|
--- |
|
base_model: v000000/L3.1-Niitorm-8B-LATCOSx2 |
|
tags: |
|
- llama |
|
- merge |
|
- llama-cpp |
|
--- |
|
|
|
# Llama-3.1-Niitorm-8B-LATCOSx2 |
|
|
|
 |
|
|
|
# Ordered by quality: |
|
* q8_0 imatrix |
|
* q8_0 |
|
* q6_k imatrix |
|
* q6_k |
|
* q5_k_m imatrix |
|
* q5_k_s imatrix |
|
* q4_k_m imatrix |
|
* q4_k_s imatrix |
|
* iq4_xs imatrix |
|
* q4_0_4_8 imatrix arm |
|
* q4_0_4_4 imatrix arm |
|
|
|
This is a test *RP* model, <b>"v000000/L3.1-Niitorm-8B-t0.0001"</b> but merged one extra time with <b>"akjindal53244/Llama-3.1-Storm-8B"</b>. Using a new merging algorithm I wrote <b>"LATCOS"</b>, which is non linear interpolation and cosine vector similarity between tensors in both magnitude and direction. |
|
This attempts to find the smoothest possible interpolation and make them work more seamlessly together by taking into account the vector direction where both models agree. The model seems a lot smarter even though it's just a bit more of storm, but also more compliant which could be a negative since it's less "dynamic". |
|
|
|
<i>imatrix data randomized bartowski, kalomeze, rp snippets, working gpt4 code, human messaging, story</i> |
