allenai/OLMo-7B-0424

Model Card for OLMo 7B April 2024

For transformers versions v4.40.0 or newer, please use OLMo 7B April 2024 HF instead.

OLMo 7B April 2024 is the latest version of the original OLMo 7B model rocking a 24 point increase in MMLU, among other evaluations improvements, from an improved version of the Dolma dataset and staged training.

OLMo is a series of Open Language Models designed to enable the science of language models. The OLMo models are trained on the Dolma dataset. We release all code, checkpoints, logs, and details involved in training these models.

Model Details

The core models released in this batch are the following:

Size	Training Tokens	Layers	Hidden Size	Attention Heads	Context Length
OLMo 1B	3 Trillion	16	2048	16	2048
OLMo 7B	2.5 Trillion	32	4096	32	2048
OLMo 7B Twin 2T	2 Trillion	32	4096	32	2048
OLMo 7B April 2024	2.05 Trillion	32	4096	32	4096

Note: OLMo 7B April 2024 also includes QKV clipping.

[Coming soon] We are releasing many checkpoints for these models, for every 1000 training steps. The naming convention is step1000-tokens4B.

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

import hf_olmo # pip install ai2-olmo
olmo = AutoModelForCausalLM.from_pretrained("allenai/OLMo-7B-0424", revision="step1000-tokens4B")

All revisions/branches are listed in the file revisions.txt. 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-7B-0424")
branches = [b.name for b in out.branches]

A few revisions were lost due to an error, but the vast majority are present.

Model Description

• Developed by: Allen Institute for AI (AI2)
• Supported by: Databricks, Kempner Institute for the Study of Natural and Artificial Intelligence at Harvard University, AMD, CSC (Lumi Supercomputer), UW
• 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: olmo at allenai dot org. Press: press at allenai dot org
• Date cutoff: Oct. 2023, with most data from Feb./March 2023 based on Dolma dataset version.

Model Sources

• Project Page: https://allenai.org/olmo
• Repositories:
  • Core repo (training, inference, fine-tuning etc.): https://github.com/allenai/OLMo
  • Evaluation code: https://github.com/allenai/OLMo-Eval
  • Further fine-tuning code: https://github.com/allenai/open-instruct
• Paper: Link
• Technical blog post: https://blog.allenai.org/olmo-1-7-7b-a-24-point-improvement-on-mmlu-92b43f7d269d
• W&B Logs: pretraining, annealing

Uses

Inference

Note: The OLMo models will shortly be included in Transformers. When the PR is merged, you will no longer need to use trust_remote_code=True or install ai2-olmo to use the model. Then, install Transformers from source.

Quickly get inference running with the following required installation:

pip install ai2-olmo

Now, proceed as usual with HuggingFace:

import hf_olmo

from transformers import AutoModelForCausalLM, AutoTokenizer
olmo = AutoModelForCausalLM.from_pretrained("allenai/OLMo-7B-0424")
tokenizer = AutoTokenizer.from_pretrained("allenai/OLMo-7B-0424")
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 the first step to build natural language generation...'

Alternatively, with the pipeline abstraction:

import hf_olmo

from transformers import pipeline
olmo_pipe = pipeline("text-generation", model="allenai/OLMo-7B-0424")
print(olmo_pipe("Language modeling is "))
>> 'Language modeling is a branch of natural language processing that aims to...'

Or, you can make this slightly faster by quantizing the model, e.g. AutoModelForCausalLM.from_pretrained("allenai/OLMo-7B-0424", torch_dtype=torch.float16, load_in_8bit=True) (requires bitsandbytes). The quantized model is more sensitive to typing / cuda, so it is recommended to pass the inputs as inputs.input_ids.to('cuda') to avoid potential issues.

Note, you may see the following error if ai2-olmo is not installed correctly, which is caused by internal Python check naming. We'll update the code soon to make this error clearer.

    raise ImportError(
ImportError: This modeling file requires the following packages that were not found in your environment: hf_olmo. Run `pip install hf_olmo`

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 repository:

torchrun --nproc_per_node=8 scripts/train.py {path_to_train_config} \
    --data.paths=[{path_to_data}/input_ids.npy] \
    --data.label_mask_paths=[{path_to_data}/label_mask.npy] \
    --load_path={path_to_checkpoint} \
    --reset_trainer_state

For more documentation, see the GitHub readme.

