FlexOlmo is a new kind of LM that unlocks a new paradigm of data collaboration. With FlexOlmo, data owners can contribute to the development of open language models without giving up control of their data. There is no need to share raw data directly, and data contributors can decide when their data is active in the model, deactivate it at any time, and receive attributions whenever it's used for inference.

Installation

We recommend using Python 3.10. First install PyTorch according to the instructions specific to your operating system and hardware.

To install dependencies, run:

git clone https://github.com/allenai/FlexOlmo.git
cd FlexOlmo
conda create -n flexolmo python=3.10
conda activate flexolmo
pip install -e ".[train,beaker,wandb]"  # for training

FlexOlmo is built using OLMo-core. OLMo-core's published Docker images contain all core and optional dependencies. You can also adapt our Dockerfile to build your own images.

Model Summary

FlexOlmo-7x7B-1T (without router training) is a Mixture-of-Experts with 33B total parameters, combining independently trained experts on public-mix, news, math, code, academic texts, creative writing, and Reddit data. The public-mix expert is trained on 1T tokens of public data while the other experts are branched from the public-mix expert and trained on 50B tokens of their respective data.

Corpus	Public	Math	News	Academic	Code	Creative Writing	Reddit
Model	Flex-public-7B-1T	Flex-math-2x7B-1T	Flex-news-2x7B-1T	Flex-pes2o-2x7B-1T	Flex-code-2x7B-1T	Flex-creative-2x7B-1T	Flex-reddit-2x7B-1T

Training scripts

All python training scripts can be found in src/scripts/train. These scripts are meant to be launched with torchrun.

Evaluation

Evaluations are built with OLMES. They can be run as follows:

bash setup_eval_env.sh
NUM_GPUS=2
bash src/scripts/eval/run_eval.sh allenai/FlexOlmo-7x7B-1T mc9 eval_results/ ${NUM_GPUS}

python src/scripts/eval/print_evals.py \
  --base-dir eval_results/ \
  --avg-core \
  --avg-gen \
  --avg-mmlu \
  --avg-mmlu-pro \
  --avg-agi-eval \
  --avg-sciriff \
  --avg-code

Evaluation snapshot

Model	MC9	Gen5	MMLU	MMLU Pro	AGIEval	BBH	Math2	NewsG	PoemG	SciRIFF5	Code4	Avg.
Prev. Public model	68.7	58.8	55.9	26.2	39.9	35.7	8.2	76.0	47.8	48.1	1.1	42.4
Individual
Math	62.5	44.3	50.6	24.1	42.0	45.6	53.1	42.6	28.0	50.7	15.8	41.8
Code	40.5	39.4	29.5	14.5	27.4	38.1	6.0	45.1	28.2	48.0	21.0	30.7
News	46.5	48.6	36.4	15.2	25.7	30.9	2.5	77.7	26.9	47.0	0.0	32.5
Creative Writing	42.7	43.9	31.5	11.6	23.3	27.6	1.7	56.9	67.5	42.4	0.0	31.7
Academic	41.0	45.2	33.8	14.8	24.1	32.4	6.5	51.8	23.0	52.0	0.0	29.5
Reddit	64.7	36.5	56.1	25.5	35.5	19.7	2.5	54.1	8.6	32.7	1.7	30.7
Combined
BTM (top-2)	68.7	57.7	59.4	28.3	43.2	44.3	23.1	73.6	54.4	46.3	24.0	47.6
🔥 FlexOlmo-7x7B-1T	70.4	60.1	60.2	30.5	44.8	46.8	47.9	78.3	66.2	53.8	14.6	52.0
FlexOlmo-7x7B-1T-RT	70.3	60.0	60.2	30.3	45.2	47.2	47.7	77.2	67.6	53.9	13.3	52.2

Note: The evaluation of the individual model refers to the dense model, not the 2x7B MoE model.

Example scripts for experiments in the paper can be found in scripts.

Citation

@misc{flexolmo,
      title={FlexOlmo: Open Language Models for Flexible Data Use}, 
      author={Weijia Shi and Akshita Bhagia and Kevin Farhat and Niklas Muennighoff and Pete Walsh and Jacob Morrison and Dustin Schwenk and Shayne Longpre and Jake Poznanski and Allyson Ettinger and Daogao Liu and Margaret Li and Mike Lewis and Wen-tau Yih and Dirk Groeneveld and Luca Soldaini and Kyle Lo and Noah A. Smith and Luke Zettlemoyer and Pang Wei Koh and Hannaneh Hajishirzi and Ali Farhadi and Sewon Min},
      year={2025},
      eprint={2507.00000},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://allenai.org/papers/flexolmo}, 
}