Downtown-Case/Deepseek-EVA-32B-DELLA-v1

Deepseek-EVA-DELLA-Merge

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

The intent is to make Deepseek R1 32B feel more like a "base model" like EVA-Gutenberg Qwen, adhering to an existing style when writing.

But consider this a quick A/B test between the new DELLA and SCE merging techniques. This is the DELLA merge, with a average density of 0.6 as found to be ideal for DARE in the SCE paper, but a higher epsilon to allow more density varaition: https://arxiv.org/abs/2408.07990

Merge Details

Merge Method

This model was merged using the DELLA merge method using /home/a/Models/Raw/Qwen_Qwen2.5-32B as a base.

Models Merged

The following models were included in the merge:

• /home/a/Models/Raw/deepseek-ai_DeepSeek-R1-Distill-Qwen-32B
• /home/a/Models/Raw/nbeerbower_EVA-Gutenberg3-Qwen2.5-32B

Configuration

The following YAML configuration was used to produce this model:

models:
  - model: /home/a/Models/Raw/Qwen_Qwen2.5-32B
  # No parameters necessary for base model
  - model: /home/a/Models/Raw/nbeerbower_EVA-Gutenberg3-Qwen2.5-32B
    parameters:
      weight: 0.5
      density: 0.5
  - model: /home/a/Models/Raw/deepseek-ai_DeepSeek-R1-Distill-Qwen-32B
    parameters:
      weight: 0.5
      density: 0.7
merge_method: della
tokenizer:
  source: "union"
base_model: /home/a/Models/Raw/Qwen_Qwen2.5-32B
chat_template: "deepseek"
parameters:
  int8_mask: true
  epsilon: 0.17
  lambda: 1
dtype: bfloat16