From b81fa93e098ca689cb6f73627968fe0b4d49decb Mon Sep 17 00:00:00 2001
From: Jonas Knobloch <jonas.knobloch@t-online.de>
Date: Tue, 7 Apr 2026 00:01:23 +0200
Subject: Run evaluator with score

---
 llm/cmd/eval/ppl.go | 14 +++++++++++---
 1 file changed, 11 insertions(+), 3 deletions(-)

(limited to 'llm/cmd/eval/ppl.go')

diff --git a/llm/cmd/eval/ppl.go b/llm/cmd/eval/ppl.go
index 261114e..8838f86 100644
--- a/llm/cmd/eval/ppl.go
+++ b/llm/cmd/eval/ppl.go
@@ -21,11 +21,19 @@ func perplexity() {
 	m := model()
 	t := tokenizer()
 
-	e := llm.NewEvaluator(m, t, func(job llm.Job, logits [][]float32, tokens []int) pplResult {
-		p, n := llm.NegLogLikelihood(logits, tokens)
+	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: p,
+			v: total,
 			n: n,
 		}
 	})
-- 
cgit v1.3.1