Edit model card

Chuxin-1.6B-Base


介绍 (Introduction)

Chuxin-1.6B-Base是16亿参数规模的模型。Chuxin-1.6B完全基于开源数据构建,在经过超大规模数据训练后,Chuxin-1.6B在各类下游任务上具有非常强的竞争力。

Chuxin-1.6B-1M是基于Chuxin-1.6B-base模型在1M窗口下训练后的结果,大海捞针实验显示其具有非常强的上下文检索能力。

如果您想了解更多关于Chuxin-1.6B开源模型的细节,我们建议您参阅我们的技术报告

Chuxin-1.6B-Base is a model with 1.6 billion parameters. Chuxin-1.6B is built entirely on open-source data. After being trained with large-scale data, Chuxin has very competitive capabilities in various downstream tasks.

Chuxin-1.6B-1M is the result of training the Chuxin-1.6B-base model with a 1M windows. Experiments such as searching for a needle in a haystack demonstrate its strong contextual retrieval abilities.

If you would like to learn more about the Chuxin-1.6B open-source model, we suggest you refer to our technical report.

快速使用(Quickstart)

您可以通过以下代码轻松调用:

You can easily call the model with the following code:

from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("chuxin-llm/Chuxin-1.6B-Base", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("chuxin-llm/Chuxin-1.6B-Base", device_map="auto", trust_remote_code=True, bf16=True).eval()
inputs = tokenizer('蒙古国的首都是乌兰巴托(Ulaanbaatar)\n冰岛的首都是雷克雅未克(Reykjavik)\n埃塞俄比亚的首都是', return_tensors='pt')
inputs = inputs.to(model.device)
pred = model.generate(**inputs, max_new_tokens=15, do_sample=False)
print(tokenizer.decode(pred.cpu()[0], skip_special_tokens=True))
# 蒙古国的首都是乌兰巴托(Ulaanbaatar)\n冰岛的首都是雷克雅未克(Reykjavik)\n埃塞俄比亚的首都是亚的斯亚贝巴(Addis Ababa)...

评测效果(Evaluation)

(常识推理和阅读理解) Common Sense Reasoning and Reading Comprehension tasks

Model size ARC-c ARC-e Boolq Copa Hellaswag OpenbookQA Piqa Sciq Winogrande Avg
Gemma 2B 48.98 78.45 69.51 84 71.73 39.8 78.02 94.3 65.51 70.03
H2O-Danube† 1.8B 35.84 62.29 65.81 - 68.20 37.6 76.93 - 61.96 -
Qwen1.5 1.8B 37.03 67.51 66.64 78 61.60 34.40 73.99 93 61.56 63.74
StableLM 2 1.6B 43.52 69.44 75.5 84 70.3 39.6 76.82 96.1 64.17 68.82
OpenLlama† 3B 34 69 68 - 49 40 75 - 62 -
CT-LLM 2B 34.81 65.49 62.45 74 54.77 33.4 71.38 90.6 57.85 60.63
TinyLLama 1.1B 34.81 67.47 63.15 74 60 34.6 73.12 88.8 58.88 61.64
OLMo 1B 34.22 67.55 61.4 82 63.96 36.4 75.1 86.7 60.3 63.07
Chuxin-1.6B-Base 1.6B 39.68 71.38 71.25 83 66.09 35.00 77.09 95 63.54 66.89

带有†的模型表示我们直接报告了相应论文中的分数,其他的则来自于我们重新测试的结果。

Models with † denote that we directly report the scores from the corresponding paper, and others are from our implementation.

Open LLM LeaderBoard

Model size ARC-c HellaSwag MMLU TruthfulQA Winogrande GSM-8k Avg Avg wo GSM
Gemma 2B 48.98 71.73 42.47 33 65.51 10.08 45.3 52.34
H2O-Danube 1.8B 39. 68 69.75 25.97 33.63 64.17 2.05 39.21 46.64
Qwen1.5† 1.8B 37.88 61.42 46.71 39.43 60.3 33.59 46.55 49.15
StableLM 2 1.6B 43.52 70.3 39.8 36.61 64.17 17.29 45.28 50.88
OpenLlama† 3B 39.9 71.6 27.1 34.8 67 0.9 40.3 48.08
CT-LLM 2B 34.81 54.77 37.81 39.81 57.85 7.35 38.73 45.01
TinyLLama 1.1B 33.87 60.31 26.04 37.32 59.51 1.44 36.42 43.41
OLMo 1B 34.22 63.96 35.44 35.53 62.67 9.86 41.81 48.2
Chuxin-1.6B-Base 1.6B 39.68 66.09 41.07 37.65 63.54 12.66 43.45 49.61

带有†的模型表示我们直接报告 Open LLM排行榜的分数,其他的则来自于我们重新测试的结果。

Models with † denote that we directly report the scores from the Open LLM Leaderboard, and others are from our implementation.

CMMLU, C-Eval and HumanEval

Model size C-Eval CMMLU HUMANEVAL
Gemma 2B 31 31.06 9.51
Qwen1.5 1.8B 59.38 57.08 23.17
StableLM 2 1.6B 29.27 30.1 7.32
CT-LLM 2B 36.78 36.4 9.15
Chuxin-1.6B-Base 1.6B 39.31 37.11 9.76

引用 (Citation)

如果你觉得我们的工作对你有帮助,欢迎引用!

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

@article{chuxin,
  title={CHUXIN: 1.6B TECHNICAL REPORT},
  author={Xiaomin Zhuang, Yufan Jiang, Qiaozhi He, Zhihua Wu},
  journal={arXiv preprint arXiv:2405.04828},
  year={2024}
}

Downloads last month
15
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Model tree for chuxin-llm/Chuxin-1.6B-Base

Quantizations
2 models