barretts/Llama-xLAM-2-8b-fc-r-slerp-attn_focus-Instruct

merge

This is a merge of pre-trained language models created using mergekit.

Merge Details

Merge Method

This model was merged using the SLERP merge method.

Models Merged

The following models were included in the merge:

• Salesforce/Llama-xLAM-2-8b-fc-r
• meta-llama/Llama-3.1-8B-Instruct

Configuration

The following YAML configuration was used to produce this model:

merge_method: slerp
base_model: meta-llama/Llama-3.1-8B-Instruct
dtype: bfloat16

models:
  - model: meta-llama/Llama-3.1-8B-Instruct
  - model: Salesforce/Llama-xLAM-2-8b-fc-r

parameters:
  t:
    # Strongly favor xLAM-2 in Q/K/V projections (for sharper function-calling)
    - filter: self_attn.q_proj
      value: 0.9
    - filter: self_attn.k_proj
      value: 0.9
    - filter: self_attn.v_proj
      value: 0.9

    # Also boost its final attention out_proj
    - filter: self_attn.out_proj
      value: 0.8

    # Keep the MLP layers more Instruct‑leaning
    - filter: mlp
      value: 0.3

    # Give xLAM’s LM head extra say for function-call formatting
    - filter: lm_head
      value: 0.7

    # All other params at an even blend
    - value: 0.5