File size: 3,488 Bytes
f310c37 8da77e3 f310c37 8da77e3 f310c37 8da77e3 f310c37 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 | ---
license: other
license_name: mixed-upstream-licenses
pretty_name: Echo Chamber math-only pretraining data (Llama-2 tokenized)
language:
- en
size_categories:
- 10B<n<100B
---
# Echo Chamber math-only pretraining data, tokenized
The pretraining sources of *Echo Chamber: RL Post-training Amplifies Behaviors Learned in Pretraining*
(Zhao, Meterez et al., COLM 2025, [arXiv:2504.07912](https://arxiv.org/abs/2504.07912)), tokenized exactly
as in the authors' code ([rosieyzh/openrlhf-pretrain](https://github.com/rosieyzh/openrlhf-pretrain),
`pretraining/data/*_to_tokens.py`): datatrove 0.4.0 `DocumentTokenizer` with the Llama-2 tokenizer
(vocab 32000; the copy shipped with the authors' checkpoints, identical to `meta-llama/Llama-2-7b-hf`),
one `</s>` (id 2) after every document, no BOS, documents shuffled with seed 0. Question-answer sources are
`question + "\n\n" + solution` with no chat template. Produced by
`PRETRAINING-PASSK/OLMo-mathonly/pretraining/passk/flame/data/tokenize_all.sh`.
## Layout
Each piece is one directory of datatrove shards: `<rank>_<sub>_shuffled.ds` (raw little-endian uint16
token ids, 2 bytes per token), `.ds.index` (document end offsets) and `.ds.metadata`. The OLMo fork reads
the `.ds` files as memmaps (`data.memmap_dtype: uint16`); `MANIFEST.tsv` lists every file with its size and
token count.
| piece | source and text construction | shards | tokens |
|---|---|---|---|
| `MetaMathQA-tokenized/` | meta-math/MetaMathQA-395K: query + \n\n + response | 1 | 0.101B |
| `algebraic-stack-tokenized/` | EleutherAI/proof-pile-2, algebraic-stack/train (documents) | 61 | 10.900B |
| `eval/gsm8k-test-tokenized/` | openai/gsm8k main test: question + \n\n + answer (evaluation only) | 1 | 0.000B |
| `eval/math500-tokenized/` | HuggingFaceH4/MATH-500 test: problem + \n\n + solution (evaluation only) | 1 | 0.000B |
| `eval/openmathinstruct1-validation-tokenized/` | nvidia/OpenMathInstruct-1, correct_solutions/validation (evaluation only) | 1 | 0.056B |
| `finemath3-tokenized/` | HuggingFaceTB/finemath, finemath-3plus (documents) | 64 | 41.760B |
| `openmathinstruct1-tokenized/` | nvidia/OpenMathInstruct-1, correct_solutions/train: question + \n\n + generated_solution | 1 | 0.369B |
| `openmathinstruct2-tokenized/` | nvidia/OpenMathInstruct-2, train (14M rows): problem + \n\n + generated_solution | 32 | 5.672B |
| `tinygsm-tokenized/` | TinyGSM/TinyGSM: question + \n\n + code | 17 | 2.669B |
| **training sources total** | | | **61.471B** |
The paper's mixtures (model-card naming: `as` Algebraic-Stack, `fm3` FineMath-3+, `tg` TinyGSM, `omi1`/`omi2`
OpenMathInstruct-1/2, `mmqa` MetaMathQA, `<n>x` = n passes) are built from these pieces as directories of
symlinks, n copies per pass (`pretraining/passk/data/make_mix.py`, the authors' `mix_datasets.py`), and trained for
one epoch at 512 x 2048 tokens per step.
Note: the authors' OpenMathInstruct-2 script globs `*.parquet`, which on the hub also matches the
`train_1M/2M/5M` subset files (duplicates of rows in `train`); this copy holds the 14M-row `train` split
only, as the paper describes.
## Fetching
```bash
bash PRETRAINING-PASSK/OLMo-mathonly/pretraining/passk/flame/data/download_from_hf.sh /tmp/$USER-mathonly-data # parallel resumable curl
```
Licenses follow the upstream datasets (ODC-By for FineMath, the Proof-Pile-2 subset licenses for
Algebraic-Stack, CC-BY-NC/CC-BY for the NVIDIA OpenMathInstruct sets, MIT for TinyGSM/MetaMathQA, MIT for
GSM8K and MATH).
|