RuHae/KletterMix-Ablations-Qwen3-Tokenized-Data

Qwen3-tokenized KletterMix ablation data

Megatron indexed training data tokenized with the Qwen3-0.6B tokenizer (vocabulary size 151,936). Each branch is a separate 12B-token corpus.

Branch	Dataset	Composition
fineweb2-de	FineWeb2-deu_Latn	12B-token German FineWeb2 subset, tokenized with the Qwen3 tokenizer.
germanweb	GermanWeb	12B-token GermanWeb subset, tokenized with the Qwen3 tokenizer.
klettermix	KletterMix	12B-token KletterMix subset restricted to examples with proxy score between 0.7 and 0.9.
klettermix-filtered-0p50	KletterMix, proxy score >= 0.50	12B-token KletterMix subset restricted to examples with proxy score at least 0.50.
klettermix-filtered-0p55	KletterMix, proxy score >= 0.55	12B-token KletterMix subset restricted to examples with proxy score at least 0.55.
klettermix-filtered-0p60	KletterMix, proxy score >= 0.60	12B-token KletterMix subset restricted to examples with proxy score at least 0.60.

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.