summaryrefslogtreecommitdiff
path: root/llm/cmd/eval
diff options
context:
space:
mode:
authorJonas Knobloch <jonas.knobloch@t-online.de>2026-06-27 20:24:38 +0200
committerJonas Knobloch <jonas.knobloch@t-online.de>2026-06-27 20:41:40 +0200
commitc87625c71737f41f42f0188ff87b8d9313c8548a (patch)
tree0d9666eaabe04d29a01b5417e2fa59c96811a17e /llm/cmd/eval
parente1e4eab2e1a682d633125f3c71cc09c3e95b30d3 (diff)
Pad token buffer
Diffstat (limited to 'llm/cmd/eval')
-rw-r--r--llm/cmd/eval/perplexity.go8
1 files changed, 6 insertions, 2 deletions
diff --git a/llm/cmd/eval/perplexity.go b/llm/cmd/eval/perplexity.go
index dd4dae3..b84b659 100644
--- a/llm/cmd/eval/perplexity.go
+++ b/llm/cmd/eval/perplexity.go
@@ -43,8 +43,11 @@ func perplexity() {
n := 0
cfg := llm.TokenBufferConfig{
- Window: 1024,
- Stride: 512,
+ Window: 1024,
+ Stride: 512,
+ PadLeft: false, // probably not fine since we don't adjust model inputs for padding
+ PadRight: true, // fine since we remove padded log probs anyway
+ PadTokenID: 50256,
}
if err := e.RunAndCollect("Perplexity", d, cfg, func(r pplResult) error {
@@ -60,6 +63,7 @@ func perplexity() {
ppl := math.Exp(avg)
fmt.Println(ppl)
+ fmt.Printf("%d tokens\n", n)
if err := m.Destroy(); err != nil {
log.Fatal(err)