Rei-24B-KTO / README.md
Delta-Vector's picture
Update README.md
db0c365 verified
metadata
library_name: transformers
tags:
  - fine-tuning
  - prose
  - KTO
  - axolotl
  - finetune
  - roleplaying
  - creative-writing
base_model:
  - Delta-Vector/Rei-24B-Base

Rei-KTO-24B

Model banner

Model Information

Rei-KTO-24B
KTO enhanced Painted Fantasy Finetune Creative Prose

A model meant to replicate the style and prose of the Anthropic Claude models, Opus and Sonnet. This model is meant for Roleplaying/Creative-writing, Has some nice smarts without being too sloppy, etc - It's pretty good. Trained in 2 steps, Firstly SFT trained on Zerofata's PaintedFantasy which i found great at anime-otaku-esque characters, and then KTO'd to improve coherency and Instruct Following

Quantized Versions

Available Downloads

  • GGUF FormatFor use with LLama.cpp & Forks (Ty Mradermacher <3)
  • EXL2 FormatFor use with TabbyAPI (Coming Soon!)

Prompting

The model is tuned with V7 Tekken formatting. A typical input would look like this:

[SYSTEM_PROMPT]system_prompt[/SYSTEM_PROMPT][INST]Hi there![/INST]Nice to meet you![INST]Can I ask a question?[/INST]

Training

Training was done in 2 steps, SFT>KTO

Access Configs
 SFT: https://wandb.ai/new-eden/Painted-Fantasy-Rei/artifacts/axolotl-config/config-u7to9d5q/v0/files/axolotl_config_f0p7vnaf.yml 
              KTO : https://wandb.ai/new-eden/Painted-Rei/artifacts/axolotl-config/config-8n37w77c/v0/files/axolotl_config_hvrd2tzn.yml
            

Training

The training was done for 2 epochs using 8 x A100s for 24 hours/p>

Credits

Thank you to Lucy Knada, Ateron, Alicat, Intervitens, Cgato, Kubernetes Bad and the rest of Anthracite.