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
|
package entropy
import (
"fmt"
"go.jknobloc.com/x/dataset"
"go.jknobloc.com/x/llm"
)
type logProb struct {
document int
token int
value float32
offset int
}
func Context(model llm.Causal, tokenizer llm.Tokenizer, data dataset.Reader) error {
evaluatorConfig := llm.EvaluatorConfig{
BatchSize: 32,
NumWorkers: 16,
}
tokenBufferConfig := llm.TokenBufferConfig{
Window: 1024,
Stride: 512,
PadLeft: false,
PadRight: false,
PadTokenID: 256,
}
eval := llm.NewEvaluator(model, tokenizer, func(job llm.Job, logProbs []float32, tokens []int) []logProb {
r := make([]logProb, len(tokens))
for i, token := range tokens {
r[i] = logProb{
document: job.Document,
token: token,
value: logProbs[i],
offset: job.Position*tokenBufferConfig.Stride + job.Seen + i,
}
}
return r
}, evaluatorConfig)
return eval.RunAndCollect("Context", data, tokenBufferConfig, func(r []logProb) error {
for _, l := range r {
fmt.Println(l) // TODO implement
}
return nil
})
}
|
