Daredevil-8B

Daredevil-8B is a mega-merge designed to maximize MMLU. On 27 May 24, it is the Llama 3 8B model with the highest MMLU score. From my experience, a high MMLU score is all you need with Llama 3 models.

It is a merge of the following models using LazyMergekit:

• nbeerbower/llama-3-stella-8B
• Hastagaras/llama-3-8b-okay
• nbeerbower/llama-3-gutenberg-8B
• openchat/openchat-3.6-8b-20240522
• Kukedlc/NeuralLLaMa-3-8b-DT-v0.1
• cstr/llama3-8b-spaetzle-v20
• mlabonne/ChimeraLlama-3-8B-v3
• flammenai/Mahou-1.1-llama3-8B
• KingNish/KingNish-Llama3-8b

Thanks to nbeerbower, Hastagaras, openchat, Kukedlc, cstr, flammenai, and KingNish for their merges. Special thanks to Charles Goddard and Arcee.ai for MergeKit.

🔎 Applications

You can use it as an improved version of meta-llama/Meta-Llama-3-8B-Instruct.

This is a censored model. For an uncensored version, see mlabonne/Daredevil-8B-abliterated.

Tested on LM Studio using the "Llama 3" preset.

⚡ Quantization

• GGUF: https://huggingface.co/mlabonne/Daredevil-8B-GGUF

🏆 Evaluation

Open LLM Leaderboard

Daredevil-8B is the best-performing 8B model on the Open LLM Leaderboard in terms of MMLU score (27 May 24).

Nous

Daredevil-8B is the best-performing 8B model on Nous' benchmark suite (evaluation performed using LLM AutoEval, 27 May 24). See the entire leaderboard here.

Model	Average	AGIEval	GPT4All	TruthfulQA	Bigbench
mlabonne/Daredevil-8B 📄	55.87	44.13	73.52	59.05	46.77
mlabonne/Daredevil-8B-abliterated 📄	55.06	43.29	73.33	57.47	46.17
mlabonne/Llama-3-8B-Instruct-abliterated-dpomix 📄	52.26	41.6	69.95	54.22	43.26
meta-llama/Meta-Llama-3-8B-Instruct 📄	51.34	41.22	69.86	51.65	42.64
failspy/Meta-Llama-3-8B-Instruct-abliterated-v3 📄	51.21	40.23	69.5	52.44	42.69
mlabonne/OrpoLlama-3-8B 📄	48.63	34.17	70.59	52.39	37.36
meta-llama/Meta-Llama-3-8B 📄	45.42	31.1	69.95	43.91	36.7

🌳 Model family tree

🧩 Configuration

models:
  - model: NousResearch/Meta-Llama-3-8B
    # No parameters necessary for base model
  - model: nbeerbower/llama-3-stella-8B
    parameters:
      density: 0.6
      weight: 0.16
  - model: Hastagaras/llama-3-8b-okay
    parameters:
      density: 0.56
      weight: 0.1
  - model: nbeerbower/llama-3-gutenberg-8B
    parameters:
      density: 0.6
      weight: 0.18
  - model: openchat/openchat-3.6-8b-20240522
    parameters:
      density: 0.56
      weight: 0.12
  - model: Kukedlc/NeuralLLaMa-3-8b-DT-v0.1
    parameters:
      density: 0.58
      weight: 0.18
  - model: cstr/llama3-8b-spaetzle-v20
    parameters:
      density: 0.56
      weight: 0.08
  - model: mlabonne/ChimeraLlama-3-8B-v3
    parameters:
      density: 0.56
      weight: 0.08
  - model: flammenai/Mahou-1.1-llama3-8B
    parameters:
      density: 0.55
      weight: 0.05
  - model: KingNish/KingNish-Llama3-8b
    parameters:
      density: 0.55
      weight: 0.05
merge_method: dare_ties
base_model: NousResearch/Meta-Llama-3-8B
dtype: bfloat16

💻 Usage

!pip install -qU transformers accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "mlabonne/Daredevil-8B"
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.bfloat16,
    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"])