README.md

metadata

license: llama3.1
library_name: transformers
tags:
  - mergekit
  - merge
base_model:
  - nvidia/Llama-3.1-Nemotron-70B-Instruct-HF
  - mlabonne/Llama-3-70B-Instruct-abliterated-LORA
model-index:
  - name: Llama-3.1-Nemotron-lorablated-70B
    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: 71.47
            name: strict accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=nbeerbower/Llama-3.1-Nemotron-lorablated-70B
          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: 48.06
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=nbeerbower/Llama-3.1-Nemotron-lorablated-70B
          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: 23.34
            name: exact match
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=nbeerbower/Llama-3.1-Nemotron-lorablated-70B
          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: 0.89
            name: acc_norm
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=nbeerbower/Llama-3.1-Nemotron-lorablated-70B
          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: 14.92
            name: acc_norm
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=nbeerbower/Llama-3.1-Nemotron-lorablated-70B
          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: 43.46
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=nbeerbower/Llama-3.1-Nemotron-lorablated-70B
          name: Open LLM Leaderboard

Llama-3.1-Nemotron-lorablated-70B

An uncensored version of nvidia/Llama-3.1-Nemotron-70B-Instruct-HF created by merging mlabonne/Llama-3-70B-Instruct-abliterated-LORA using task arithmetic.

Method

This model was created using mergekit.

From Ubuntu 24.04 (as root):

apt update
apt install pipx
git clone https://github.com/arcee-ai/mergekit.git
cd mergekit && pipx install -e .

mergekit-yaml config.yaml Llama-3.1-Nemotron-lorablated-70B --allow-crimes --lora-merge-cache=./cache

See @mlabonne's Llama-3.1-70B-Instruct-lorablated for more details on how the LoRA was extracted.

Configuration

The following YAML configuration was used to produce this model:

base_model: nvidia/Llama-3.1-Nemotron-70B-Instruct-HF+mlabonne/Llama-3-70B-Instruct-abliterated-LORA
dtype: bfloat16
merge_method: task_arithmetic
parameters:
  normalize: false
slices:
- sources:
  - layer_range: [0, 80]
    model: nvidia/Llama-3.1-Nemotron-70B-Instruct-HF+mlabonne/Llama-3-70B-Instruct-abliterated-LORA
    parameters:
      weight: 1.0

Acknowlegements

Thanks to @mlabonne, @grimjim, and @failspy for pioneering this technique for uncensoring models.

Compute provided by Hetzner and funded by Schneewolf Labs.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	33.69
IFEval (0-Shot)	71.47
BBH (3-Shot)	48.06
MATH Lvl 5 (4-Shot)	23.34
GPQA (0-shot)	0.89
MuSR (0-shot)	14.92
MMLU-PRO (5-shot)	43.46