Cran-May/PCB-NRSheared-2

merge

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

Merge Details

Merge Method

This model was merged using the DELLA merge method using huihui-ai/Mistral-Small-24B-Instruct-2501-abliterated as a base.

Models Merged

The following models were included in the merge:

• AlSamCur123/Mistral-Small3-24B-InstructContinuedFine
• trashpanda-org/MS-24B-Instruct-Mullein-v0

Configuration

The following YAML configuration was used to produce this model:

# 文件名: pcb_della_merge_12b.yaml
merge_method: della  # 仍然使用 DELLA，因为它是一个强大的自适应剪裁方法
base_model: huihui-ai/Mistral-Small-24B-Instruct-2501-abliterated
models:
  - model: trashpanda-org/MS-24B-Instruct-Mullein-v0
    parameters:
      weight: 1.0
      # PCB策略：限制层影响范围 + 动态竞争平衡
      layers:
        - layers: "8-16"
          parameter_name: density
          value: 0.4
        - layers: "8-16"
          parameter_name: epsilon
          value: 0.15
        - layers: "8-16"
          parameter_name: lambda
          value: 1.5
        - layers: "17-24"
          parameter_name: density
          value: 0.2
      variance_threshold: 0.3

  - model: AlSamCur123/Mistral-Small3-24B-InstructContinuedFine
    parameters:
      weight: 1.0
      # 强化指令理解层
      layers:
        - layers: "0-12"
          parameter_name: density
          value: 0.7
        - layers: "0-12"
          parameter_name: epsilon
          value: 0.05
        - layers: "0-12"
          parameter_name: lambda
          value: 2.0
      variance_threshold: 0.25

  - model: huihui-ai/Mistral-Small-24B-Instruct-2501-abliterated
    parameters:
      weight: 1.0
      # 基模型参数保护策略
      density: 0.9
      layers:
        - layers: "12-24"
          parameter_name: density
          value: 1.0

parameters:
  global_density: 0.55      # 全局剪裁密度（PCB平衡点）
  intra_balance: true
  variance_threshold: 0.2
  epsilon_range: [0.1, 0.2]

tokenizer:
  source: base

generation_config:
  eos_token_id: 2
  pad_token_id: 2
  repetition_penalty: 1.15
  top_k: 40
  temperature: 0.8

# 参数压缩设置 (目标 12-13B 模型)
architecture:
  hidden_size: 3072         # 显著降低 hidden_size (原始 5120 -> 4096 -> 3072)
  intermediate_size: 8256    # 相应调整 intermediate_size (比例保持不变)
  num_attention_heads: 24   # 相应减少 attention heads (比例保持不变)
  num_hidden_layers: 30     # 层数保持 30 层 (适度压缩)