From 6e6a94b12d6d43c1ce4b8e587742627af6d86a43 Mon Sep 17 00:00:00 2001 From: Jonas Knobloch Date: Fri, 11 Sep 2026 18:36:45 +0200 Subject: WIP --- research/knobloch/cmd/train/main.go | 39 ++++++++++++++++---------------- research/knobloch/cmd/train/serialize.go | 4 +++- 2 files changed, 22 insertions(+), 21 deletions(-) (limited to 'research/knobloch/cmd') diff --git a/research/knobloch/cmd/train/main.go b/research/knobloch/cmd/train/main.go index 72f7e00..ebe8c54 100644 --- a/research/knobloch/cmd/train/main.go +++ b/research/knobloch/cmd/train/main.go @@ -10,25 +10,24 @@ import ( "github.com/jonasknobloch/mbpe" "go.jknobloc.com/x/dataset" - "go.jknobloc.com/x/profile" "go.jknobloc.com/x/shelf" "go.jknobloc.com/x/tokenizer/bpe" "go.jknobloc.com/x/tokenizer/bpe/split" ) func main() { - stop := profile.CPU() + // stop := profile.CPU() - // train() + train() // serialize() - profile.Mem() + // profile.Mem() - stop() + // stop() } func train() { - out := shelf.Abs("results/knobloch/minipile") + out := shelf.Abs("results/knobloch/fineweb_verify") if err := os.MkdirAll(out, os.ModePerm); err != nil { log.Fatal(err) @@ -44,7 +43,7 @@ func train() { return m }() - // mbpe.InvertWeightFunction = true + mbpe.InvertWeightFunction = true // mbpe.UseSimpleClashes = true newTrainer := func(segmenter mbpe.Segmenter, alpha float64) *mbpe.MBPETrainer { @@ -58,24 +57,24 @@ func train() { b.SetMatcher(split.NewFSA()) - return mbpe.NewMBPETrainer(b, segmenter, alpha, mbpe.NewMBPE(), 1<<17, alphabet) + return mbpe.NewMBPETrainer(b, segmenter, alpha, mbpe.NewMBPE(), 1<<19, alphabet) } configs := []struct { float64 string }{ - {0.0, "m000_minipile_v2"}, - {0.1, "m010_minipile_v2"}, - {0.2, "m020_minipile_v2"}, - {0.3, "m030_minipile_v2"}, - {0.4, "m040_minipile_v2"}, - {0.5, "m050_minipile_v2"}, - {0.6, "m060_minipile_v2"}, - {0.7, "m070_minipile_v2"}, - {0.8, "m080_minipile_v2"}, - {0.9, "m090_minipile_v2"}, - {1.0, "m100_minipile_v2"}, + {0.0, "mi000_fineweb"}, + {0.1, "mi010_fineweb"}, + {0.2, "mi020_fineweb"}, + {0.3, "mi030_fineweb"}, + {0.4, "mi040_fineweb"}, + {0.5, "mi050_fineweb"}, + {0.6, "mi060_fineweb"}, + {0.7, "mi070_fineweb"}, + {0.8, "mi080_fineweb"}, + {0.9, "mi090_fineweb"}, + {1.0, "mi100_fineweb"}, } for i, c := range configs { @@ -86,7 +85,7 @@ func train() { if dictErr := t.LoadDict(dict); dictErr != nil { var reader dataset.Reader - if r, err := dataset.NewParquetReader(shelf.Abs("data/minipile/train")); err != nil { + if r, err := dataset.NewParquetReader(shelf.Abs("data/fineweb-edu/sample-10BT/train")); err != nil { log.Fatal(err) } else { reader = r diff --git a/research/knobloch/cmd/train/serialize.go b/research/knobloch/cmd/train/serialize.go index 9fbf968..5536450 100644 --- a/research/knobloch/cmd/train/serialize.go +++ b/research/knobloch/cmd/train/serialize.go @@ -23,7 +23,9 @@ func serialize() { paths = ps } - steps := []int{100512, 50256, 32768, 16384, 8192} + // steps := []int{100512, 50256, 32768, 16384, 8192} + // steps := []int{131072} + steps := []int{256} outRoot := shelf.Abs("tokenizers") -- cgit v1.3.1