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)
	}
}