GGUF version request (Q4_K_M / Q4_K_S)
Could you please provide a GGUF version of this model? I'd be happy with either Q4_K_M or Q4_K_S quantization. Thank you.
Could you please provide a GGUF version of this model? I'd be happy with either Q4_K_M or Q4_K_S quantization. Thank you.
https://huggingface.co/llmfan46/Qwen3.5-35B-A3B-heretic-GGUF
Could you please provide a GGUF version of this model? I'd be happy with either Q4_K_M or Q4_K_S quantization. Thank you.
Could you please provide a GGUF version of this model? I'd be happy with either Q4_K_M or Q4_K_S quantization. Thank you.
https://huggingface.co/llmfan46/Qwen3.5-35B-A3B-heretic-GGUF
Really good, I'm using Q5.
@llmfan46 Thank you.
@Nuke1229 I must thank you so much, it is actually really good. I was testing it with several images, it follows the system instruction very well. Some times it captions the poses wrong, some times it mentions a genital that is not visible but still better than other uncensored version from other author imo.
I'm using KoboldCPP and I'm not sure if my template is correct or not or if can improve even more but its working nice.
@Nuke1229 I must thank you so much, it is actually really good. I was testing it with several images, it follows the system instruction very well. Some times it captions the poses wrong, some times it mentions a genital that is not visible but still better than other uncensored version from other author imo.
I'm using KoboldCPP and I'm not sure if my template is correct or not or if can improve even more but its working nice.
@Noire1 I use jan.ai, so I’m not entirely sure, but if you’re doing NSFW captions, I recommend running the images through a tagger first. Then, feed both the tags and the image into the LLM. Use the tags as a guide and reinforce it in the system prompt with something like: "Analyze the image and write a description based on these tags, but use your own judgment and verify if the tags are accurate."
To be honest, the system prompt needs to be longer and more detailed. For genitals or small details, you should use 'leading' instructions—hinting that 'there might be [X] present, so look closely.' This forces the model to increase its inference time (thinking time) and focus its attention more effectively.
@Nuke1229 I must thank you so much, it is actually really good. I was testing it with several images, it follows the system instruction very well. Some times it captions the poses wrong, some times it mentions a genital that is not visible but still better than other uncensored version from other author imo.
I'm using KoboldCPP and I'm not sure if my template is correct or not or if can improve even more but its working nice.@Noire1 I use jan.ai, so I’m not entirely sure, but if you’re doing NSFW captions, I recommend running the images through a tagger first. Then, feed both the tags and the image into the LLM. Use the tags as a guide and reinforce it in the system prompt with something like: "Analyze the image and write a description based on these tags, but use your own judgment and verify if the tags are accurate."
To be honest, the system prompt needs to be longer and more detailed. For genitals or small details, you should use 'leading' instructions—hinting that 'there might be [X] present, so look closely.' This forces the model to increase its inference time (thinking time) and focus its attention more effectively.
@Nuke1229 Thank you, I'm using Jan.Ai for the first time and I'm impressed, the quality is even better than KoboldCPP!!!