Maths for the anneal
The last part of training, when the learning rate comes down (the anneal, or decay phase), is where a model gains the most from the best data. This folder holds the maths and reasoning text for that phase: 9.8B tokens.
Everything here is kept as plain documents (a problem followed by its worked solution, or a page of maths), not as chat turns. Chat data stays in tuning, after pretraining.
Sources
| source | tokens | what it is | licence |
|---|---|---|---|
MegaMath Q&A (LLM360/MegaMath), megamath_qa/ |
6.53B | maths questions with step-by-step answers, about 300 tokens each | ODC-By |
MathCoder2 (Dolmino mix), mathcoder2/ |
2.97B | maths with reasoning and code | Apache 2.0 |
TuluMath (Dolmino mix), tulu_math/ |
0.25B | persona word problems with solutions | ODC-By |
DolminoSynthMath (Dolmino mix), dolmino_math_synth/ |
0.03B | basic arithmetic and GSM-style problems | ODC-By |
CodeSearchNet maths filter (Dolmino mix), codesearchnet_owmfilter/ |
0.002B | maths-related code | ODC-By |
GSM8K training set (Dolmino mix), gsm8k/ |
0.001B | grade-school problems with worked answers | MIT |
Left out: Metamath-owmfilter (CC BY-SA); MegaMath Web-Pro (maths web pages that Nemotron-CC-Math already covers, with no shared URLs to remove the repeats); TinyGSM and TinyGSM-MIND (problems solved as Python programs: a single format that costs model capacity and does not carry over to plain-answer benchmarks); and long reasoning traces from large models (thousands of tokens per answer, too long for a model of this size).
Format
The same as the text shards: Gemma 4 tokens, 3 bytes each, every document <bos> text <eos>, with a .json and a
.counts.npz next to each shard. One folder per source.