StarDust-12b-v2 / README.md
leaderboard-pr-bot's picture
Adding Evaluation Results
45a3562 verified
|
raw
history blame
6.21 kB
metadata
language:
  - en
license: apache-2.0
tags:
  - chat
  - mistral
  - roleplay
  - creative-writing
base_model:
  - nbeerbower/mistral-nemo-bophades-12B
  - anthracite-org/magnum-v2-12b
  - Sao10K/MN-12B-Lyra-v3
  - Gryphe/Pantheon-RP-1.6-12b-Nemo
pipeline_tag: text-generation
model-index:
  - name: StarDust-12b-v2
    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: 56.29
            name: strict accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=Luni/StarDust-12b-v2
          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: 34.95
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=Luni/StarDust-12b-v2
          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: 5.97
            name: exact match
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=Luni/StarDust-12b-v2
          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: 5.82
            name: acc_norm
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=Luni/StarDust-12b-v2
          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: 14.26
            name: acc_norm
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=Luni/StarDust-12b-v2
          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: 27.1
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=Luni/StarDust-12b-v2
          name: Open LLM Leaderboard

image/png

image/png

StarDust-12b-v2

Quants

Description | Usecase

  • The result of this merge is in my opinion a more vibrant and less generic sonnet inspired prose, it's able to be gentle and harsh where asked.
  • The v2 uses the non-kto magnum which tends to have less "claudeism" (making the story feel rather repetitive)
  • Note on Non-Kto: There is a very big gap between people preferring and disliking the KTO. To make things easier, you can still use Luni/StarDust-12b-v1 which has the KTO version.
  • In early testing users have reported a much better experience in longer roleplays and a abillity to add a creative touch to the stable experiencve.

Just like with v1:

  • This model is intended to be used as a Role-playing model.
  • Its direct conversational output is... I can't even say it's luck, it's just not made for it.
  • Extension to Conversational output: The Model is designed for roleplay, direct instructing or general purpose is NOT recommended.

Initial Feedback

  • Initial feedback has proven the model to be a solid "go-to" choice for creative storywriting
  • The prose has been certified as "amazing" with many making it their default model.

Prompting

ChatML has proven to be the BEST choice.

Both Mistral and ChatML should work though I had better results with ChatML: ChatML Example:

"""<|im_start|>user
Hi there!<|im_end|>
<|im_start|>assistant
Nice to meet you!<|im_end|>
<|im_start|>user
Can I ask a question?<|im_end|>
<|im_start|>assistant
"""

Merge Details

Merge Method

This model was merged using the DARE TIES merge method using Sao10K/MN-12B-Lyra-v3 as a base.

Models Merged

The following models were included in the merge:

Special Thanks

Special thanks to the SillyTilly and myself for helping me find the energy to finish this.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric Value
Avg. 24.06
IFEval (0-Shot) 56.29
BBH (3-Shot) 34.95
MATH Lvl 5 (4-Shot) 5.97
GPQA (0-shot) 5.82
MuSR (0-shot) 14.26
MMLU-PRO (5-shot) 27.10