1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
|
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
cfg := llm.TokenBufferConfig{
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 {
total += r.v
n += r.n
return nil
}); err != nil {
log.Fatal(err)
}
avg := total / float64(n)
ppl := math.Exp(avg)
fmt.Println(ppl)
fmt.Printf("%d tokens\n", n)
if err := m.Destroy(); err != nil {
log.Fatal(err)
}
}
|
