NEXS-llama-3.1-8b-multislerp

NEXS Llama-3.1-8B Multi-SLERP Merge

A multi-SLERP merge of LlamaForCausalLM domain experts covering finance, medical, legal, safety/toxicity, and truthfulness, produced with mergekit. Part of the NEXS multi-SLERP merge collection.

Method

Multi-SLERP (multislerp) performs barycentric spherical interpolation on a hypersphere for more than two models: it projects the models into the tangent space at their weighted Euclidean mean, interpolates, and projects back. Here it is run in task-vector space — each source's delta from the shared base model is computed, the deltas are spherically averaged with equal weight (normalize_weights: true, eps: 1e-8), and the result is added back to the base. Merging was done with mergekit.

Variants had minor vocab differences (128256/128257/128258); tokenizer_source: base reconciled all embeddings to the base tokenizer.

Sources

Base model (task-vector reference): meta-llama/Llama-3.1-8B-Instruct

Merged variants (equal weight 1.0 each):

• mukaj/Llama-3.1-Hawkish-8B
• TsinghuaC3I/Llama-3.1-8B-UltraMedical
• MaziyarPanahi/calme-2.3-legalkit-8b
• K-intelligence/Llama-SafetyGuard-Content-Binary
• HiTZ/Llama-3.1-8B-Instruct-multi-truth-judge

mergekit config

merge_method: multislerp
base_model: meta-llama/Llama-3.1-8B-Instruct
tokenizer_source: base
dtype: float32
out_dtype: bfloat16
parameters:
  normalize_weights: true
  eps: 1.0e-8
models:
  - model: mukaj/Llama-3.1-Hawkish-8B
    parameters: {weight: 1.0}
  - model: TsinghuaC3I/Llama-3.1-8B-UltraMedical
    parameters: {weight: 1.0}
  - model: MaziyarPanahi/calme-2.3-legalkit-8b
    parameters: {weight: 1.0}
  - model: K-intelligence/Llama-SafetyGuard-Content-Binary
    parameters: {weight: 1.0}
  - model: HiTZ/Llama-3.1-8B-Instruct-multi-truth-judge
    parameters: {weight: 1.0}