PetroGPT

GitHub Repository: PetroGPT & WestSeverus-7B-DPO-v2

PetroGPT

focusing on pretrain and finetune domain specific LLMs in Gas/Oil Refinement/Chemistry domains including MoE (Mixture of Expert) model -- WestSeverus-7B-DPO-v2 is a powerful multi-domain language model, combining Mistral-7B, FerdanoGPT (math-focused), and WestLake for scientific, mathematical, and programming tasks. Ranked #1 on Open LLM leaderboard, available on Hugging Face.

Description:

WestSeverus-7B-DPO-v2 is an advanced multi-domain language model developed by combining Mistral-7B, FerdanoGPT (math-focused), and WestLake (context and reasoning-enhanced). The model excels in scientific, mathematical, and programming tasks, leveraging the latest advancements in model fusion techniques such as DARE (Drop And Rescale). It achieved a top ranking on the Open LLM leaderboard for one week and is publicly available for use and research on Hugging Face.

• Key Capabilities:

  • Scientific knowledge
  • Mathematical reasoning
  • Programming (coding tasks)

• Fusion Method: DARE (Drop And Rescale) to effectively combine the strengths of multiple specialized models.

• Hugging Face Repo: PetroGPT/WestSeverus-7B-DPO-v2

• Reference: WestSeverus model is cited in the paper "Language Models are Super Mario: Absorbing Abilities from Homologous Models as a Free Lunch" (arXiv: 2311.03099). See the reference here: arXiv:2311.03099

---

WestSeverus-7B-DPO-v2

WestSeverus-7B-DPO-v2 is a state-of-the-art large language model that integrates multiple specialized models to achieve exceptional performance in scientific knowledge, mathematical reasoning, and programming tasks. It is built upon Mistral-7B, enhanced with mathematical reasoning capabilities from FerdanoGPT, and further improved with scientific and programming knowledge from WestLake. The model utilizes the DARE (Drop And Rescale) technique for model fusion, achieving outstanding results on multi-task benchmarks.

Features

• Scientific Knowledge: The model excels in scientific domains including physics, chemistry, and general knowledge.
• Mathematical Reasoning: Strong performance on tasks requiring advanced mathematical reasoning, such as GSM8K and MATH.
• Programming: Efficient at solving programming challenges from datasets like HumanEval and MBPP.

Model Training and Fusion

• Base Model: Mistral-7B
• Enhancements:
  • FerdanoGPT (for math capabilities)
  • WestLake (for scientific and programming expertise)
• Fusion Technique: DARE (Drop And Rescale) is used to merge the models effectively, dropping and rescaling parameters to maintain performance across tasks.

How to Use

You can easily load and use the model with Hugging Face's transformers library:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "PetroGPT/WestSeverus-7B-DPO-v2"
model = AutoModelForCausalLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

input_text = "What is the integral of x^2?"
inputs = tokenizer(input_text, return_tensors="pt")
outputs = model.generate(inputs['input_ids'])

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

License

This repository is licensed under the MIT License.

References

• WestSeverus model is cited in the paper "Language Models are Super Mario: Absorbing Abilities from Homologous Models as a Free Lunch". You can access the paper on arXiv.
• The model is publicly available on Hugging Face at PetroGPT/WestSeverus-7B-DPO-v2.