grimjim's picture
Adding Evaluation Results (#2)
ecd12fe verified
metadata
license: cc-by-nc-4.0
library_name: transformers
tags:
  - mergekit
  - merge
base_model:
  - openlynn/Llama-3-Soliloquy-8B-v2
  - grimjim/Llama-3-Instruct-8B-SPPO-Iter3-SimPO-merge
  - grimjim/llama-3-aaditya-OpenBioLLM-8B
  - grimjim/llama-3-Nephilim-v1-8B
pipeline_tag: text-generation
model-index:
  - name: llama-3-Nephilim-v2-8B
    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: 39.22
            name: strict accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=grimjim/llama-3-Nephilim-v2-8B
          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: 29.9
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=grimjim/llama-3-Nephilim-v2-8B
          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: 9.52
            name: exact match
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=grimjim/llama-3-Nephilim-v2-8B
          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: 6.6
            name: acc_norm
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=grimjim/llama-3-Nephilim-v2-8B
          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: 7.89
            name: acc_norm
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=grimjim/llama-3-Nephilim-v2-8B
          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: 29.35
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=grimjim/llama-3-Nephilim-v2-8B
          name: Open LLM Leaderboard

llama-3-Nephilim-v2-8B

This repo contains is a merge of pre-trained language models created using mergekit. GGUF quants are available here.

Task arithmetic was used to add the contributions of three models to v1 as a base model. The resulting model should be intelligent and attentive to context, and significantly more varied in its outputs compared to v1. The majority contribution was from a merge of two models showcasing recent advances in preference optimization (princeton-nlp/Llama-3-Instruct-8B-SimPO and UCLA-AGI/Llama-3-Instruct-8B-SPPO-Iter3), along with minority contributions from a highly-trained roleplay model and a fine-tuned biomedical model.

Care should be taken when using this model, as it is possible that harmful outputs may be generated. Given that this model is derivative, responsible use is further mandated by the WhiteRabbitNeo Usage Restrictions Extension to the Llama-3 License. This model is further subject to CC-BY-NC-4.0 by default, meaning that commercial use is restricted, barring an alternative licensing agreement.

Tested with 8k context length and Instruct prompting.

Example context templates variants tested with Llama 3 can be downloaded here; their corresponding Instruct prompts can be downloaded here.

During testing, sampler settings were temp=1, minP=0.01, and smooth sampling (factor=0.23, curve=4.32), all of which can be downloaded as a single JSON file.

Built with Meta Llama 3.

Merge Details

Merge Method

This model was merged using the task arithmetic merge method using grimjim/llama-3-Nephilim-v1-8B as a base.

Models Merged

The following models were included in the merge:

Configuration

The following YAML configuration was used to produce this model:

base_model: grimjim/llama-3-Nephilim-v1-8B
dtype: bfloat16
merge_method: task_arithmetic
parameters:
  normalize: false
slices:
- sources:
  - layer_range: [0, 32]
    model: grimjim/llama-3-Nephilim-v1-8B
  - layer_range: [0, 32]
    model: grimjim/Llama-3-Instruct-8B-SPPO-Iter3-SimPO-merge
    parameters:
      weight: 0.9
  - layer_range: [0, 32]
    model: openlynn/Llama-3-Soliloquy-8B-v2
    parameters:
      weight: 0.1
  - layer_range: [0, 32]
    model: grimjim/llama-3-aaditya-OpenBioLLM-8B
    parameters:
      weight: 0.1

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric Value
Avg. 20.41
IFEval (0-Shot) 39.22
BBH (3-Shot) 29.90
MATH Lvl 5 (4-Shot) 9.52
GPQA (0-shot) 6.60
MuSR (0-shot) 7.89
MMLU-PRO (5-shot) 29.35