Noticed some weird behavior in 4bpw exl2, not sure if this is contained or a model related issue. However after seeing some recent bugfixes regarding the targetting of lm training heads among a few other things, i will be attemping to retrain this for comparitive sake.

Base model: (Falcon3-10B)

Prompt format: ChatML

<|im_start|>system
{system_prompt}<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant

The following YAML configuration was used to produce this model: (SLERP merge method)

slices:
  - sources:
      - model: Nitral-Archive/nightwing3-r64-1-latest_test-train-10B
        layer_range: [0, 40]
      - model: Nitral-Archive/nightwing3-r64-2-latest_test-train-10B
        layer_range: [0, 40]
merge_method: slerp
base_model: Nitral-Archive/nightwing3-r64-1-latest_test-train-10B
parameters:
  t:
    - filter: self_attn
      value: [0, 0.5, 0.3, 0.7, 1]
    - filter: mlp
      value: [1, 0.5, 0.7, 0.3, 0]
    - value: 0.420
dtype: bfloat16