Honkware's picture
Update README.md
668dd3b
|
raw
history blame
912 Bytes
---
license: other
---
# Manticore-13b-Landmark-GPTQ
## Key Features
- **[Landmark Attention](https://arxiv.org/pdf/2305.16300v1.pdf)**
- **[Large Context Size (~18k)](https://i.ibb.co/tLLGLNc/image.jpg)**
- **[4-Bit Quantization](https://arxiv.org/pdf/2210.17323.pdf)**
## Composition
Manticore-13b-Landmark-GPTQ is a blend of:
- [Manticore-13B](https://huggingface.co/openaccess-ai-collective/manticore-13b)
- [Manticore-13B-Landmark-QLoRA](https://huggingface.co/Honkware/Manticore-13b-Landmark-QLoRA)
- [AutoGPTQ](https://github.com/PanQiWei/AutoGPTQ)
## Using [Oobabooga](https://github.com/oobabooga/text-generation-webui)
- Trust Remote Code - **(Enabled)**
- Add the bos_token to the beginning of prompts - **(Disabled)**
- Truncate the prompt up to this length - **(Increased)**
## Landmark Training Code
See [GitHub](https://github.com/eugenepentland/landmark-attention-qlora) for the training code.