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

Evaluation

Core model results for the new and original 7B model are found below.

Task	Llama-7b	Llama2-7b	Falcon-7b	Mpt-7b	OLMo-7B	Llama2-13b	OLMo 7B April 2024
arc_c	44.5	48.5	47.5	46.5	48.5	52.8	42.5
arc_e	67.9	69.5	70.4	70.5	65.4	73.7	67.2
boolq	75.4	80.2	74.6	74.2	73.4	82.2	83.7
copa	91.0	86.0	86.0	85.0	90.0	90.0	86.0
hellaswag	76.2	76.8	75.9	77.6	76.4	78.6	75.5
openbookqa	51.2	48.4	53.0	48.6	50.4	51.8	50.0
piqa	77.2	76.7	78.5	77.3	78.4	79.0	77.5
sciq	93.9	94.5	93.9	93.7	93.8	95.5	96.7
winogrande	70.5	69.4	68.9	69.9	67.9	73.5	69.8
truthfulQA (MC2)	33.9	38.5	34.0	33.0	36.0	36.8	35.8
MMLU (5 shot MC)	31.5	45.0	24.0	30.8	28.3	55.5	52.0
GSM8k	10.0	12.0	4.0	4.5	8.5	25.0	29.0
Full average	60.3	62.1	59.2	59.3	59.8	66.2	63.8

And for the 1B model:

task	random	StableLM 2 1.6b*	Pythia 1B	TinyLlama 1.1B	OLMo 1B (ours)
arc_challenge	25	43.81	33.11	34.78	34.45
arc_easy	25	63.68	50.18	53.16	58.07
boolq	50	76.6	61.8	64.6	60.7
copa	50	84	72	78	79
hellaswag	25	68.2	44.7	58.7	62.5
openbookqa	25	45.8	37.8	43.6	46.4
piqa	50	74	69.1	71.1	73.7
sciq	25	94.7	86	90.5	88.1
winogrande	50	64.9	53.3	58.9	58.9
Average	36.11	68.41	56.44	61.48	62.42

*Unlike OLMo, Pythia, and TinyLlama, StabilityAI has not disclosed yet the data StableLM was trained on, making comparisons with other efforts challenging.

Model Details

Data

For training data details, please see the Dolma documentation. This model uses the new April 2024 version with more data sources, better deduplication, and quality filtering. During the annealing phase we use a higher quality subset of Dolma with a linearly decaying learning rate to 0.

Staged training / annealing

In contrast to OLMo 1.0, we trained OLMo April 2024 with a two-stage curriculum:

• In the first stage, we trained the model from scratch on the Dolma 1.7 dataset. We set a cosine learning rate schedule with a warmup of 2500 steps, a peak learning rate of 3e-4, and a cosine decay to 3e-5 after 3T tokens. We cut off this stage after 2T tokens, when the learning rate is still high.
• At this point we switch to the second stage, in which we train on a higher-quality subset of Dolma 1.7 (see below) for another 50B tokens, while linearly decaying the learning rate to 0. Our high-quality subset includes (1) using all available Wikipedia, OpenWebMath and Flan data, (2) removing Dolma CC, CC News, and Megawika, and (3) rebalancing remaining sources to achieve approximately equal proportions of each. See exact token counts and relative proportions of this second stage mix below. Both stages contribute equally to the final performance of the OLMo model. After the first stage, OLMo 1.7 already outperforms OLMo 1.0. The second stage consistently adds 2 to 3 points of performance on top.

Architecture

OLMo 7B architecture with peer models for comparison.

