Can, we play with this? There is no model card. 😋
@Nitral-AI , Can we play with this model? There is no model card yet. 🤔
It probably just finished getting fine-tuned. Let the man work...
@Nitral-AI , Can we play with this model? There is no model card yet. 🤔
Of course you can play with it.
It probably just finished getting fine-tuned. Let the man work...
Yea, when i upload the weights from local here it takes up to 8 hours sometimes. So i may or may not have been asleep while it was uploading.
This model seems more coherent - even more so than Hathor_Aleph-L3-8B-v0.72 and Hathor_Fractionate-L3-8B-v.05. Do you get the same feeling? Please, tell me what is different about this model. I like to hear the about your model's developments. 😋
This model seems more coherent than Hathor_Aleph-L3-8B-v0.72 and Hathor_Fractionate-L3-8B-v.05. Do you get the same feeling? Please, tell me what is different about this model. I like to hear the about your model's developments. 😋
Honestly this is basically 0.72, (slightly different hparams and sections of the same datasets) + 0.1 back merged for RP.
I like it! 😋
I have a hypothesis: the incorporation of E(RP) data helps intelligent LLMs (Large Language Models) learn how to creatively synthesize ideas that require imagination and the ability to combine disparate ideas into a new ideas. But, perhaps this is just the ability of an intelligent base model? Or, perhaps this is just the result of training the LLM on other creative data? 🤔
I think there might be some merit to your hypothesis. I know that both turboderp (ExLlamaV2 and CAT model series creator), and KoboldAI (KoboldCPP and PsyFighter model series creator) add medical and veterinarian datasets to the training data, and it improves roleplay capabilities of their models in a multitude of ways, especially in fantasy settings that involve limbs in any sort of manner (character getting chewed by a dragon, or being patched up), or anthropomorphic creatures, or, well... furries as some call it.
Another example is the inclusion of bad words inside the dataset. Without them, it would be impossible for the model to, so to speak, semantically understand how the bad words connect in a sentence and why it is important to avoid them during token soup prediction.
How does ERP data fit into this? Honestly, I don't know, but I would not be surprised if it had some effect on the quality of the output outside ERP itself.