merge

This is a merge of pre-trained language models created using mergekit.

Code credit: this excellent medium blog

Merge Details

Merge Method

This model was merged using the DARE TIES merge method using CultriX/NeuralTrix-7B-dpo as a base.

Models Merged

The following models were included in the merge:

mlabonne/NeuralBeagle14-7B
HuggingFaceH4/zephyr-7b-alpha

Benchmarks

Open LLM Leaderboard

Model	Average	ARC	HellaSwag	MMLU	TruthfulQA	Winogrande	GSM8K
mayacinka/NeuralZephyr-Beagle-7B	71.57	68.6	86.38	64.67	65.17	81.14	63.46

Configuration

The following YAML configuration was used to produce this model:

models:
  - model: CultriX/NeuralTrix-7B-dpo
  - model: HuggingFaceH4/zephyr-7b-alpha
    parameters:
      density: 0.83
      weight: 0.4
  - model: mlabonne/NeuralBeagle14-7B
    parameters: 
      density: 0.83
      weight: 0.6
merge_method: dare_ties
base_model: CultriX/NeuralTrix-7B-dpo
parameters:
  int8_mask: true
dtype: bfloat16

Inference

# pip install transformers

from transformers import AutoTokenizer
import transformers
import torch

model = "mayacinka/NeuralZephyr-Beagle-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"])

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	71.57
AI2 Reasoning Challenge (25-Shot)	68.60
HellaSwag (10-Shot)	86.38
MMLU (5-Shot)	64.67
TruthfulQA (0-shot)	65.17
Winogrande (5-shot)	81.14
GSM8k (5-shot)	63.46

mayacinka
/

NeuralZephyr-Beagle-7B