Nexesenex/Llama_3.x_70b_SmarTricks_v1.30_flat

about

Successor of SmartTricks 1.30.

• Flat means a single level merge, not a 3 stages merges.
• This, to avoid models being "averaged over and over" until it becomes too "soupy", even if benching right.
• In other words, limiting the number of consecutive merges to preserve as much as possible the features of each model.

---

benchs

PPL much better than Smartricks v1.30.

IK_LLama.CPP Benchs in IQ6_K:

• PPL-512 WikiEng Text 564 : 3.33
• ARC-C 299 : 60.20
• ARC-E 570 : 82.28
• Hellaswag 200 : 88
• Winogrande 1263 : 81.53
• MMLU : 46.64

MMLU results are much lower than they should be on LlamaCPP. This is a constant quirk since 2024.

---

merge

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

Merge Details

Merge Method

This model was merged using the Model Stock merge method using huihui-ai/Llama-3.3-70B-Instruct-abliterated as a base.

Models Merged

The following models were included in the merge:

• hitachi-nlp/Llama-3.1-70B-FLDx2
• TheDrummer/Fallen-Llama-3.3-R1-70B-v1
• huihui-ai/Llama-3.1-Nemotron-70B-Instruct-HF-abliterated
• huihui-ai/Llama-3.1-Tulu-3-70B-abliterated

Configuration

The following YAML configuration was used to produce this model:

merge_method: model_stock
models:
  - model: TheDrummer/Fallen-Llama-3.3-R1-70B-v1
    parameters:
      weight: 1.0
  - model: huihui-ai/Llama-3.1-Nemotron-70B-Instruct-HF-abliterated
    parameters:
      weight: 1.0
  - model: huihui-ai/Llama-3.1-Tulu-3-70B-abliterated
    parameters:
      weight: 1.0
  - model: hitachi-nlp/Llama-3.1-70B-FLDx2
    parameters:
      weight: 1.0
base_model: huihui-ai/Llama-3.3-70B-Instruct-abliterated
dtype: bfloat16
out_dtype: bfloat16
parameters:
  int8_mask: true
  normalize: true
  rescale: false
  filter_wise: false
  smooth: false
  allow_negative_weights: false
chat_template: auto
tokenizer:
  source: union