from @MedleMedler:
I found out that there other LLM loaders with more settings, which are useful for better control of the ai model.
Top-p (top_p): Also known as nucleus sampling, this parameter controls the cumulative probability distribution cutoff. The model will only consider the top p% of tokens with the highest probabilities for sampling. Reducing this value helps in controlling the generation quality by avoiding low-probability tokens. Default: 0.9.
Top-k (top_k): Limits the number of highest probability tokens considered for each step of the generation. A value of 0 means no limit. This parameter can prevent the model from focusing too narrowly on the top choices, promoting diversity in the generated text. Default: 50.
Repetition Penalty (repetition_penalty): Adjusts the likelihood of tokens that have already appeared in the output, discouraging repetition. Values greater than 1 penalize tokens that have been used, making them less likely to appear again. Default: 1.2.
Trust Remote Code (trust_remote_code): A security parameter that allows or prevents the execution of remote code within loaded models. It is crucial for safely using models from untrusted or unknown sources. Setting this to True may introduce security risks. Default: False.
Torch Data Type (torch_dtype): Specifies the tensor data type for calculations within the model. Options include "float32", "bfloat16", "float16", "float64", or "auto" for automatic selection based on device capabilities. Using "bfloat16" or "float16" can significantly reduce memory usage and increase computation speed on compatible hardware. Default: "auto".
I think that integrating these settings in the GripTape config nodes, will improve the output quality. Not sure if all these settings are useful and if there are other settings which will help to fine tune the ai model more, for better, more/less consistent results.
from @MedleMedler: I found out that there other LLM loaders with more settings, which are useful for better control of the ai model.
Top-p (top_p): Also known as nucleus sampling, this parameter controls the cumulative probability distribution cutoff. The model will only consider the top p% of tokens with the highest probabilities for sampling. Reducing this value helps in controlling the generation quality by avoiding low-probability tokens. Default: 0.9.
Top-k (top_k): Limits the number of highest probability tokens considered for each step of the generation. A value of 0 means no limit. This parameter can prevent the model from focusing too narrowly on the top choices, promoting diversity in the generated text. Default: 50.
Repetition Penalty (repetition_penalty): Adjusts the likelihood of tokens that have already appeared in the output, discouraging repetition. Values greater than 1 penalize tokens that have been used, making them less likely to appear again. Default: 1.2.
Trust Remote Code (trust_remote_code): A security parameter that allows or prevents the execution of remote code within loaded models. It is crucial for safely using models from untrusted or unknown sources. Setting this to True may introduce security risks. Default: False.
Torch Data Type (torch_dtype): Specifies the tensor data type for calculations within the model. Options include "float32", "bfloat16", "float16", "float64", or "auto" for automatic selection based on device capabilities. Using "bfloat16" or "float16" can significantly reduce memory usage and increase computation speed on compatible hardware. Default: "auto".
I think that integrating these settings in the GripTape config nodes, will improve the output quality. Not sure if all these settings are useful and if there are other settings which will help to fine tune the ai model more, for better, more/less consistent results.