README.md

metadata

license: apache-2.0
datasets:
  - openbmb/UltraInteract_sft
  - openbmb/UltraInteract_pair
  - openbmb/UltraFeedback
tags:
  - reasoning
  - preference_learning
  - nca
pipeline_tag: text-generation

Eurus: A suite of open-source LLMs optimized for reasoning

Introduction • Evaluation

Links

• 📜 Paper
• 🤗 Eurus Collection
• 🤗 UltraInteract
  • SFT
  • Preference Learning
• GitHub Repo

Introduction

Eurux-8x22B-KTO is SFT and KTO fine-tuned from Mixtral-8x22B on all multi-turn trajectory pairs in UltraInteract and all pairs in UltraFeedback.

It achieves superb reasoning performance as well as exellent chat & instruction-following capabilities.

Evaluation

We conducted overall coding, math, reasoning, knowledge, instruction-following and chat benchmarking. Results are shown below, with the best scores in open-source models bolded:

Models/Benchmarks	Coding			Math			Reasoning	Knowledge	Ins-Following	Chat
	HumanEval	MBPP	LeetCode	GSMPLUS	MATH	TheoremQA	BBH (CoT)	MMLU	IFEval	MT-Bench
GPT-3.5-Turbo	76.8	82.5	23.3	61.2	37.8	35.6	70.1	70.0	56.6	7.94
GPT-4	85.4	83.5	41.8	85.6	69.7	52.4	86.7	86.4	79.7	8.96
Mixtral-8x7B-Ins	50.6	50.1	5.6	49.6	25.9	20.4	73.5	70.3	48.8	8.30
DS-LM-67B-Chat	70.7	65.7	20.0	65.0	41.0	17.9	78.9	72.3	52.7	8.35
QWen-1.5-72B	71.3	56.9	15.6	65.4	43.4	18.5	78.0	72.9	53.4	8.61
Eurus-70b-NCA	79.3	71.9	33.3	62.8	41.7	32.6	80.0	59.4	49.2	7.54
Eurux-8x22b-KTO	71.3	68.9	29.4	68.3	48.4	35.3	83.6	75.9	67.1	8.58
Eurux-8x22b-NCA	75.0	69.7	35.0	68.1	49.0	35.5	83.5	75.6	67.1	8.46

Usage

# pip install 'transformers>=4.39.3'
# pip install accelerate

import torch
from transformers import pipeline

pipe = pipeline(
    "text-generation",
    model="openbmb/Eurux-8x22b-kto",
    device_map="auto",
    torch_dtype=torch.bfloat16,
)
messages = [
    {"role": "user", "content": "What does Eurus mean?"},
]
outputs = pipe(
    messages,
    max_new_tokens=512,
    do_sample=True,
    temperature=0.7,
    top_k=50,
    top_p=0.95,
)
print(outputs[0]["generated_text"][-1]["content"])

We apply tailored prompts for coding and math, consistent with UltraInteract data formats:

Coding

[INST] Write Python code to solve the task:
{Instruction} [/INST]

Math-CoT

[INST] Solve the following math problem step-by-step.
Simplify your answer as much as possible. Present your final answer as \\boxed{Your Answer}.
{Instruction} [/INST]

Math-PoT

[INST] Tool available:
[1] Python interpreter
When you send a message containing Python code to python, it will be executed in a stateful Jupyter notebook environment.
Solve the following math problem step-by-step.
Simplify your answer as much as possible.
{Instruction} [/INST]

Citation

@misc{yuan2024advancing,
      title={Advancing LLM Reasoning Generalists with Preference Trees}, 
      author={Lifan Yuan and Ganqu Cui and Hanbin Wang and Ning Ding and Xingyao Wang and Jia Deng and Boji Shan and Huimin Chen and Ruobing Xie and Yankai Lin and Zhenghao Liu and Bowen Zhou and Hao Peng and Zhiyuan Liu and Maosong Sun},
      year={2024},
      eprint={2404.02078},
      archivePrefix={arXiv},
      primaryClass={cs.AI}
}