matsuo-llm-advanced-phase-f5a

merged_model

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/Qwen2.5-7B-Instruct as a base.

Models Merged

The following models were included in the merge:

• ./outputs/phase_d/merged_model
• ./outputs/phase_e2b/merged_model

Configuration

The following YAML configuration was used to produce this model:

# F5-a: DARE-TIES Low Density (Phase D 0.65 + E2-b 0.35, density=0.35)
#
# F4シリーズの教訓:
#   F4-a (density=0.5, D:0.6/E2b:0.4): DB 52.44%, ALF 56%(pre) → 52%(official)
#   F4-b (density=0.5, D:0.7/E2b:0.3): DB 52.99%, ALF 52%(pre)
#   F4-c (density=0.7, D:0.6/E2b:0.4): DB 54.02%, ALF 52%(pre)
#
# パターン: E2-b影響↑ or density↑ → DB↑ ALF↓
# 仮説: density↓ → DAREが小さな変更を除去 → E2-bのDB特化バイアス弱化 → ALF format維持
#
# F5-a戦略: 最低密度(0.35) + Phase D重視(0.65)
# 期待: ALF ≥ 56%(pre-eval), DB ≥ 52.5%

models:
  - model: ./outputs/phase_d/merged_model
    parameters:
      density: 0.35
      weight: 0.65
  - model: ./outputs/phase_e2b/merged_model
    parameters:
      density: 0.35
      weight: 0.35
merge_method: dare_ties
base_model: Qwen/Qwen2.5-7B-Instruct
parameters:
  normalize: true
  int8_mask: true
dtype: bfloat16