	OLMo 7B	Llama 2 7B	OpenLM 7B	Falcon 7B	PaLM 8B
d_model	4096	4096	4096	4544	4096
num heads	32	32	32	71	16
num layers	32	32	32	32	32
MLP ratio	~8/3	~8/3	~8/3	4	4
LayerNorm type	non-parametric LN	RMSNorm	parametric LN	parametric LN	parametric LN
pos embeddings	RoPE	RoPE	RoPE	RoPE	RoPE
attention variant	full	GQA	full	MQA	MQA
biases	none	none	in LN only	in LN only	none
block type	sequential	sequential	sequential	parallel	parallel
activation	SwiGLU	SwiGLU	SwiGLU	GeLU	SwiGLU
sequence length	2048	4096	2048	2048	2048
batch size (instances)	2160	1024	2048	2304	512
batch size (tokens)	~4M	~4M	~4M	~4M	~1M
weight tying	no	no	no	no	yes

Hyperparameters

AdamW optimizer parameters are shown below.

Size	Peak LR	Betas	Epsilon	Weight Decay
1B	4.0E-4	(0.9, 0.95)	1.0E-5	0.1
7B	3.0E-4	(0.9, 0.99)	1.0E-5	0.1

Optimizer settings comparison with peer models.

	OLMo 7B	Llama 2 7B	OpenLM 7B	Falcon 7B
warmup steps	5000	2000	2000	1000
peak LR	3.0E-04	3.0E-04	3.0E-04	6.0E-04
minimum LR	3.0E-05	3.0E-05	3.0E-05	1.2E-05
weight decay	0.1	0.1	0.1	0.1
beta1	0.9	0.9	0.9	0.99
beta2	0.95	0.95	0.95	0.999
epsilon	1.0E-05	1.0E-05	1.0E-05	1.0E-05
LR schedule	linear	cosine	cosine	cosine
gradient clipping	global 1.0	global 1.0	global 1.0	global 1.0
gradient reduce dtype	FP32	FP32	FP32	BF16
optimizer state dtype	FP32	most likely FP32	FP32	FP32

Bias, Risks, and Limitations

Like any base language model or fine-tuned model without safety filtering, it is relatively easy for a user to prompt these models to generate harmful and generally sensitive content. Such content can also be produced unintentionally, especially in the case of bias, so we recommend users consider the risks of applications of this technology.

Otherwise, many facts from OLMo or any LLM will often not be true, so they should be checked.

Citation

BibTeX:

@article{Groeneveld2023OLMo,
  title={OLMo: Accelerating the Science of Language Models},
  author={Groeneveld, Dirk and Beltagy, Iz and Walsh, Pete and Bhagia, Akshita and Kinney, Rodney and Tafjord, Oyvind and Jha, Ananya Harsh and Ivison, Hamish and Magnusson, Ian and Wang, Yizhong and Arora, Shane and Atkinson, David and Authur, Russell and Chandu, Khyathi and Cohan, Arman and Dumas, Jennifer and Elazar, Yanai and Gu, Yuling and Hessel, Jack and Khot, Tushar and Merrill, William and Morrison, Jacob and Muennighoff, Niklas and Naik, Aakanksha and Nam, Crystal and Peters, Matthew E. and Pyatkin, Valentina and Ravichander, Abhilasha and Schwenk, Dustin and Shah, Saurabh and Smith, Will and Subramani, Nishant and Wortsman, Mitchell and Dasigi, Pradeep and Lambert, Nathan and Richardson, Kyle and Dodge, Jesse and Lo, Kyle and Soldaini, Luca and Smith, Noah A. and Hajishirzi, Hannaneh},
  journal={Preprint},
  year={2024}
}

APA:

Groeneveld, D., Beltagy, I., Walsh, P., Bhagia, A., Kinney, R., Tafjord, O., Jha, A., Ivison, H., Magnusson, I., Wang, Y., Arora, S., Atkinson, D., Authur, R., Chandu, K., Cohan, A., Dumas, J., Elazar, Y., Gu, Y., Hessel, J., Khot, T., Merrill, W., Morrison, J., Muennighoff, N., Naik, A., Nam, C., Peters, M., Pyatkin, V., Ravichander, A., Schwenk, D., Shah, S., Smith, W., Subramani, N., Wortsman, M., Dasigi, P., Lambert, N., Richardson, K., Dodge, J., Lo, K., Soldaini, L., Smith, N., & Hajishirzi, H. (2024). OLMo: Accelerating the Science of Language Models. Preprint.

Model Card Contact

For errors in this model card, contact Nathan, {nathanl} at allenai dot org.