Original model: https://huggingface.co/HiroseKoichi/L3-8B-Lunar-Stheno

L3-8B-Lunar-Stheno

L3-8B-Lunaris-v1 is definitely a significant improvement over L3-8B-Stheno-v3.2 in terms of situational awareness and prose, but it's not without issues: the response length can sometimes be very long, causing it to go on a rant; it tends to not take direct action, saying that it will do something but never actually doing it; and its performance outside of roleplay took a hit.

This merge fixes all of those issues, and I'm genuinely impressed with the results. While I did use a SLERP merge to create this model, there was no blending of the models; all I did was replace L3-8B-Stheno-v3.2's weights with L3-8B-Lunaris-v1's.

Experimental Quants Included

There's a full set of quants available, but half of them include an experimental quantization method, which is indicated by the prefix f16 before the quantization level. These quants keep the embeddings and output tensors at f16, but is otherwise the same as the rest.

The experimental variants should theoretically be higher quality than their standard equivalent, but any feedback is welcome.

Details

License: llama3
Instruct Format: llama-3
Context Size: 8K

Models Used

Merge Config

models:
    - model: Sao10K/L3-8B-Stheno-v3.2
    - model: Sao10K/L3-8B-Lunaris-v1
merge_method: slerp
base_model: Sao10K/L3-8B-Stheno-v3.2
parameters:
  t:
    - filter: self_attn
      value: 0
    - filter: mlp
      value: 1
    - value: 0
dtype: bfloat16