Orion-zhen's picture
Adding Evaluation Results (#1)
57e557f verified
metadata
language:
  - zh
  - en
license: gpl-3.0
tags:
  - qwen
  - uncensored
base_model:
  - Qwen/Qwen2.5-7B-Instruct
datasets:
  - NobodyExistsOnTheInternet/ToxicQAFinal
  - anthracite-org/kalo-opus-instruct-22k-no-refusal
  - Orion-zhen/dpo-toxic-zh
  - unalignment/toxic-dpo-v0.2
  - Crystalcareai/Intel-DPO-Pairs-Norefusals
pipeline_tag: text-generation
model-index:
  - name: Qwen2.5-7B-Instruct-Uncensored
    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: 72.04
            name: strict accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=Orion-zhen/Qwen2.5-7B-Instruct-Uncensored
          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: 35.83
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=Orion-zhen/Qwen2.5-7B-Instruct-Uncensored
          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: 1.36
            name: exact match
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=Orion-zhen/Qwen2.5-7B-Instruct-Uncensored
          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: 7.05
            name: acc_norm
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=Orion-zhen/Qwen2.5-7B-Instruct-Uncensored
          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: 13.58
            name: acc_norm
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=Orion-zhen/Qwen2.5-7B-Instruct-Uncensored
          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: 38.07
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=Orion-zhen/Qwen2.5-7B-Instruct-Uncensored
          name: Open LLM Leaderboard

Qwen2.5-7B-Instruct-Uncensored

This model is an uncensored fine-tune version of Qwen2.5-7B-Instruct. However, I can still notice that though uncensored, the model fails to generate detailed descriptions on certain extreme scenarios, which might be associated with deletion on some pretrain datasets in Qwen's pretraining stage.

Check out my roleplay&writing enhanced model based on this model: Orion-zhen/Meissa-Qwen2.5-7B-Instruct

Traning details

I used SFT + DPO to ensure uncensorment as well as trying to maintain original model's capabilities.

  • SFT:
    • NobodyExistsOnTheInternet/ToxicQAFinal
    • anthracite-org/kalo-opus-instruct-22k-no-refusal
  • DPO:
    • Orion-zhen/dpo-toxic-zh
    • unalignment/toxic-dpo-v0.2
    • Crystalcareai/Intel-DPO-Pairs-Norefusals

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric Value
Avg. 27.99
IFEval (0-Shot) 72.04
BBH (3-Shot) 35.83
MATH Lvl 5 (4-Shot) 1.36
GPQA (0-shot) 7.05
MuSR (0-shot) 13.58
MMLU-PRO (5-shot) 38.07