BigWeave-v12-90b / README.md
llmixer's picture
Adding Evaluation Results (#2)
8a9b813 verified
metadata
language:
  - en
license: llama2
tags:
  - Xwin
  - Euryale 1.3
  - Platypus2
  - WinterGoddess
  - frankenmerge
  - dare
  - ties
  - 90b
pipeline_tag: conversational
model-index:
  - name: BigWeave-v12-90b
    results:
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: AI2 Reasoning Challenge (25-Shot)
          type: ai2_arc
          config: ARC-Challenge
          split: test
          args:
            num_few_shot: 25
        metrics:
          - type: acc_norm
            value: 68.09
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=llmixer/BigWeave-v12-90b
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: HellaSwag (10-Shot)
          type: hellaswag
          split: validation
          args:
            num_few_shot: 10
        metrics:
          - type: acc_norm
            value: 87.7
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=llmixer/BigWeave-v12-90b
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MMLU (5-Shot)
          type: cais/mmlu
          config: all
          split: test
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 69.41
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=llmixer/BigWeave-v12-90b
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: TruthfulQA (0-shot)
          type: truthful_qa
          config: multiple_choice
          split: validation
          args:
            num_few_shot: 0
        metrics:
          - type: mc2
            value: 61.35
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=llmixer/BigWeave-v12-90b
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: Winogrande (5-shot)
          type: winogrande
          config: winogrande_xl
          split: validation
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 81.22
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=llmixer/BigWeave-v12-90b
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: GSM8k (5-shot)
          type: gsm8k
          config: main
          split: test
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 47.38
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=llmixer/BigWeave-v12-90b
          name: Open LLM Leaderboard

BigWeave v12 90B

The BigWeave models aim to identify merge settings equaling or surpassing the performance of Goliath-120b. The version number merely tracks various attempts and is not a quality indicator. Only results demonstrating good performance are retained and shared.

This version is a DARE-TIES merge of two passthrough merges: Xwin-LM-70b-v0.1 + Euryale-1.3-70b (BigWeave v6) and Platypus2-70b-instruct + WinterGoddess-1.4x-70b (BigWeave v8). Both models individually show strong performance, and the merged model achieves even lower perplexity than each model separately.

The 90b size allows for 4bit quants to fit into 48GB of VRAM.

Prompting Format

Vicuna and Alpaca.

Merge process

The models used in the merge are Xwin-LM-70b-v0.1, Euryale-1.3-70b, Platypus2-70b-instruct and WinterGoddess-1.4x-70b.

Merge configuration: ``` slices: - sources: - model: Xwin-LM/Xwin-LM-70B-V0.1 layer_range: [0,12] - sources: - model: Sao10K/Euryale-1.3-L2-70B layer_range: [9,14] - sources: - model: Xwin-LM/Xwin-LM-70B-V0.1 layer_range: [12,62] - sources: - model: Sao10K/Euryale-1.3-L2-70B layer_range: [54,71] - sources: - model: Xwin-LM/Xwin-LM-70B-V0.1 layer_range: [62,80] merge_method: passthrough dtype: float16

slices: - sources: - model: garage-bAInd/Platypus2-70B-instruct layer_range: [0,12] - sources: - model: Sao10K/WinterGoddess-1.4x-70B-L2 layer_range: [9,14] - sources: - model: garage-bAInd/Platypus2-70B-instruct layer_range: [12,62] - sources: - model: Sao10/WinterGoddess-1.4x-70B-L2 layer_range: [54,71] - sources: - model: garage-bAInd/Platypus2-70B-instruct layer_range: [62,80] merge_method: passthrough dtype: float16

models: - model: llmixer/BigWeave-v8-90b parameters: weight: 0.5 density: 0.25 merge_method: dare_ties base_model: llmixer/BigWeave-v6-90b dtype: float16


# Acknowledgements
[@Xwin-LM](https://huggingface.co/Xwin-LM) For creating Xwin

[@Sao10K](https://huggingface.co/Sao10K) For creating Euryale and WinterGoddess

[@garage-bAInd](https://huggingface.co/garage-bAInd) For creating Platypus2

[@alpindale](https://huggingface.co/alpindale) For creating the original Goliath

[@chargoddard](https://huggingface.co/chargoddard) For developing [mergekit](https://github.com/cg123/mergekit).

# [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard)
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/details_llmixer__BigWeave-v12-90b)

|             Metric              |Value|
|---------------------------------|----:|
|Avg.                             |69.19|
|AI2 Reasoning Challenge (25-Shot)|68.09|
|HellaSwag (10-Shot)              |87.70|
|MMLU (5-Shot)                    |69.41|
|TruthfulQA (0-shot)              |61.35|
|Winogrande (5-shot)              |81.22|
|GSM8k (5-shot)                   |47.38|