allenai/OLMo-2-0325-32B

Model Details

Model Card for OLMo 2 32B

We introduce OLMo 2 32B, the largest model in the OLMo 2 family. OLMo 2 was pre-trained on OLMo-mix-1124 and uses Dolmino-mix-1124 for mid-training.

OLMo 2 is the latest in a series of Open Language Models designed to enable the science of language models. We have released all code, checkpoints, logs, and associated training details on GitHub.

Size	Training Tokens	Layers	Hidden Size	Attention Heads	Context Length
OLMo 2-7B	4 Trillion	32	4096	32	4096
OLMo 2-13B	5 Trillion	40	5120	40	4096
OLMo 2-32B	6 Trillion	64	5120	40	4096

The core models released in this batch include the following:

Stage	OLMo 2 32B	OLMo 2 13B	OLMo 2 7B
Base Model	allenai/OLMo-2-0325-32B	allenai/OLMo-2-1124-13B	allenai/OLMo-2-1124-7B
SFT	allenai/OLMo-2-0325-32B-SFT	allenai/OLMo-2-1124-13B-SFT	allenai/OLMo-2-1124-7B-SFT
DPO	allenai/OLMo-2-0325-32B-DPO	allenai/OLMo-2-1124-13B-DPO	allenai/OLMo-2-1124-7B-DPO
Final Models (RLVR)	allenai/OLMo-2-0325-32B-Instruct	allenai/OLMo-2-1124-13B-Instruct	allenai/OLMo-2-1124-7B-Instruct
Reward Model (RM)		(Same as 7B)	allenai/OLMo-2-1124-7B-RM

Installation

OLMo 2 32B is supported in transformers v4.48 or higher:

pip install transformers>=4.48

If using vLLM, you will need to install from the main branch until v0.7.4 is released. Please

Inference

You can use OLMo with the standard HuggingFace transformers library:

from transformers import AutoModelForCausalLM, AutoTokenizer
olmo = AutoModelForCausalLM.from_pretrained("allenai/OLMo-2-0325-32B")
tokenizer = AutoTokenizer.from_pretrained("allenai/OLMo-2-0325-32B")
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])
>> 'Language modeling is  a key component of any text-based application, but its effectiveness...'

For faster performance, you can quantize the model using the following method:

AutoModelForCausalLM.from_pretrained("allenai/OLMo-2-0325-32B",
    torch_dtype=torch.float16,
    load_in_8bit=True)  # Requires bitsandbytes

The quantized model is more sensitive to data types and CUDA operations. To avoid potential issues, it's recommended to pass the inputs directly to CUDA using:

inputs.input_ids.to('cuda')

We have released checkpoints for these models. For pretraining, the naming convention is stage1-stepXXX-tokensYYYB. For checkpoints with ingredients of the soup, the naming convention is stage2-ingredientN-stepXXX-tokensYYYB

To load a specific model revision with HuggingFace, simply add the argument revision:

olmo = AutoModelForCausalLM.from_pretrained("allenai/OLMo-2-0325-32B", revision="step250000-tokens2098B")

Or, you can access all the revisions for the models via the following code snippet:

from huggingface_hub import list_repo_refs
out = list_repo_refs("allenai/OLMo-2-0325-32B")
branches = [b.name for b in out.branches]

Fine-tuning

Model fine-tuning can be done from the final checkpoint (the main revision of this model) or many intermediate checkpoints. Two recipes for tuning are available.

1. Fine-tune with the OLMo-core repository:

torchrun --nproc-per-node=8 ./src/scripts/official/OLMo2-0325-32B-train.py run01

You can override most configuration options from the command-line. For example, to override the learning rate you could launch the script like this:

torchrun --nproc-per-node=8 ./src/scripts/train/OLMo2-0325-32B-train.py run01 --train_module.optim.lr=6e-3

For more documentation, see the GitHub readme.

2. Further fine-tuning support is being developing in AI2's Open Instruct repository. Details are here.

Model Description

• Developed by: Allen Institute for AI (Ai2)
• Model type: a Transformer style autoregressive language model.
• Language(s) (NLP): English
• License: The code and model are released under Apache 2.0.
• Contact: Technical inquiries: [email protected]. Press: [email protected]
• Date cutoff: Dec. 2023.

Model Sources

• Project Page: https://allenai.org/olmo
• Repositories:
  • Core repo (training, inference, fine-tuning etc.): https://github.com/allenai/OLMo-core
  • Evaluation code: https://github.com/allenai/OLMo-Eval
  • Further fine-tuning code: https://github.com/allenai/open-instruct
• Paper: https://arxiv.org/abs/2501.00656

Evaluation

Core model results for OLMo 2 32B are found below.

