File size: 6,387 Bytes
c57e142 ac80538 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 |
---
license: other
language:
- zh
- en
---
# BlueLM
<p align="center">
🖥 <a href="https://github.com/vivo-ai-lab/BlueLM" target="_blank">github</a> • 📜 <a href="https://huggingface.co/vivo-ai/BlueLM-7B-Chat/blob/main/MODEL_LICENSE" target="_blank">LICENSE</a> • 🎯 <a href="https://developers.vivo.com/product/ai/bluelm" target="_blank">vivo Developers</a> • 🗨 <a href="https://github.com/vivo-ai-lab/BlueLM/blob/main/resources/wechat.png" target="_blank">WeChat</a>
</p>
## 模型介绍/Introduction
BlueLM 是由 vivo AI 全球研究院自主研发的大规模预训练语言模型，本次发布包含 7B 基础模型和 7B 对话模型，同时我们开源了支持 **32K** 的长文本基础模型和对话模型。
- **更大量的优质数据**：高质量语料库进行训练，规模达到了 **2.6 万亿** 的 token 数，该语料库包含中文、英文以及少量日韩数据。
- **更优的效果**：其中 BlueLM-7B-Chat 在 **C-Eval** 和 **CMMLU** 上均取得领先结果，对比同尺寸开源模型中具有较强的竞争力。
- **长文本支持**：BlueLM-7B-Base-32K 和 BlueLM-7B-Chat-32K 均支持 **32K** 长文本，在保持基础能力相当情况下，能够支持更长上下文理解。
- **协议说明**：BlueLM 系列欢迎开发者进行学术研究和商业应用。
BlueLM is a large-scale open-source language model independently developed by the vivo AI Lab. This release includes 2K and 32K context length versions for both Base and Chat models.
- **High-quality Data**: BlueLM is trained on a high-quality data with 2.6 trillion tokens. Our train corpus mainly consists of Chinese and English data, with a small amount of Japanese and Korean data.
- **Stronger Performance**: BlueLM-7B-Chat achieves a strong competitive performance in C-Eval and CMMLU benchmarks of the same size.
- **Longer Context**: We have extended the context length of both BlueLM-7B-Base-32K and BlueLM-7B-Chat-32K models from 2K to 32K. The models can support longer context understanding while maintaining the same basic capabilities.
- **Model License**: BlueLM weights are open for academic research and commercial use.
本次发布基座模型下载链接见：
The release versions and hugging face download links are listed in the table below:
| | Base Model | Chat Model | 4bits Quantized Chat Model |
|:---:|:--------------------:|:--------------------:|:--------------------------:|
| 7B-2k | [BlueLM-7B-Base](https://huggingface.co/vivo-ai/BlueLM-7B-Base) | [BlueLM-7B-Chat](https://huggingface.co/vivo-ai/BlueLM-7B-Chat) | [BlueLM-7B-Chat-4bits](https://huggingface.co/vivo-ai/BlueLM-7B-Chat-4bits) |
| 7B-32K | [BlueLM-7B-Base-32K](https://huggingface.co/vivo-ai/BlueLM-7B-Base-32K) | [BlueLM-7B-Chat-32K](https://huggingface.co/vivo-ai/BlueLM-7B-Chat-32K) | - |
## 评测结果/Benchmark Results
为了保证模型评测的一致性，我们采用 [OpenCompass](https://opencompass.org.cn/leaderboard-llm) 进行相关榜单的评测。我们分别在 C-Eval、MMLU、CMMLU、GaoKao、AGIEval、BBH、GSM8K、MATH 和 HumanEval 榜单对 BlueLM 的通用能力、数学能力和代码能力进行了测试。
To ensure the consistency of model evaluation, we use [OpenCompass](https://opencompass.org.cn/leaderboard-llm) to evaluate the performance on relevant leaderboards. We conducted extensive tests on C-Eval, MMLU, CMMLU, GaoKao, AGIEval, BBH, GSM8K, MATH and HumanEval datasets across general ability, mathematical ability and coding ability.
| Model | **C-Eval** | **MMLU** | **CMMLU** | **Gaokao** | **AGIEval** | **BBH** | **GSM8K** | **MATH** | **HumanEval** |
|:------------------|:-----------|:---------|:----------|:-----------|:------------|:--------|:----------|:---------|:--------------|
| | 5-shot | 5-shot | 5-shot | 0-shot | 0-shot | 3-shot | 4-shot | 5-shot | 0-shot |
| GPT-4 | 69.9 | 86.4 | 71.2 | 72.3 | 55.1 | 86.7 | 91.4 | 45.8 | 74.4 |
| ChatGPT | 52.5 | 70.0 | 53.9 | 51.1 | 39.9 | 70.1 | 78.2 | 28 | 73.2 |
| LLaMA2-7B | 32.5 | 45.3 | 31.8 | 18.9 | 21.8 | 38.2 | 16.7 | 3.3 | 12.8 |
| ChatGLM2-6B(Base) | 51.7 | 47.9 | 50.0 | - | - | 33.7 | 32.4 | 6.5 | - |
| Baichuan2-7B | 56.3 | 54.7 | 57.0 | 34.8 | 34.6 | 41.8 | 24.6 | 5.4 | 17.7 |
| BlueLM-7B-Base | 67.5 | 55.2 | 66.6 | 58.9 | 43.4 | 41.7 | 27.2 | 6.2 | 18.3 |
| BlueLM-7B-Chat | 72.7 | 50.7 | 74.2 | 48.7 | 43.4 | 65.6 | 51.9 | 13.4 | 21.3 |
## 推理部署/Inference and Deployment
```python
>>> import torch
>>> from transformers import AutoModelForCausalLM, AutoTokenizer
>>> tokenizer = AutoTokenizer.from_pretrained("vivo-ai/BlueLM-7B-Chat", trust_remote_code=True, use_fast=False)
>>> model = AutoModelForCausalLM.from_pretrained("vivo-ai/BlueLM-7B-Chat", device_map="cuda:0", torch_dtype=torch.bfloat16, trust_remote_code=True)
>>> model = model.eval()
>>> inputs = tokenizer("[|Human|]:三国演义的作者是谁？[|AI|]:", return_tensors="pt")
>>> inputs = inputs.to("cuda:0")
>>> pred = model.generate(**inputs, max_new_tokens=64, repetition_penalty=1.1)
>>> print(tokenizer.decode(pred.cpu()[0], skip_special_tokens=True))
三国演义的作者是谁？ 《三国演义》是元末明初小说家罗贯中创作的长篇小说。
```
更多使用说明，请参考我们的 [Github 仓库](https://github.com/vivo-ai-lab/BlueLM)。
For more instructions, please refer to our [Github Repo](https://github.com/vivo-ai-lab/BlueLM).
## 协议/License
社区使用代码依照 [Apache-2.0](https://www.apache.org/licenses/LICENSE-2.0) 协议开源，且使用 BlueLM 模型权重需要遵循 [vivo_BlueLM模型许可协议](https://huggingface.co/vivo-ai/BlueLM-7B-Chat/blob/main/MODEL_LICENSE)。
Our code is licensed under the [Apache-2.0](https://www.apache.org/licenses/LICENSE-2.0) and [Community License for BlueLM Model](https://huggingface.co/vivo-ai/BlueLM-7B-Chat/blob/main/MODEL_LICENSE). |