88dbac521b0de3db441c37f8e946cc2b0d3669 Mon Sep 17 00:00:00 2001 From: Jonas Knobloch Date: Mon, 9 Mar 2026 18:59:37 +0100 Subject: Add ppl command --- llm/cmd/ppl/ppl.go | 69 ++++++++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 69 insertions(+) create mode 100644 llm/cmd/ppl/ppl.go (limited to 'llm/cmd') diff --git a/llm/cmd/ppl/ppl.go b/llm/cmd/ppl/ppl.go new file mode 100644 index 0000000..622644a --- /dev/null +++ b/llm/cmd/ppl/ppl.go @@ -0,0 +1,69 @@ +package main + +import ( + "fmt" + "log" + + "github.com/jonasknobloch/mbpe" + "github.com/jonasknobloch/x/dataset" + "github.com/jonasknobloch/x/gpt2" + "github.com/jonasknobloch/x/llm" +) + +func main() { + d := data() + m := model() + t := tokenizer() + + e := llm.NewEvaluator() + + e.SetTokenizer(t) + e.AddModel(m) + + ppl, err := e.Perplexity(d) + + if err != nil { + log.Fatal(err) + } + + fmt.Println(ppl) +} + +func data() *dataset.Reader { + var miniPile *dataset.Reader + + if r, err := dataset.NewReader("dataset/cmd/dataset/tmp/minipile/validation"); err != nil { + log.Fatal(err) + } else { + miniPile = r + } + + return miniPile +} + +func model() *gpt2.Model { + m := gpt2.NewModel("gpt2/models/base/model.onnx", "0") + + if err := m.Init(); err != nil { + log.Fatal(err) + } + + return m +} + +func tokenizer() *mbpe.Tokenizer { + m := mbpe.NewMBPE() + + if err := m.Load("gpt2/models/base/vocab.json", "gpt2/models/base/merges.txt"); err != nil { + log.Fatal(err) + } + + t := mbpe.NewTokenizer(m) + + byteLevel := mbpe.NewByteLevel(false) + + t.SetPreTokenizer(byteLevel) + t.SetDecoder(byteLevel) + + return t +} -- cgit v1.3.1