Seeking Help from the Community : Request for Restoring Blagoveshchensk_14B_V3 Safetensors Files !
I really like your work but I have a huge request for you - please, I would really like to get access to download version V3 of the safetensors files of this model of yours. I understand it was there before but now the link to it is not available and there is only version V4 ((can you help with this?
I would be very grateful.
Hello, I'm afraid V3 is no longer available, as I cleaned the profile of outdated models. All that remains of V3 is quants from the honourable mradermacher. Unfortunately, the exact "recipe" for V3 has also been lost, so I can't reproduce V3.
To improve the quality of future merges, I would like to ask you what you liked so much in V3?
Good evening. oh, what a pity that only quanta remained (((((((((((( just V3 (within the family of 14B models) in my tasks shows much greater empathy in the context of understanding the emotional state of the interlocutor in generating an emotionally decorated answer to questions - when playing the role of a given role character (with the same input parameters). V4 (in my tests) shows itself much better in aspects of the logic of thinking and the stability and quality of construction of the text of the answer, as well as the same brevity of the answer and summarization and those tasks, but is inferior to V3 in the emotional empathic response - in my tests. Please tell me, if I create Lora for V4 and then use it for V3, will it work? As I understand it, they had the same Queen 2.5 14B base or something else has changed a lot for the compatibility of Lora between them ? P\s: I would really like to see in V15 those aspects that I was looking forward to in terms of generation? responses with emotions as in V3 and empathic understanding model of the emotional state of the interlocutor.
And then I have another question for the community : friends, maybe someone has previously downloaded weight files of version Blagoveshchensk_14B_V3 on their computer? It would be great if someone could help restore the lost source files of the original version of the development of a good model
Thanks for the detailed feedback! The stability and correctness of the text construction were the main reason for the creation of V4. Understanding emotions and an emotionally tinged response are very interesting metrics, and I will definitely add them to the testing aspects of future models.
I'm not that familiar with the creation of text LoRA yet, so I can't tell you. The architecture of the models is the same - Qwen 2.5. It would probably be more rational to create LoRA to improve the emotionality of V4 than to try to stabilize V3.
I recommend renaming the discussion so that people can see your request for V3 weights.