---
datasets:
- emozilla/yarn-train-tokenized-32k-mistral
metrics:
- perplexity
library_name: transformers
license: apache-2.0
language:
- en
---

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

[Preprint (arXiv)](https://arxiv.org/abs/2309.00071)  
[GitHub](https://github.com/jquesnelle/yarn)
![yarn](https://raw.githubusercontent.com/jquesnelle/yarn/solar/data/proofpile-long-small-solar.csv.png)

## 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](https://huggingface.co/upstage/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

```python
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`
```sh
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](https://huggingface.co/mistralai/Mistral-7B-v0.1) | 8k | 3.09 |  2.96 | - | - | - |
| [Yarn-Mistral-7b-64k](https://huggingface.co/NousResearch/Yarn-Mistral-7b-64k) | 64k | 3.18 | 3.04 | 2.65 | 2.44 | 2.20 |
| [Yarn-Mistral-7b-128k](https://huggingface.co/NousResearch/Yarn-Mistral-7b-128k) | 128k | 3.21 | 3.08 | 2.68 | 2.47 | 2.24 |
| [SOLAR-10.7B-v1.0](https://huggingface.co/upstage/SOLAR-10.7B-v1.0) | 4k | 3.07 | - | - | - | - |
| [Yarn-Solar-10b-32k](https://huggingface.co/NousResearch/Yarn-Solar-10b-32k) | 32k | 3.09 | 2.95 | 2.57 | 2.31 | - |
| **[Yarn-Solar-10b-64k](https://huggingface.co/NousResearch/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](https://huggingface.co/mistralai/Mistral-7B-v0.1) | 8k | 59.98 | 83.31 | 64.16 | 42.15 |
| [Yarn-Mistral-7b-64k](https://huggingface.co/NousResearch/Yarn-Mistral-7b-64k) | 64k | 59.38 | 81.21 | 61.32 | 42.50 |
| [Yarn-Mistral-7b-128k](https://huggingface.co/NousResearch/Yarn-Mistral-7b-128k) | 128k | 58.87 | 80.58 | 60.64 | 42.46 |
| [SOLAR-10.7B-v1.0](https://huggingface.co/upstage/SOLAR-10.7B-v1.0) | 4k | 61.95 | 84.60 | 65.48 | 45.04 |
| [Yarn-Solar-10b-32k](https://huggingface.co/NousResearch/Yarn-Solar-10b-32k) | 32k | 59.64 | 83.65 | 64.36 | 44.82 |
| **[Yarn-Solar-10b-64k](https://huggingface.co/NousResearch/Yarn-Solar-10b-64k)** | **64k** |  **59.21** | **83.08** | **63.57** | **45.70** |

## Collaborators

 - [bloc97](https://github.com/bloc97): Methods, paper and evals
 - [@theemozilla](https://twitter.com/theemozilla): Methods, paper, model training, and evals
 - [@EnricoShippole](https://twitter.com/EnricoShippole): Model training
 - [honglu2875](https://github.com/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](https://www.fz-juelich.de/en/ias/jsc/systems/supercomputers/juwels) supercomputer.