OLMo: Open Language Model

OLMo is a repository for training and using AI2's state-of-the-art open language models. It is designed by scientists, for scientists.

Installation

First, install PyTorch following the instructions specific to your operating system.

For training and fine-tuning, we recommend installing from source:

git clone https://github.com/allenai/OLMo.git
cd OLMo
pip install -e .[all]

You can also install from PyPI with:

pip install ai2-olmo

Pretraining

OLMo pretraining follows a two-stage training procedure. In the first stage, we train on large amounts of mostly web-based data: OLMo-mix-1124 In the second stage, we train on a smaller amount of high-quality, targeted data: Dolmino-mix-1124

You can find all the checkpoints, at minimum every 1000 training steps, on Huggingface:

• Huggingface for the 7B variant
• Huggingface for the 13B variant

Steps to reproduce

To reproduce any of the training processes described below, run this:

torchrun --nproc_per_node=8 scripts/train.py {path_to_train_config}

For the training config, use any of the configs listed below.

If you want to override any of the settings in the training config without having to write a new config every time, you can do this:

torchrun --nproc_per_node=8 scripts/train.py {path_to_train_config} \
  --setting1=value \
  --setting2=value \
  --setting3.subsetting1=value

The training configs below refer to training data that gets streamed in live over HTTP. To reproduce at large scale, we recommend downloading the files locally and changing the paths to point to your local file system.

Note: Some of the files that the training configs refer to are still being uploaded (as of 2024-11-27). They should all appear in the next few days as the uploads complete.

Stage 1

Stage 1 is the biggest stage, where we train on 4T or 5T tokens on largely web-based data.

Stage 2 for the 7B

For the 7B model, we train three times with different data order on 50B high quality tokens, and then average ("soup") the models.

The training configs linked here are set up to download the latest checkpoint after stage 1, and start training from there.

Stage 2 for the 13B

For the 13B model, we train three times with different data order on 100B high quality tokens, and one more time on 300B high quality tokens. Then we average ("soup") the models.

The training configs linked here are set up to download the latest checkpoint after stage 1, and start training from there.

Instruction tuned variants

For instruction tuned variants of these models, go to

• OLMo2 7B Instruct
• OLMo2 13B Instruct

Inference

You can use our Hugging Face integration to run inference on the OLMo Transformers checkpoints:

from transformers import AutoModelForCausalLM, AutoTokenizer
olmo = AutoModelForCausalLM.from_pretrained("allenai/OLMo-2-1124-7B")
tokenizer = AutoTokenizer.from_pretrained("allenai/OLMo-2-1124-7B")
message = ["Language modeling is "]
inputs = tokenizer(message, return_tensors='pt', return_token_type_ids=False)
# optional verifying cuda
# inputs = {k: v.to('cuda') for k,v in inputs.items()}
# olmo = olmo.to('cuda')
response = olmo.generate(**inputs, max_new_tokens=100, do_sample=True, top_k=50, top_p=0.95)
print(tokenizer.batch_decode(response, skip_special_tokens=True)[0])

Alternatively, with the Hugging Face pipeline abstraction:

from transformers import pipeline
olmo_pipe = pipeline("text-generation", model="allenai/OLMo-2-1124-7B")
print(olmo_pipe("Language modeling is"))

Quantization

olmo = AutoModelForCausalLM.from_pretrained("allenai/OLMo-2-1124-7B", torch_dtype=torch.float16, load_in_8bit=True)  # requires bitsandbytes

The quantized model is sensitive to input types and CUDA handling. To avoid potential issues, we recommend explicitly converting input IDs to CUDA using: inputs.input_ids.to('cuda')

Evaluation

Additional tools for evaluating OLMo models are available at the OLMo Eval repo.

Citing

@article{OLMo,
  title={OLMo: Accelerating the Science of Language Models},
  author={Dirk Groeneveld and Iz Beltagy and Pete Walsh and Akshita Bhagia and Rodney Kinney and Oyvind Tafjord and A. Jha and Hamish Ivison and Ian Magnusson and Yizhong Wang and Shane Arora and David Atkinson and Russell Authur and Khyathi Raghavi Chandu and Arman Cohan and Jennifer Dumas and Yanai Elazar and Yuling Gu and Jack Hessel and Tushar Khot and William Merrill and Jacob Daniel Morrison and Niklas Muennighoff and Aakanksha Naik and Crystal Nam and Matthew E. Peters and Valentina Pyatkin and Abhilasha Ravichander and Dustin Schwenk and Saurabh Shah and Will Smith and Emma Strubell and Nishant Subramani and Mitchell Wortsman and Pradeep Dasigi and Nathan Lambert and Kyle Richardson and Luke Zettlemoyer and Jesse Dodge and Kyle Lo and Luca Soldaini and Noah A. Smith and Hanna Hajishirzi},
  year={2024},
  url={https://api.semanticscholar.org/CorpusID:267365485},
  journal={arXiv preprint},
}