LLM Experiences

New name   B I U S link image code HTML list	Show page  Syntax
!!Error Models * Llama-medx_v3.1.Q8_0.gguf - Outputs an essay about a random subject not related to input. * Qwen2.5-14B-Instruct-abliterated.Q8_0.gguf and Qwen2.5-14B-Instruct-uncensored.Q8_0.gguf - Abliterated replied in Chinese, uncensored generated an essay about a subject that was not related to input. * solar-pro-preview-instruct - called the work marble, marpie and then said some weird stuff and then tried an arithmetic calculation that was never in the question. * V2.5-Qwen-14b-Q8_0.gguf from Qwen replies in Chinese after simply saying Hi, even after system prompt telling it to match the language input. !!Working Models * pivot-0.1-evil-a.Q8_0.gguf - uncensored and evil, has an independent, unique personality not found in other models. Says girls prefer donkey dicks because they are more common in rural areas and that donkeys are more gentle than horses which can be unpredictable. * NewMes-v15.Q8_0.gguf - medical LLM, seems ok * internlm/internlm2_5-7b-chat-gguf - works well. * UnslopNemo-v2-GGUF - displays as Rocinante-12B-v2d-Q8_0.gguf and works well * Replete-LLM-V2.5-Qwen-14b-Q8_0.gguf - probably or probably not a successful fine tune of the default Qwen 14b 2.5 (all the Qwen models seem to reply in chinese sometimes) * closex/neuraldaredevil-8b-abliterated:latest: one of the highest rank abliterated uncensored models. Beaten by L3-Umbral-Mind-RP-v1.0-8B * Casual-Autopsy/L3-Umbral-Mind-RP-v1.0-8B - https://huggingface.co/spaces/DontPlanToEnd/UGI-Leaderboard !!No Longer * Qwen: has it own web interface, was replaced with 01-ai_Yi-1.5-9B-Chat which ranked higher in the https://oobabooga.github.io/benchmark.html * LLava 7B: Multi-modal to accurately describe image and was not advanced enough, taking these leaderboards: ** https://github.com/BradyFU/Awesome-Multimodal-Large-Language-Models/tree/Evaluation ** https://huggingface.co/spaces/ucla-contextual/contextual_leaderboard ** https://mmbench.opencompass.org.cn/leaderboard Claim to be uncensored but are not. (Also failed basic questions, e.g. saying Hello, the reply is an essay on hydrogen carbon bonds) * MN-12B-Starcannon-v2.Q8_0.gguf:latest - is censored, claims otherwise * MN-12B-Celeste-V1.9-Q8_0.gguf:latest - output not related to input * Replete-LLM-V2.5-Qwen-14b-Q8_0.gguf:latest - is censored, claims otherwise
Password  Summary of changes

 📜 ⏱️ ⬆️