Edit model card

Qwen2-72B

Branch Names

bpw = Bits per Weight hb = Bits for the lm_head layer

Quantization Details

Quantized via Exllamav2 Version: 0.1.6 All versions in this repo were quantized with the setting Rope Scale=4

Original Model Card

Introduction

Qwen2 is the new series of Qwen large language models. For Qwen2, we release a number of base language models and instruction-tuned language models ranging from 0.5 to 72 billion parameters, including a Mixture-of-Experts model. This repo contains the 72B Qwen2 base language model.

Compared with the state-of-the-art opensource language models, including the previous released Qwen1.5, Qwen2 has generally surpassed most opensource models and demonstrated competitiveness against proprietary models across a series of benchmarks targeting for language understanding, language generation, multilingual capability, coding, mathematics, reasoning, etc.

For more details, please refer to our blog, GitHub, and Documentation.

Model Details

Qwen2 is a language model series including decoder language models of different model sizes. For each size, we release the base language model and the aligned chat model. It is based on the Transformer architecture with SwiGLU activation, attention QKV bias, group query attention, etc. Additionally, we have an improved tokenizer adaptive to multiple natural languages and codes.

Requirements

The code of Qwen2 has been in the latest Hugging face transformers and we advise you to install transformers>=4.37.0, or you might encounter the following error:

KeyError: 'qwen2'

Usage

We do not advise you to use base language models for text generation. Instead, you can apply post-training, e.g., SFT, RLHF, continued pretraining, etc., on this model.

Performance

The evaluation of base models mainly focuses on the model performance of natural language understanding, general question answering, coding, mathematics, scientific knowledge, reasoning, multilingual capability, etc.

The datasets for evaluation include:

English Tasks: MMLU (5-shot), MMLU-Pro (5-shot), GPQA (5shot), Theorem QA (5-shot), BBH (3-shot), HellaSwag (10-shot), Winogrande (5-shot), TruthfulQA (0-shot), ARC-C (25-shot)

Coding Tasks: EvalPlus (0-shot) (HumanEval, MBPP, HumanEval+, MBPP+), MultiPL-E (0-shot) (Python, C++, JAVA, PHP, TypeScript, C#, Bash, JavaScript)

Math Tasks: GSM8K (4-shot), MATH (4-shot)

Chinese Tasks: C-Eval (5-shot), CMMLU (5-shot)

Multilingual Tasks: Multi-Exam (M3Exam 5-shot, IndoMMLU 3-shot, ruMMLU 5-shot, mMMLU 5-shot), Multi-Understanding (BELEBELE 5-shot, XCOPA 5-shot, XWinograd 5-shot, XStoryCloze 0-shot, PAWS-X 5-shot), Multi-Mathematics (MGSM 8-shot), Multi-Translation (Flores-101 5-shot)

Qwen2-72B performance

Datasets DeepSeek-V2 Mixtral-8x22B Llama-3-70B Qwen1.5-72B Qwen1.5-110B Qwen2-72B
Architecture MoE MoE Dense Dense Dense Dense
#Activated Params 21B 39B 70B 72B 110B 72B
#Params 236B 140B 70B 72B 110B 72B
English
MMLU 78.5 77.8 79.5 77.5 80.4 84.2
MMLU-Pro - 49.5 52.8 45.8 49.4 55.6
GPQA - 34.3 36.3 36.3 35.9 37.9
Theorem QA - 35.9 32.3 29.3 34.9 43.1
BBH 78.9 78.9 81.0 65.5 74.8 82.4
HellaSwag 87.8 88.7 88.0 86.0 87.5 87.6
WindoGrande 84.8 85.0 85.3 83.0 83.5 85.1
ARC-C 70.0 70.7 68.8 65.9 69.6 68.9
TruthfulQA 42.2 51.0 45.6 59.6 49.6 54.8
Coding
HumanEval 45.7 46.3 48.2 46.3 54.3 64.6
MBPP 73.9 71.7 70.4 66.9 70.9 76.9
EvalPlus 55.0 54.1 54.8 52.9 57.7 65.4
MultiPL-E 44.4 46.7 46.3 41.8 52.7 59.6
Mathematics
GSM8K 79.2 83.7 83.0 79.5 85.4 89.5
MATH 43.6 41.7 42.5 34.1 49.6 51.1
Chinese
C-Eval 81.7 54.6 65.2 84.1 89.1 91.0
CMMLU 84.0 53.4 67.2 83.5 88.3 90.1
Multilingual
Mulit-Exam 67.5 63.5 70.0 66.4 75.6 76.6
Multi-Understanding 77.0 77.7 79.9 78.2 78.2 80.7
Multi-Mathematics 58.8 62.9 67.1 61.7 64.4 76.0
Multi-Translation 36.0 23.3 38.0 35.6 36.2 37.8

Citation

If you find our work helpful, feel free to give us a cite.

@article{qwen2,
  title={Qwen2 Technical Report},
  year={2024}
}
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Examples
Unable to determine this model's library. Check the docs .