Cleanup eval command

* Remove logprob extraction

1 files changed, 62 insertions, 0 deletions

diff --git a/llm/cmd/eval/perplexity.go b/llm/cmd/eval/perplexity.go
new file mode 100644
index 0000000..ba92c99
--- /dev/null
+++ b/llm/cmd/eval/perplexity.go
@@ -0,0 +1,62 @@
+package main
+
+import (
+	"fmt"
+	"log"
+	"math"
+
+	"go.jknobloc.com/x/llm"
+)
+
+type pplResult struct {
+	v float64
+	n int
+}
+
+func perplexity() {
+	d := data()
+
+	m := model()
+	t := tokenizer()
+
+	e := llm.NewEvaluator(m, t, func(job llm.Job, logProbs []float32, tokens []int) pplResult {
+		total := float64(0)
+
+		n := 0
+
+		for _, p := range logProbs {
+			total -= float64(p)
+
+			n++
+		}
+
+		return pplResult{
+			v: total,
+			n: n,
+		}
+	}, llm.EvaluatorConfig{
+		BatchSize:  1,
+		NumWorkers: 4,
+	})
+
+	total := float64(0)
+	n := 0
+
+	if err := e.RunAndCollect("Perplexity", d, 1024, 512, func(r pplResult) error {
+		total += r.v
+		n += r.n
+
+		return nil
+	}); err != nil {
+		log.Fatal(err)
+	}
+
+	avg := total / float64(n)
+	ppl := math.Exp(avg)
+
+	fmt.Println(ppl)
+
+	if err := m.Destroy(); err != nil {
+		log.Fatal(err)
+	}
+}