Model Card: Yarn-Solar-10b-64k

Preprint (arXiv)
GitHub

Model Description

Yarn-Solar-10b-64k is a state-of-the-art language model for long context, further pretrained on two billion long context tokens using the YaRN extension method. It is an extension of SOLAR-10.7B-v1.0 and supports a 64k token context window.

To use, pass trust_remote_code=True when loading the model, for example

model = AutoModelForCausalLM.from_pretrained("NousResearch/Yarn-Solar-10b-64k",
  attn_implementation="flash_attention_2",
  torch_dtype=torch.bfloat16,
  device_map="auto",
  trust_remote_code=True)

In addition you will need to use the latest version of transformers

pip install git+https://github.com/huggingface/transformers

Benchmarks

Long context benchmarks:

Model	Context Window	4k PPL	8k PPL	16k PPL	32k PPL	64k PPL
Mistral-7B-v0.1	8k	3.09	2.96	-	-	-
Yarn-Mistral-7b-64k	64k	3.18	3.04	2.65	2.44	2.20
Yarn-Mistral-7b-128k	128k	3.21	3.08	2.68	2.47	2.24
SOLAR-10.7B-v1.0	4k	3.07	-	-	-	-
Yarn-Solar-10b-32k	32k	3.09	2.95	2.57	2.31	-
Yarn-Solar-10b-64k	64k	3.13	2.99	2.61	2.34	2.15

Short context benchmarks showing that quality degradation is minimal:

Model	Context Window	ARC-c	Hellaswag	MMLU	Truthful QA
Mistral-7B-v0.1	8k	59.98	83.31	64.16	42.15
Yarn-Mistral-7b-64k	64k	59.38	81.21	61.32	42.50
Yarn-Mistral-7b-128k	128k	58.87	80.58	60.64	42.46
SOLAR-10.7B-v1.0	4k	61.95	84.60	65.48	45.04
Yarn-Solar-10b-32k	32k	59.64	83.65	64.36	44.82
Yarn-Solar-10b-64k	64k	59.21	83.08	63.57	45.70

Collaborators

• bloc97: Methods, paper and evals
• @theemozilla: Methods, paper, model training, and evals
• @EnricoShippole: Model training
• honglu2875: Paper and evals

The authors would like to thank LAION AI for their support of compute for this model. It was trained on the JUWELS supercomputer.