Model	Training FLOPs	Average	ARC/C	HSwag	WinoG	MMLU	DROP	NQ	AGIEval	GSM8k	MMLUPro	TriviaQA
Open weights models
Llama-2-13B	1.6 · 10^23	54.1	67.3	83.9	74.9	55.7	45.6	38.4	41.5	28.1	23.9	81.3
Mistral-7B-v0.3	n/a	58.8	78.3	83.1	77.7	63.5	51.8	37.2	47.3	40.1	30	79.3
Llama-3.1-8B	7.2 · 10^23	61.8	79.5	81.6	76.6	66.9	56.4	33.9	51.3	56.5	34.7	80.3
Mistral-Nemo-12B	n/a	66.9	85.2	85.6	81.5	69.5	69.2	39.7	54.7	62.1	36.7	84.6
Qwen-2.5-7B	8.2 · 10^23	67.4	89.5	89.7	74.2	74.4	55.8	29.9	63.7	81.5	45.8	69.4
Gemma-2-9B	4.4 · 10^23	67.8	89.5	87.3	78.8	70.6	63	38	57.3	70.1	42	81.8
Mistral-Small-24B	n/a	75.2	93.3	91.3	77.8	80.7	74.4	42.3	69.1	79.7	54.2	88.8
Gemma-2-27B	2.1 · 10^24	71.3	90.7	88.4	74.5	75.7	70.1	44.7	61.5	75.7	44.7	87.4
Qwen-2.5-14B	1.6 · 10^24	72.2	94.0	94.0	80.0	79.3	51.5	37.3	71.0	83.4	52.8	79.1
Qwen-2.5-32B	3.5 · 10^24	74.9	95.6	96.0	84.0	83.1	53.1	37.0	78.0	83.3	59.0	79.9
Partially open models
StableLM-2-12B	2.9 · 10^23	62.2	81.9	84.5	77.7	62.4	55.5	37.6	50.9	62	29.3	79.9
Zamba-2-7B	n/c	65.2	92.2	89.4	79.6	68.5	51.7	36.5	55.5	67.2	32.8	78.8
Fully open models
Amber-7B	0.5 · 10^23	35.2	44.9	74.5	65.5	24.7	26.1	18.7	21.8	4.8	11.7	59.3
OLMo-7B	1.0 · 10^23	38.3	46.4	78.1	68.5	28.3	27.3	24.8	23.7	9.2	12.1	64.1
MAP-Neo-7B	2.1 · 10^23	49.6	78.4	72.8	69.2	58	39.4	28.9	45.8	12.5	25.9	65.1
OLMo-0424-7B	0.9 · 10^23	50.7	66.9	80.1	73.6	54.3	50	29.6	43.9	27.7	22.1	58.8
DCLM-7B	1.0 · 10^23	56.9	79.8	82.3	77.3	64.4	39.3	28.8	47.5	46.1	31.3	72.1
OLMo-2-1124-7B	1.8 · 10^23	62.9	79.8	83.8	77.2	63.7	60.8	36.9	50.4	67.5	31.0	78
OLMo-2-1124-13B	4.6 · 10^23	68.3	83.5	86.4	81.5	67.5	70.7	46.7	54.2	75.1	35.1	81.9
OLMo-2-0325-32B	1.3 · 10^24	72.9	90.4	89.7	78.7	74.9	74.3	50.2	61.0	78.8	43.3	88.0

• Columns ARC/C through NQ represent metrics tracked during OLMo 2 development.
• Columns AGIEval through TriviaQA represent unseen evals.

Model Details

Pretraining

	OLMo 2 32B	OLMo 2 13B	OLMo 2 7B
Pretraining Stage 1	6 trillion tokens (1.5 epoch)	5 trillion tokens (1.2 epochs)	4 trillion tokens (1 epoch)
Pretraining Stage 2	100B tokens (2 runs) 300B tokens (1 run) merged	100B tokens (3 runs) 300B tokens (1 run) merged	50B tokens (3 runs) merged
Post-training	SFT + DPO + PPO (preference mix)	SFT + DPO + PPO (preference mix)	SFT + DPO + PPO (preference mix)

Stage 1: Initial Pretraining

• Dataset: OLMo-mix-1124 (3.9T tokens)
• Coverage: 95%+ of total pretraining budget
• 32B Model: ~1.5 epoch

Stage 2: Fine-tuning

• Dataset: Dolmino-Mix-1124
• Two training mixes:
  • 100B tokens
  • 300B tokens
• Mix composition: 50% high-quality web data + academic/Q&A/instruction/math content

Model Merging

• 32B Model: 3 versions on 100B mix + 1 version on 300B mix, merged for final checkpoint

Bias, Risks, and Limitations

Like any base or fine-tuned language model, AI can be prompted by users to generate harmful and sensitive content. Such content may also be produced unintentionally, especially in cases involving bias, so we recommend that users consider the risks when applying this technology. Additionally, many statements from OLMo or any LLM are often inaccurate, so facts should be verified.

Citation

@misc{olmo20242olmo2furious,
      title={{2 OLMo 2 Furious}},
      author={Team OLMo and Pete Walsh and Luca Soldaini and Dirk Groeneveld and Kyle Lo and Shane Arora and Akshita Bhagia and Yuling Gu and Shengyi Huang and Matt Jordan and Nathan Lambert and Dustin Schwenk and Oyvind Tafjord and Taira Anderson and David Atkinson and Faeze Brahman and Christopher Clark and Pradeep Dasigi and Nouha Dziri and Michal Guerquin and Hamish Ivison and Pang Wei Koh and Jiacheng Liu and Saumya Malik and William Merrill and Lester James V. Miranda and Jacob Morrison and Tyler Murray and Crystal Nam and Valentina Pyatkin and Aman Rangapur and Michael Schmitz and Sam Skjonsberg and David Wadden and Christopher Wilhelm and Michael Wilson and Luke Zettlemoyer and Ali Farhadi and Noah A. Smith and Hannaneh Hajishirzi},
      year={2024},
      eprint={2501.00656},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2501.00656},
}

Model Card Contact

For errors in this model card, contact [email protected].