GLM-130B: An Open Bilingual Pre-trained Model
Aohan Zeng, Xiao Liu, Zhengxiao Du, Zihan Wang, Hanyu Lai, Ming Ding, Zhuoyi Yang, Yifan Xu, Wendi Zheng, Xiao Xia, Weng Lam Tam, Zixuan Ma, Yufei Xue, Jidong Zhai, WenGuang Chen, Peng Zhang, Yuxiao Dong, Jie Tang
Code Available — Be the first to reproduce this paper.
ReproduceCode
- github.com/thudm/glm-130bOfficialIn paperpytorch★ 7,664
- github.com/thudm/chatglm2-6bpytorch★ 15,640
- github.com/thudm/chatglmpytorch★ 13,746
- github.com/thudm/chatglm3pytorch★ 13,746
- github.com/modelscope/modelscopepytorch★ 8,811
- github.com/THUDM/GLMpytorch★ 3,463
- github.com/jackaduma/ChatGLM-LoRA-RLHF-PyTorchpytorch★ 140
- github.com/2023-MindSpore-4/Code12/tree/main/MindFormers/glmmindspore★ 0
- github.com/2023-MindSpore-4/Code12/tree/main/MindFormers/glm3mindspore★ 0
Abstract
We introduce GLM-130B, a bilingual (English and Chinese) pre-trained language model with 130 billion parameters. It is an attempt to open-source a 100B-scale model at least as good as GPT-3 (davinci) and unveil how models of such a scale can be successfully pre-trained. Over the course of this effort, we face numerous unexpected technical and engineering challenges, particularly on loss spikes and divergence. In this paper, we introduce the training process of GLM-130B including its design choices, training strategies for both efficiency and stability, and engineering efforts. The resultant GLM-130B model offers significant outperformance over GPT-3 175B (davinci) on a wide range of popular English benchmarks while the performance advantage is not observed in OPT-175B and BLOOM-176B. It also consistently and significantly outperforms ERNIE TITAN 3.0 260B -- the largest Chinese language model -- across related benchmarks. Finally, we leverage a unique scaling property of GLM-130B to reach INT4 quantization without post training, with almost no performance loss, making it the first among 100B-scale models and more importantly, allowing its effective inference on 4RTX 3090 (24G) or 8RTX 2080 Ti (11G) GPUs, the most affordable GPUs required for using 100B-scale models. The GLM-130B model weights are publicly accessible and its code, training logs, related toolkit, and lessons learned are open-sourced at https://github.com/THUDM/GLM-130B/.
Tasks
Benchmark Results
| Dataset | Model | Metric | Claimed | Verified | Status |
|---|---|---|---|---|---|
| BIG-bench-lite | GLM-130B (3-shot) | Accuracy | 15.11 | — | Unverified |
| BIG-bench-lite | GLM-130B (1-shot) | Accuracy | 14.91 | — | Unverified |
| BIG-bench-lite | GLM-130B (0-shot) | Accuracy | 13.31 | — | Unverified |
| CLUE (AFQMC) | GLM-130B | Accuracy | 71.2 | — | Unverified |
| CLUE (AFQMC) | ERNIE 3.0 Titan-260B | Accuracy | 69 | — | Unverified |
| CLUE (C3) | ERNIE 3.0 Titan-260B | Accuracy | 54.9 | — | Unverified |
| CLUE (C3) | GLM-130B | Accuracy | 77.5 | — | Unverified |
| CLUE (CMNLI) | GLM-130B | Accuracy | 77 | — | Unverified |
| CLUE (CMNLI) | ERNIE 3.0 Titan-260B | Accuracy | 51.7 | — | Unverified |
| CLUE (CMRC2018) | GLM-130B | Accuracy | 55.7 | — | Unverified |
| CLUE (CMRC2018) | ERNIE 3.0 Titan-260B | Accuracy | 16.6 | — | Unverified |
| CLUE (DRCD) | ERNIE 3.0 Titan-260B | Accuracy | 29.5 | — | Unverified |
| CLUE (DRCD) | GLM-130B | Accuracy | 77.1 | — | Unverified |
| CLUE (OCNLI_50K) | ERNIE 3.0 Titan-260B | Accuracy | 44.6 | — | Unverified |
| CLUE (OCNLI_50K) | GLM-130B | Accuracy | 74.7 | — | Unverified |
| CLUE (WSC1.1) | ERNIE 3.0 Titan-260B | Accuracy | 81.1 | — | Unverified |
| CLUE (WSC1.1) | GLM-130B | Accuracy | 83.9 | — | Unverified |
| FewCLUE (BUSTM) | ERNIE 3.0 Titan-260B | Accuracy | 64.4 | — | Unverified |
| FewCLUE (BUSTM) | GLM-130B | Accuracy | 77.5 | — | Unverified |
| FewCLUE (CHID-FC) | ERNIE 3.0 Titan-260B | Accuracy | 87.1 | — | Unverified |
| FewCLUE (CHID-FC) | GLM-130B | Accuracy | 90.1 | — | Unverified |
| FewCLUE (CLUEWSC-FC) | GLM-130B | Accuracy | 77.4 | — | Unverified |
| FewCLUE (CLUEWSC-FC) | ERNIE 3.0 Titan-260B | Accuracy | 53.5 | — | Unverified |
| FewCLUE (EPRSTMT) | ERNIE 3.0 Titan-260B | Accuracy | 88.8 | — | Unverified |
| FewCLUE (EPRSTMT) | GLM-130B | Accuracy | 92.5 | — | Unverified |
| FewCLUE (OCNLI-FC) | ERNIE 3.0 Titan-260B | Accuracy | 53.8 | — | Unverified |
| FewCLUE (OCNLI-FC) | GLM-130B | Accuracy | 73.8 | — | Unverified |
| LAMBADA | GLM-130B (bidirectional attention) | Accuracy | 80.2 | — | Unverified |
| The Pile | GPT-3 | Bits per byte | 0.74 | — | Unverified |
| The Pile | Jurassic-1 | Bits per byte | 0.65 | — | Unverified |
| The Pile | GLM-130B | Bits per byte | 0.63 | — | Unverified |