![image/png](https://cdn-uploads.huggingface.co/production/uploads/64adfd277b5ff762771e4571/ijVNJF9HePkQCjejXZLcI.png) Back from the dead! Hoping to make something cool to share with everyone! Introducing Crimson Dawn! Built atop the impressive [Mistral-Nemo-Base-2407](https://huggingface.co/mistralai/Mistral-Nemo-Base-2407); Crimson Dawn was built with the idea that AI should not be a boring bland generic assistant, but something that you can connect with on a more personal level. Something that can be interesting in a Roleplay, but useful as an assistant too. # Quants! [exl2](https://huggingface.co/lucyknada/Epiculous_Crimson_Dawn-V0.1-exl2) / [gguf](https://huggingface.co/mradermacher/Crimson_Dawn-V0.1-GGUF) ## Prompting Crimson Dawn was trained with the Mistral Instruct template, therefore it should be prompted in a similar way that you would prompt any other mistral based model. If you are using GGUF I strongly advise using ChatML, for some reason that quantization performs better using ChatML. ``` "[INST] Prompt goes here [/INST]<\s>" ``` ### Context and Instruct [Magnum-123B-Context.json](https://files.catbox.moe/rkyqwg.json)
[Magnum-123B-Instruct.json](https://files.catbox.moe/obb5oe.json)
~~[Mistral-Custom-Context.json](https://files.catbox.moe/l9w0ry.json)~~
~~[Mistral-Custom-Instruct.json](https://files.catbox.moe/9xiiwb.json)~~
*** NOTE ***
There have been reports of the quantized model misbehaving with the mistral prompt, if you are seeing issues it may be worth trying ChatML Context and Instruct templates. ### Current Top Sampler Settings [Violet_Twilight-Nitral-Special](https://files.catbox.moe/ot54u3.json)- Considered the best settings!
[Crimson_Dawn-Nitral-Special](https://files.catbox.moe/8xjxht.json)
[Crimson_Dawn-Magnum-Style](https://files.catbox.moe/lc59dn.json) ### Tokenizer If you are using SillyTavern, please set the tokenizer to API (WebUI/ koboldcpp) ## Training Training was done twice over 2 epochs each on two 2x [NVIDIA A6000 GPUs](https://www.nvidia.com/en-us/design-visualization/rtx-a6000/) using LoRA. A two-phased approach was used in which the base model was trained 2 epochs on RP data, the LoRA was then applied to base. Finally, the new modified base was trained 2 epochs on instruct, and the new instruct LoRA was applied to the modified base, resulting in what you see here. [](https://github.com/axolotl-ai-cloud/axolotl) ## Special Thanks Special thanks to my friends over at Anthracite! Without their help and Kalomaze starting the synthetic data script, none of this would have been possible. Also want to thank my friends in The Chaotic Neutrals for their friendship, support, and guidance.