GamerUntouch's picture
Update README.md
67a6497
|
raw
history blame
1.11 kB
metadata
license: other

See LICENSE file for license. This is a collection of merged, then converted to 4bit LLaMA models trained on the storytelling dataset I used for the storytelling LoRAs.

UPDATE: 04/04 Cleaned data and retrained to 32 groupsize and safetensors. Formatting oddities seem to have been wiped out. Format: Nothing notable, chapters separated by *** therefore may mess some things up.

UPDATE: 2024-04-18 Retrained and merged using updated LoRAs.

To merge and convert, used:

transformers 4.28.1.
gptq cuda branch 5731aa1
llamacpp master branch 8944a13

Notes for usage.

- These models are not instruct LoRAs. They are designed to supplement existing story data.
- There will likely be some bleedthrough on locations and names, this is especially notable if you use with very little context.
- There isn't any large notable formatting, ### seperated stories in the dataset, and *** seperated chapters.

Currently transferring models over.

7B safetensors 4bit - 
7B ggml 4bit -

13B safetensors 4bit - UPLOADED
13B ggml 4bit -

30B safetensors 4bit - 
30B ggml 4bit -