Title: Introducing Omningotex-7b: The World's Most Accurate 7B LLM
Today, I'm excited to share the creation of a groundbreaking language model, "liminerity/Omningotex-7b-slerp." This model has achieved an impressive accuracy rate of 76.33%, making it the most accurate 7B LLM in the world. The journey to create Omningotex-7b-slerp began with an experimental process called "merging." I started with a model named "ingot-7b-slerp," which was created by merging two other LLMs, "blurred-beagle-7b-slerp" (by myself, liminerity) and "Macaroni-7b-Tied" (by andrijdavid), a total of eight times over. After the successful creation of ingot-7b-slerp, I proceeded to merge it with another model, "dpo-binarized-NeuralTrix-7B" by eren23, using gradient slerp. The resulting model, "binarized-ingotrix-slerp-7b," achieved an accuracy rate of 76.04%. To further enhance the model's performance, I decided to merge "binarized-ingotrix-slerp-7b" with "dpo-binarized-NeutrixOmnibe-7B" by eren23 once again. The resulting model, "Omningotex-7b," is now the most accurate 7B LLM available. This breakthrough in LLM accuracy was achieved through a combination of careful experimentation and a deep understanding of the underlying algorithms and techniques. I believe that Omningotex-7b-slerp's success demonstrates the potential for further advancements in the field of natural language processing and artificial intelligence. I look forward to sharing more updates and insights as I continue to explore the possibilities of LLMs and push the boundaries of what is possible in the world of AI. Stay tuned for more exciting developments in the future!
A huge thank you to Maxime Labonne and his creation of LazyMergeKit colab project. Use of it helped me gain a further grasp of the concepts at play and led to the creation of this model. I'm sure it won't be number 1 for long which excited me even more!
Next, I set out to learn how to fine-tune with the resources I have available. My next overall goal is to try and find a way to produce a smaller model with high accuracy either through merging down using fewer layers after each merge. I may need to include finetuning between each merge or merging larger more accurate models into a smaller base while maintaining accuracy and performance. Every version of "TinyMistral" I come by seems to be bricked in the sense it spits out nonsense. Thank you for your time If you read this all the way.
Omningotex-7B-slerp
Omningotex-7B-slerp is a merge of the following models using LazyMergekit:
🧩 Configuration
slices:
- sources:
- model: liminerity/binarized-ingotrix-slerp-7b
layer_range: [0, 32]
- model: eren23/dpo-binarized-NeutrixOmnibe-7B
layer_range: [0, 32]
merge_method: slerp
base_model: liminerity/binarized-ingotrix-slerp-7b
parameters:
t:
- filter: self_attn
value: [0, 0.5, 0.3, 0.7, 1]
- filter: mlp
value: [1, 0.5, 0.7, 0.3, 0]
- value: 0.5
dtype: bfloat16
💻 Usage
!pip install -qU transformers accelerate
from transformers import AutoTokenizer
import transformers
import torch
model = "liminerity/Omningotex-7b-slerp"
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. | 76.33 |
AI2 Reasoning Challenge (25-Shot) | 73.29 |
HellaSwag (10-Shot) | 88.96 |
MMLU (5-Shot) | 64.69 |
TruthfulQA (0-shot) | 76.32 |
Winogrande (5-shot) | 84.21 |
GSM8k (5-shot) | 70.51 |
- Downloads last month
- 74
Model tree for liminerity/Omningotex-7b-slerp
Collection including liminerity/Omningotex-7b-slerp
Evaluation results
- normalized accuracy on AI2 Reasoning Challenge (25-Shot)test set Open LLM Leaderboard73.290
- normalized accuracy on HellaSwag (10-Shot)validation set Open LLM Leaderboard88.960
- accuracy on MMLU (5-Shot)test set Open LLM Leaderboard64.690
- mc2 on TruthfulQA (0-shot)validation set Open LLM Leaderboard76.320
- accuracy on Winogrande (5-shot)validation set Open LLM Leaderboard84.210
- accuracy on GSM8k (5-shot)test set Open LLM Leaderboard70.510