README.md

metadata

library_name: transformers
tags:
  - mergekit
  - merge
base_model:
  - allknowingroger/Qwenslerp2-14B
  - rombodawg/Rombos-LLM-V2.6-Qwen-14b
  - VAGOsolutions/SauerkrautLM-v2-14b-DPO
  - Qwen/Qwen2.5-14B
  - CultriX/Qwen2.5-14B-Wernicke
model-index:
  - name: Qwestion-14B
    results:
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: IFEval (0-Shot)
          type: HuggingFaceH4/ifeval
          args:
            num_few_shot: 0
        metrics:
          - type: inst_level_strict_acc and prompt_level_strict_acc
            value: 63.18
            name: strict accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=CultriX/Qwestion-14B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: BBH (3-Shot)
          type: BBH
          args:
            num_few_shot: 3
        metrics:
          - type: acc_norm
            value: 48.76
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=CultriX/Qwestion-14B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MATH Lvl 5 (4-Shot)
          type: hendrycks/competition_math
          args:
            num_few_shot: 4
        metrics:
          - type: exact_match
            value: 31.72
            name: exact match
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=CultriX/Qwestion-14B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: GPQA (0-shot)
          type: Idavidrein/gpqa
          args:
            num_few_shot: 0
        metrics:
          - type: acc_norm
            value: 15.77
            name: acc_norm
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=CultriX/Qwestion-14B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MuSR (0-shot)
          type: TAUR-Lab/MuSR
          args:
            num_few_shot: 0
        metrics:
          - type: acc_norm
            value: 17.22
            name: acc_norm
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=CultriX/Qwestion-14B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MMLU-PRO (5-shot)
          type: TIGER-Lab/MMLU-Pro
          config: main
          split: test
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 49.14
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=CultriX/Qwestion-14B
          name: Open LLM Leaderboard
license: apache-2.0
language:
  - en
metrics:
  - accuracy
pipeline_tag: text-generation

merge

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:

• allknowingroger/Qwenslerp2-14B
• rombodawg/Rombos-LLM-V2.6-Qwen-14b
• VAGOsolutions/SauerkrautLM-v2-14b-DPO
• CultriX/Qwen2.5-14B-Wernicke

Configuration

The following YAML configuration was used to produce this model:

models:
  - model: CultriX/Qwen2.5-14B-Wernicke
    parameters:
      weight: 0.55         # Backbone model for conversational ability and GPQA
      density: 0.80        # Retain most critical parameters for stability and strength
  - model: VAGOsolutions/SauerkrautLM-v2-14b-DPO
    parameters:
      weight: 0.20         # High IFEval and MMLU-PRO performance with minimized weaknesses
      density: 0.60        # Focus on impactful parameters for specific benchmarks
  - model: rombodawg/Rombos-LLM-V2.6-Qwen-14b
    parameters:
      weight: 0.25         # Enhanced emphasis on reasoning-heavy tasks like MUSR and MATH
      density: 0.70        # Retain reasoning-intensive parameters for improved benchmarks
  - model: allknowingroger/Qwenslerp2-14B
    parameters:
      weight: 0.15         # General stabilizer for consistency across all tasks
      density: 0.65        # Focus on balance and avoiding redundancy
base_model: Qwen/Qwen2.5-14B
merge_method: dare_ties
parameters:
  normalize: true          # Ensure parameter scale consistency
  int8_mask: true          # Optimize for memory and compute efficiency
dtype: bfloat16
tokenizer_source: Qwen/Qwen2.5-14B-Instruct
adaptive_merge_parameters:
  task_weights:
    IFEval: 1.0            # Maintain high IFEval performance
    MATH: 1.3              # Prioritize reasoning and calculation-heavy tasks
    GPQA: 1.1              # Boost factual recall and reasoning accuracy
    MUSR: 1.2              # Enhance logical reasoning and factual understanding
    MMLU-PRO: 1.0          # Retain consistent knowledge representation
  smoothing_factor: 0.15   # Fine-tune blending for stable transitions between tasks
gradient_clipping: 1.0      # Prevent over-contribution from any single model

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	37.63
IFEval (0-Shot)	63.18
BBH (3-Shot)	48.76
MATH Lvl 5 (4-Shot)	31.72
GPQA (0-shot)	15.77
MuSR (0-shot)	17.22
MMLU-PRO (5-shot)	49.14