listtowardslight/replete-storm-GGUF

Replete Storm is an experimental della merge of the highly useful Replete-LLM-V2-Llama-3.1-8b and the outstanding akjindal53244's Llama-3.1-Storm-8B, which is based in part on Arcee AI's highly scoring Llama-Spark. Credits go to the authors and creators of those models.

This merge is crafted to favor Replete in the early and final layers, but favor Llama Storm midway, as you can see in the density and weight coefficients in the mergekit YAML below. The goal is to leave little fine-tuning for the specific qualities of each model to assert themselves: the quality conversation conversation, function-calling, and truthfulness, together with Replete's GPQA performance.

---

license: llama3.1 datasets: - arcee-ai/The-Tome - Replete-AI/The_Living_AI_Dataset - Replete-AI/code_bagel_hermes-2.5 language: - en base_model: NousResearch/Meta-Llama-3.1-8B

models:
  - model: Replete-AI/Replete-LLM-V2-Llama-3.1-8b
    parameters:
      density: [ 0.80, 0.60, 0.50, 0.40, 0.50, 0.60, 0.80 ]
      epsilon: [ 0.10, 0.10, 0.10, 0.10, 0.10, 0.10, 0.10 ]
      weight:  [ 0.80, 0.65, 0.60, 0.55, 0.60, 0.65, 0.70 ]
      lambda: 0.85
  - model: akjindal53244/Llama-3.1-Storm-8B
    parameters:
      density: [ 0.40, 0.60, 0.70, 0.80, 0.70, 0.60, 0.40 ]
      epsilon: [ 0.10, 0.10, 0.10, 0.10, 0.10, 0.10, 0.10 ]
      weight:  [ 0.40, 0.70, 0.80, 0.90, 0.80, 0.70, 0.40 ]
      lambda: 0.7
merge_method: della
base_model: NousResearch/Meta-Llama-3.1-8B
parameters:
  int8_mask: true
  normalize: true
  rescale: true
dtype: float16
tokenizer_source: union
name: replete-storm