---
license: apache-2.0
---

# XGen

Official research release for the family of **XGen** models (`7B`) by Salesforce AI Research:

*Title*: [Long Sequence Modeling with XGen: A 7B LLM Trained on 8K Input Sequence Length](https://blog.salesforceairesearch.com/xgen/)

## Models

* [XGen-7B-4K-Base](https://huggingface.co/Salesforce/xgen-7b-4k-base): Pretrained XGen-7B model trained under 4K sequence length.
  * License: Apache-2.0
* [XGen-7B-8K-Base](https://huggingface.co/Salesforce/xgen-7b-8k-base): Pretrained XGen-7B model trained under 8K sequence length.
  * License: Apache-2.0

The training data for the models are tokenized with OpenAI Tiktoken library.
To use this model, install tiktoken library via `pip`:

```sh
pip install tiktoken
```

The models can be used as auto-regressive samplers as follows:

```python
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Salesforce/xgen-7b-8k-base", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("Salesforce/xgen-7b-8k-base", torch_dtype=torch.bfloat16)
inputs = tokenizer("The world is", return_tensors="pt")
sample = model.generate(**inputs, max_length=128)
print(tokenizer.decode(sample[0]))
```

## Citation

```bibtex
@misc{XGen,
  title={Long Sequence Modeling with XGen: A 7B LLM Trained on 8K Input Sequence Length},
  author={Salesforce AI Research},
  howpublished={Salesforce AI Research Blog},
  year={2023},
  url={https://blog.salesforceairesearch.com/xgen-7b/}
}
```