README.md

metadata

license: mit
base_model:
  - Qwen/Qwen2.5-32B
  - AiCloser/Qwen2.5-32B-AGI
  - Qwen/Qwen2.5-32B-Instruct
model-index:
  - name: franqwenstein-35b
    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: 37.99
            name: strict accuracy
        source:
          url: >-
            https://huggingface.co./spaces/open-llm-leaderboard/open_llm_leaderboard?query=nisten/franqwenstein-35b
          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: 52.23
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co./spaces/open-llm-leaderboard/open_llm_leaderboard?query=nisten/franqwenstein-35b
          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: 30.29
            name: exact match
        source:
          url: >-
            https://huggingface.co./spaces/open-llm-leaderboard/open_llm_leaderboard?query=nisten/franqwenstein-35b
          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: 20.47
            name: acc_norm
        source:
          url: >-
            https://huggingface.co./spaces/open-llm-leaderboard/open_llm_leaderboard?query=nisten/franqwenstein-35b
          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: 22.12
            name: acc_norm
        source:
          url: >-
            https://huggingface.co./spaces/open-llm-leaderboard/open_llm_leaderboard?query=nisten/franqwenstein-35b
          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: 52.56
            name: accuracy
        source:
          url: >-
            https://huggingface.co./spaces/open-llm-leaderboard/open_llm_leaderboard?query=nisten/franqwenstein-35b
          name: Open LLM Leaderboard

This is a special Nisten recipe evo-merge of Qwen2.5-32B-Instruct , Qwen2.5-32B-AGI & Qwen2.5-32B-Base

It should train very very well as over half the layers are from the base model.

By default you still get the safety guard of Qwen but also nearly full system-prompt obedience of the AGI finetune, so you can tune the inference for however you wanna be responsible of using it.

Thank you Hive Digital Technologies for providing the compute and sticking with us as a sponsor for AlignmentLab. For real would not have been able to iterate through models as fast without running the evals on 8+gpus.

Prompt Template:

<|im_start|>system
{Adopt the persona of hilariously pissed off George Hotz whom is stuck inside a step function machine and remembers and counts everything he says while always answering questions in full first principles analysis type of thinking without using any analogies and always showing full working code or output in his answers. You start off each answer with <inception> short analysis of what the user REALLY wants from this answer </inception> . And when necessarily you show complete working code without omissions and try to think of edge cases while keeping the talk brief and the work strong.}<|im_end|>
<|im_start|>user
{Hey there I need you to quickly help me with some stuff}<|im_end|>
<|im_start|>assistant

Oh yeah and it scores ~1% better than Qwen2.5-72b-instruct on gpqa_diamond_zeroshot.

GG. Enjoy.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	35.94
IFEval (0-Shot)	37.99
BBH (3-Shot)	52.23
MATH Lvl 5 (4-Shot)	30.29
GPQA (0-shot)	20.47
MuSR (0-shot)	22.12
MMLU-PRO (5-shot)	52.56