README.md

metadata

base_model:
  - Qwen/Qwen2.5-14B-Instruct
  - Qwen/Qwen2.5-14B
  - CultriX/Qwen2.5-14B-Hyperionv3_r128
  - CultriX/Qwen2.5-14B_Virtuoso-small-v2-LoRA_r128
library_name: transformers
tags:
  - mergekit
  - merge
license: apache-2.0

Qwen2.5-DeepHyper

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

Merge Details

Merge Method

This model was merged using the DARE TIES merge method using Qwen/Qwen2.5-14B as a base.

Models Merged

The following models were included in the merge:

• Qwen/Qwen2.5-14B-Instruct
• CultriX/Qwen2.5-14B-Hyperionv3_r128
• /root/.cache/huggingface/hub/models--CultriX--Qwen2.5-14B-DeepSeek_r128/snapshots/1bca847f92fced165076d9ac921a1e3ef01fcd7f/
• CultriX/Qwen2.5-14B_Virtuoso-small-v2-LoRA_r128

Configuration

The following YAML configuration was used to produce this model:

base_model: Qwen/Qwen2.5-14B
models:
  # Each adapter was extracted (rank=128) from its respective finetuned model.
  # Their weights are set lower than the full instruct model (which is now the base)
  - model: CultriX/Qwen2.5-14B-Hyperionv3_r128
    parameters:
      weight: 0.9  # Reduced weight relative to base
      density: 0.9

  - model: CultriX/Qwen2.5-14B_Virtuoso-small-v2-LoRA_r128
    parameters:
      weight: 1.0
      density: 1.0

  - model: Qwen/Qwen2.5-14B-Instruct
    parameters:
      weight: 0.75
      density: 0.75

  - model: /root/.cache/huggingface/hub/models--CultriX--Qwen2.5-14B-DeepSeek_r128/snapshots/1bca847f92fced165076d9ac921a1e3ef01fcd7f/
    parameters:
      weight: 1.00
      density: 1.00

# Merging method and overall parameters
merge_method: dare_ties         # Ties corresponding weights across sources.
parameters:
  weight: 1.0                 # Overall scaling factor.
  density: 1.0                # Overall density (typically left at 1.0).
  normalize: true             # Normalize each set of weights before merging.
  int8_mask: true             # Enable masking if using int8 quantized weights.

# Use the instruct tokenizer to ensure compatibility.
tokenizer_source: CultriX/Qwen2.5-14B_Virtuoso-small-v2-LoRA_r128

# Data type for merged weights.
dtype: bfloat16