metadata

language:
  - en
license: cc-by-nc-4.0
tags:
  - merge
  - lazymergekit
  - dpo
  - rlhf
  - autoquant
  - gptq
dataset:
  - mlabonne/truthy-dpo-v0.1
  - mlabonne/distilabel-intel-orca-dpo-pairs
  - mlabonne/chatml-OpenHermes2.5-dpo-binarized-alpha
base_model:
  - mlabonne/NeuralMonarch-7B
model-index:
  - name: AlphaMonarch-7B
    results:
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: AI2 Reasoning Challenge (25-Shot)
          type: ai2_arc
          config: ARC-Challenge
          split: test
          args:
            num_few_shot: 25
        metrics:
          - type: acc_norm
            value: 73.04
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=mlabonne/AlphaMonarch-7B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: HellaSwag (10-Shot)
          type: hellaswag
          split: validation
          args:
            num_few_shot: 10
        metrics:
          - type: acc_norm
            value: 89.18
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=mlabonne/AlphaMonarch-7B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MMLU (5-Shot)
          type: cais/mmlu
          config: all
          split: test
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 64.4
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=mlabonne/AlphaMonarch-7B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: TruthfulQA (0-shot)
          type: truthful_qa
          config: multiple_choice
          split: validation
          args:
            num_few_shot: 0
        metrics:
          - type: mc2
            value: 77.91
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=mlabonne/AlphaMonarch-7B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: Winogrande (5-shot)
          type: winogrande
          config: winogrande_xl
          split: validation
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 84.69
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=mlabonne/AlphaMonarch-7B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: GSM8k (5-shot)
          type: gsm8k
          config: main
          split: test
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 66.72
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=mlabonne/AlphaMonarch-7B
          name: Open LLM Leaderboard

👑 AlphaMonarch-7B

tl;dr: AlphaMonarch-7B is a new DPO merge that retains all the reasoning abilities of the very best merges and significantly improves its conversational abilities. Kind of the best of both worlds in a 7B model. 🎉

AlphaMonarch-7B is a DPO fine-tuned of mlabonne/NeuralMonarch-7B using the argilla/OpenHermes2.5-dpo-binarized-alpha preference dataset.

It is based on a merge of the following models using LazyMergekit:

Special thanks to Jon Durbin, Intel, Argilla, and Teknium for the preference datasets.

Try the demo: https://huggingface.co/spaces/mlabonne/AlphaMonarch-7B-GGUF-Chat

🔍 Applications

This model uses a context window of 8k. I recommend using it with the Mistral Instruct chat template (works perfectly with LM Studio).

If you use SillyTavern, you might want to tweak the inference parameters. Here's what LM Studio uses as a reference: temp 0.8, top_k 40, top_p 0.95, min_p 0.05, repeat_penalty 1.1.

It is one of the very best 7B models in terms of instructing following and reasoning abilities and can be used for conversations, RP, and storytelling. Note that it tends to have a quite formal and sophisticated style, but it can be changed by modifying the prompt.

⚡ Quantized models

Thanks to LoneStriker for the GPTQ, AWQ, and EXL2 quants.

🏆 Evaluation

Nous

AlphaMonarch-7B is the best-performing 7B model on Nous' benchmark suite (evaluation performed using LLM AutoEval). See the entire leaderboard here.

Model	Average	AGIEval	GPT4All	TruthfulQA	Bigbench
AlphaMonarch-7B 📄	62.74	45.37	77.01	78.39	50.2
NeuralMonarch-7B 📄	62.73	45.31	76.99	78.35	50.28
Monarch-7B 📄	62.68	45.48	77.07	78.04	50.14
teknium/OpenHermes-2.5-Mistral-7B 📄	52.42	42.75	72.99	52.99	40.94
mlabonne/NeuralHermes-2.5-Mistral-7B 📄	53.51	43.67	73.24	55.37	41.76
mlabonne/NeuralBeagle14-7B 📄	60.25	46.06	76.77	70.32	47.86
mlabonne/NeuralOmniBeagle-7B 📄	62.3	45.85	77.26	76.06	50.03
eren23/dpo-binarized-NeuralTrix-7B 📄	62.5	44.57	76.34	79.81	49.27
CultriX/NeuralTrix-7B-dpo 📄	62.5	44.61	76.33	79.8	49.24

EQ-bench

AlphaMonarch-7B is also outperforming 70B and 120B parameter models on EQ-bench by Samuel J. Paech, who kindly ran the evaluations.

MT-Bench

########## First turn ##########
                                    score
model                       turn         
gpt-4                       1     8.95625
OmniBeagle-7B               1     8.31250
AlphaMonarch-7B             1     8.23750
claude-v1                   1     8.15000
NeuralMonarch-7B            1     8.09375
gpt-3.5-turbo               1     8.07500
claude-instant-v1           1     7.80000

########## Second turn ##########
                                     score
model                       turn          
gpt-4                       2     9.025000
claude-instant-v1           2     8.012658
OmniBeagle-7B               2     7.837500
gpt-3.5-turbo               2     7.812500
claude-v1                   2     7.650000
AlphaMonarch-7B             2     7.618750
NeuralMonarch-7B            2     7.375000

########## Average ##########
                                score
model                                
gpt-4                        8.990625
OmniBeagle-7B                8.075000
gpt-3.5-turbo                7.943750
AlphaMonarch-7B              7.928125
claude-instant-v1            7.905660
claude-v1                    7.900000
NeuralMonarch-7B             7.734375
NeuralBeagle14-7B            7.628125

Open LLM Leaderboard

AlphaMonarch-7B is one of the best-performing non-merge 7B models on the Open LLM Leaderboard:

🌳 Model Family Tree

💻 Usage

!pip install -qU transformers accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "mlabonne/AlphaMonarch-7B"
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"])