Datasets:

allenai

/

dolma

like 862

Follow

Ai2 1.85k

Hi!

Is it possible to release the data mix used for the second stage of Olmo 1.7 training (as mentioned here)?

Even though this is just a combination of the Dolma 1.7 data sources, having this as a separate data split would make it much more convenient to use, and would add a lot of value to the community since (as far as I know) there are currently no public high-quality second-stage-annealing datasets.

In the meantime, a couple of questions for trying to create this mix myself:

1. Where is the table with mix weights under the heading "Staged training data and learning rate" in the blogpost sourced from? I wasn't able to find the mix weights in the Dolma or Olmo 1.7 READMEs (as a sidenote, I also assume the quantities are billions of tokens, not millions as denoted in the table).
2. Was the Dolma data shuffled before sharding, i.e., is it enough to download a subsample of the shards (where a shard is e.g. https://olmo-data.org/dolma-v1_7/c4-filtered/c4-0007.json.gz) to subsample the dataset, or would I have to download the entire dataset, shuffle, then subsample?

Thanks a lot for any help!