thiflaf/Qwen3-4B-Instruct-Thinking-DARE

merged_model_dare

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/Qwen3-4B-Instruct-2507 as a base.

Models Merged

The following models were included in the merge:

• Qwen/Qwen3-4B-Thinking-2507

Configuration

The following YAML configuration was used to produce this model:

models:
  - model: Qwen/Qwen3-4B-Instruct-2507
    # Base model, không cần tham số riêng
  - model: Qwen/Qwen3-4B-Thinking-2507
    parameters:
      density: 0.6       # Tỷ lệ layer được merge (0.5–0.7 thường tốt)
      weight: 0.5        # Mức đóng góp của Thinking model (0.3–0.6 tùy mục tiêu)
merge_method: dare_ties
base_model: Qwen/Qwen3-4B-Instruct-2507
parameters:
  int8_mask: true        # Giúp giảm RAM/GPU usage khi merge
  rescale: true          # Cân bằng norm sau merge
  normalize: true        # Chuẩn hóa delta trước cộng
dtype: bfloat16