README.md

metadata

language:
  - en
base_model:
  - unsloth/Llama-3.2-3B-Instruct
license: llama3.2
model-index:
  - name: LLaMa-3.2-Instruct-JankMix-v0.2-SFT-3B
    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: 62.92
            name: strict accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=PJMixers-Dev/LLaMa-3.2-Instruct-JankMix-v0.2-SFT-3B
          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: 23.34
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=PJMixers-Dev/LLaMa-3.2-Instruct-JankMix-v0.2-SFT-3B
          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: 11.33
            name: exact match
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=PJMixers-Dev/LLaMa-3.2-Instruct-JankMix-v0.2-SFT-3B
          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: 3.02
            name: acc_norm
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=PJMixers-Dev/LLaMa-3.2-Instruct-JankMix-v0.2-SFT-3B
          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: 4.87
            name: acc_norm
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=PJMixers-Dev/LLaMa-3.2-Instruct-JankMix-v0.2-SFT-3B
          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: 23.5
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=PJMixers-Dev/LLaMa-3.2-Instruct-JankMix-v0.2-SFT-3B
          name: Open LLM Leaderboard

A much further trained version, this time done with full finetuning instead of DoRA. Similar ~50/50 mix of completion and instruct data.

Note: This likely has refusals like PJMixers-Dev/LLaMa-3.2-Instruct-JankMix-v0.1-SFT-3B since no focus was put on removing refusals. I'm working on a KTO DoRA to solve this, and possibly improve roleplay performance.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	21.50
IFEval (0-Shot)	62.92
BBH (3-Shot)	23.34
MATH Lvl 5 (4-Shot)	11.33
GPQA (0-shot)	3.02
MuSR (0-shot)	4.87
MMLU-PRO (5-shot)	23.50