README.md

metadata

language:
  - en
license: other
library_name: transformers
tags:
  - mergekit
  - merge
  - lazymergekit
base_model:
  - Qwen/Qwen2.5-32B-Instruct
license_name: tongyi-qianwen
license_link: https://huggingface.co/Qwen/Qwen2-72B-Instruct/blob/main/LICENSE
pipeline_tag: text-generation
model-index:
  - name: BigQwen2.5-52B-Instruct
    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: 79.29
            name: strict accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=mlabonne/BigQwen2.5-52B-Instruct
          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: 59.81
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=mlabonne/BigQwen2.5-52B-Instruct
          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: 17.82
            name: exact match
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=mlabonne/BigQwen2.5-52B-Instruct
          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: 6.94
            name: acc_norm
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=mlabonne/BigQwen2.5-52B-Instruct
          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: 10.45
            name: acc_norm
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=mlabonne/BigQwen2.5-52B-Instruct
          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: 50.22
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=mlabonne/BigQwen2.5-52B-Instruct
          name: Open LLM Leaderboard

BigQwen2.5-52B-Instruct

BigQwen2.5-52B-Instruct is a Qwen/Qwen2-32B-Instruct self-merge made with MergeKit.

It applies the mlabonne/Meta-Llama-3-120B-Instruct recipe.

I made it due to popular demand but I haven't tested it so use it at your own risk. ¯\_(ツ)_/¯

🔍 Applications

It might be good for creative writing tasks. I recommend a context length of 32k but you can go up to 131,072 tokens in theory.

🧩 Configuration

The following YAML configuration was used to produce this model:

slices:
- sources:
  - layer_range: [0, 16]
    model: Qwen/Qwen2.5-32B-Instruct
- sources:
  - layer_range: [8, 24]
    model: Qwen/Qwen2.5-32B-Instruct
- sources:
  - layer_range: [16, 32]
    model: Qwen/Qwen2.5-32B-Instruct
- sources:
  - layer_range: [24, 40]
    model: Qwen/Qwen2.5-32B-Instruct
- sources:
  - layer_range: [32, 48]
    model: Qwen/Qwen2.5-32B-Instruct
- sources:
  - layer_range: [40, 56]
    model: Qwen/Qwen2.5-32B-Instruct
- sources:
  - layer_range: [56, 64]
    model: Qwen/Qwen2.5-32B-Instruct
merge_method: passthrough
dtype: bfloat16

💻 Usage

!pip install -qU transformers accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "mlabonne/BigQwen2.5-52B-Instruct"
messages = [{"role": "user", "content": "What is a large language model?"}]

tokenizer = AutoTokenizer.from_pretrained(model)
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    torch_dtype=torch.float16,
    device_map="auto",
)

outputs = pipeline(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
print(outputs[0]["generated_text"])

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	37.42
IFEval (0-Shot)	79.29
BBH (3-Shot)	59.81
MATH Lvl 5 (4-Shot)	17.82
GPQA (0-shot)	6.94
MuSR (0-shot)	10.45
MMLU-PRO (5-shot)	50.22