EVA-Yi-1.5-9B-32K-V1

EVA Yi 1.5 9B v1

<p> A RP/storywriting focused model, full-parameter finetune of Yi-1.5-9B-32K on mixture of synthetic and natural data.<br> A continuation of nothingiisreal's Celeste 1.x series, made to improve stability and versatility, without losing unique, diverse writing style of Celeste. </p> <p> <h3>Quants: (GGUF is not recommended, lcpp breaks tokenizer fix)</h3> <ul> <li><a href=https://huggingface.co/bartowski/EVA-Yi-1.5-9B-32K-V1-GGUF>IMatrix GGUF by bartowski</a></li> <li><a href=https://huggingface.co/mradermacher/EVA-Yi-1.5-9B-32K-V1-GGUF>Static GGUF by Mradermacher</a></li> <li><a href=https://huggingface.co/bartowski/EVA-Yi-1.5-9B-32K-V1-exl2>EXL2 by bartowski</a></li> </ul> We recommend using original BFloat16 weights, quantization seems to affect Yi significantly more than other model architectures. </p> <p> Prompt format is ChatML.<br> <h3>Recommended sampler values:</h3>

• Temperature: 1
• Min-P: 0.05

<h3>Recommended SillyTavern presets (via CalamitousFelicitousness):</h3>

• Context
• Instruct and System Prompt

</p> <p> <br> <h3> Training data: </h3> <ul> <li>Celeste 70B 0.1 data mixture minus Opus Instruct subset. See that model's <a href=https://huggingface.co/nothingiisreal/L3.1-70B-Celeste-V0.1-BF16>card</a> for details.</li> <li>Kalomaze's Opus_Instruct_25k dataset, filtered for refusals.</li></ul> <h3> Hardware used: </h3> <ul><li>4x3090Ti for 5 days.</li></ul><br> </p> Model was trained by Kearm and Auri. <h4>Special thanks:</h4><ul> <li>to Lemmy, Gryphe, Kalomaze and Nopm for the data</li> <li>to ALK, Fizz and CalamitousFelicitousness for Yi tokenizer fix</li> <li>and to InfermaticAI's community for their continued support for our endeavors</li></ul>