All credits belong to the author.
If you like these also check out FantasiaFoundry's GGUF-Quantization-Script.
It stands for Importance Matrix, a technique used to improve the quality of quantized models.
[1]
The Imatrix is calculated based on calibration data, and it helps determine the importance of different model activations during the quantization process. The idea is to preserve the most important information during quantization, which can help reduce the loss of model performance and lead to better performance, especially when the calibration data is diverse.
[2]
For --imatrix data, included imatrix.dat was used.
Using llama.cpp-b2327:
Base⇢ GGUF(F16)⇢ Imatrix-Data(F16)⇢ GGUF(Imatrix-Quants)
The new IQ3_S quant-option has shown to be better than the old Q3_K_S, so I added that instead of the later. Only supported in koboldcpp-1.59.1 or higher.
If you want any specific quantization to be added, feel free to ask.
This is an attempt to create a model that combines multiple "established" 7Bs and a very small WIP private dataset with Eros' raw creative power. In terms of instruction formats, ChatML and Alpaca work best. The merge isn't purely ChatML, and as such, my previous attempts to integrate it with ChatML strings out of the box were Sisyphean and uninformed.
- sources:
- model: localfultonextractor/sumitest02
layer_range: [0, 32]
- model: tavtav/eros-7b-test
layer_range: [0, 32]
merge_method: slerp
base_model: localfultonextractor/sumitest02
parameters:
t:
- filter: self_attn
value: [0, 0.2, 0.4, 0.55, 0.8]
- filter: mlp
value: [0.7, 0.3, 0.4, 0.3, 0]
- value: 0.37 # fallback for rest of tensors
dtype: float16