--- license: other license_name: xt-aurora-license license_link: LICENSE language: - en tags: - conversational - chat - rp - roleplay - friend - slm - small - slim - slender - general - creative co2_eq_emissions: emissions: 5 source: "ML CO2" training_type: "fine-tuning" hardware_used: "1 GTX1060-3GB" --- ![image/png](https://cdn-uploads.huggingface.co/production/uploads/65ca8c3c5495933ab066c33c/-fUpi6P30Lmlx-iasdGaX.png) Buy Me a Coffee at ko-fi.com ![image/png](https://cdn-uploads.huggingface.co/production/uploads/65ca8c3c5495933ab066c33c/ZWbL5O4B5WWIwm0CWAmKU.png) About this model: This model, XT_AURORA-OpenBeta-V0.4, is by us, XeTute. The model was finetuned ontop of the previos beta-verion[XT_AURORA-OpenBeta-V0.3-GGUF]. This version[Beta V0.4] achieves better performance in being[or only acting] as a helpful friend, logical thinking[its still pretty "dumb"] and grammar without any adapters. About XT_AURORA: XT_AURORA is a series of SLMs[Slender Language Models], which all aim to provide a friendly, human-like conversation. The serie is limited by its size[about 1.1B Params], but we still try to get the best possible output. The context-length is 2048 tokens, but it can be upscaled using rope, with the cost being slightly less logic. About this version[V0.4]: * High quality output[sometimes outperforms 3B models in HumanEval], as long as the context size is under 2049 Tokens. * We provide a system prompt[Files and Versions --> chat_template]. The SLM was partly trained using that template, so the output is better if you use the prompt at start. * AURORA expects the chat template to be Vicuna[{{user}}: {some input}\nAURORA: {some output}\n{{user}}]. The model will only work correctly with this format. * Recommended temperature is from 0.3 to 0.5. * Improved chat quality in general chat, roleplaying, etc. * Math and other factual stuff often spits out false "facts", since the model was trained on in-context-learning. It is able to understand some complex words and use them correctly. All in one, AURORA's aim is to provide a digital friend, which is also accessible to humans with low-end devices. Using KoboldCPP, we got the model running[using termux] on a POCO X5 Pro 5G[CPU only, Octa Core]. We saw ~5 Tokens generation per second, ~15 Tokens processing per second. Please support us: X: GitHub: Subdomain on Neocities: We wish you a friendly chat with AURORA <3