KaraKaraWitch/Llama-SoraEToAtetaTegami-3.1X-70B
Llama-SoraEToAtetaTegami-3.1X-70B by KaraKaraWitch is a 70 billion parameter language model merged using the TIES method, based on EVA-LLaMA-3.33-70B-v0.0 and incorporating Blackroot/Mirai-70B-1.0. This model, despite being described as finicky and temperamental, exhibits surprisingly good multilingual capabilities. It is noted for its unique blend of source models, potentially leading to "cursed things" and requiring careful reprompting for optimal responses.
Loading preview...
Model Overview
Llama-SoraEToAtetaTegami-3.1X-70B is a 70 billion parameter language model created by KaraKaraWitch through a merge of pre-trained models using Mergekit. It is based on EVA-UNIT-01/EVA-LLaMA-3.33-70B-v0.0 and integrates Blackroot/Mirai-70B-1.0 using the TIES merge method.
Key Characteristics
- Merge Composition: Combines
EVA-LLaMA-3.33-70B-v0.0(40% density, 0.5 weight) andMirai-70B-1.0(60% density, 0.5 weight). - Multilingual Capabilities: Despite its experimental nature, the model demonstrates surprisingly good multilingual performance.
- Temperamental Behavior: Users should be aware that the model is described as "very finicky and temperamental," often requiring multiple reprompting attempts to achieve desired responses.
- Repetition Penalty: Similar to its Mirai component, the model does not respond well to repetition penalties.
Recommended Settings
For optimal performance, the following inference settings are suggested:
- Temperature: 1.6
- Top A: 0.05
- Min P: 0.05
This model is presented as an experimental merge, exploring the combination of different Llama 3.x variants, and is noted for its potential to generate "cursed things" due to its unique blend.