OLMo 3-tokenized KletterMix ablation data
Megatron indexed training data tokenized with the OLMo 3 tokenizer (vocabulary size 100,278). Each branch contains exactly 12B tokens in 120M-token shards.
| Branch | Dataset | Composition |
|---|---|---|
climbmix-only |
ClimbMix only | Exactly 12B tokens: 100% ClimbMix. |
klettermix-05 |
5% KletterMix | Exactly 12B tokens: 95% ClimbMix (80% base plus 15% extra) and 5% KletterMix. |
klettermix-10 |
10% KletterMix | Exactly 12B tokens: 90% ClimbMix (80% base plus 10% extra) and 10% KletterMix. |
klettermix-15 |
15% KletterMix | Exactly 12B tokens: 85% ClimbMix (80% base plus 5% extra) and 15% KletterMix. |
klettermix-20 |
20% KletterMix | Exactly 12B tokens: 80% ClimbMix and 20% KletterMix. |
fineweb2-de-15 |
15% FineWeb2-deu_Latn | Exactly 12B tokens: 85% ClimbMix and 15% FineWeb2-deu_Latn. |
fineweb2-de-20 |
20% FineWeb2-deu_Latn | Exactly 12B tokens: 80% ClimbMix and 20% FineWeb2-deu_Latn. |
germanweb-15 |
15% GermanWeb | Exactly 12B tokens: 85% ClimbMix and 15% GermanWeb. |
germanweb-20 |
20% GermanWeb | Exactly 12B tokens: 80% ClimbMix and 20% GermanWeb. |
The branch names are part of the reproducibility interface. The private
KletterMix_Ablations
repository records the local source paths, exact branch commits, and the
submodule mapping. The tokenized data can be consumed by Megatron using the
.bin/.idx prefix in the selected branch.
- Downloads last month
- 30