模型:

DeepPavlov/distilrubert-small-cased-conversational

英文

distilrubert-small-cased-conversational

Conversational DistilRuBERT-smalldistilrubert-small-cased-conversational(俄语,有大小写,2层,768隐藏层,12个注意头,107M个参数)在OpenSubtitles[1]、 Dirty Pikabu 和Taiga语料库的社交媒体部分[2](作为 Conversational RuBERT )上进行了训练。它可以被视为 Conversational DistilRuBERT-base 的小版本。

我们的DistilRuBERT-small深受[3]、[4]的启发。具体而言,我们使用了以下内容:

  • KL损失(老师和学生输出logits之间的损失)
  • MLM损失(tokens标签和学生输出logits之间的损失)
  • 余弦嵌入损失(老师编码器的六个连续隐藏状态的平均值与学生的一个隐藏状态之间的损失)
  • MSE损失(老师编码器的六个连续注意力图的平均值与学生的一个注意力图之间的损失)

该模型经过了大约80小时的训练,使用了8个nVIDIA Tesla P100-SXM2.0 16Gb。

为了评估推理速度的改进,我们对随机序列上的老师和学生模型进行了测试,序列长度为512,batch_size为16(用于吞吐量)和batch_size为1(用于延迟)。所有测试都在Intel(R) Xeon(R) CPU E5-2698 v4 @ 2.20GHz和nVIDIA Tesla P100-SXM2.0 16Gb上执行。

Model Size, Mb. CPU latency, sec. GPU latency, sec. CPU throughput, samples/sec. GPU throughput, samples/sec.
Teacher (RuBERT-base-cased-conversational) 679 0.655 0.031 0.3754 36.4902
Student (DistilRuBERT-small-cased-conversational) 409 0.1656 0.015 0.9692 71.3553

为了评估模型的质量,我们在分类、NER和问答任务上对DistilRuBERT-small进行了微调。得分和微调模型的存档可以在 DeepPavlov docs 中找到。此外,结果也可以在 paper 的表格1&2中找到,以及性能基准和训练细节。

引用

如果您发现该模型对您的研究有用,请引用 this 论文:

@misc{https://doi.org/10.48550/arxiv.2205.02340,
  doi = {10.48550/ARXIV.2205.02340},
  url = {https://arxiv.org/abs/2205.02340},
  author = {Kolesnikova, Alina and Kuratov, Yuri and Konovalov, Vasily and Burtsev, Mikhail},
  keywords = {Computation and Language (cs.CL), Machine Learning (cs.LG), FOS: Computer and information sciences, FOS: Computer and information sciences},
  title = {Knowledge Distillation of Russian Language Models with Reduction of Vocabulary},
  publisher = {arXiv},
  year = {2022},
  copyright = {arXiv.org perpetual, non-exclusive license}
}

[1]: P. Lison and J. Tiedemann, 2016, OpenSubtitles2016: Extracting Large Parallel Corpora from Movie and TV Subtitles. In Proceedings of the 10th International Conference on Language Resources and Evaluation (LREC 2016)

[2]: Shavrina T., Shapovalova O. (2017) TO THE METHODOLOGY OF CORPUS CONSTRUCTION FOR MACHINE LEARNING: «TAIGA» SYNTAX TREE CORPUS AND PARSER. in proc. of “CORPORA2017”, international conference , Saint-Petersbourg, 2017.

[3]: Sanh, V., Debut, L., Chaumond, J., & Wolf, T. (2019). DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108.

[4]: https://github.com/huggingface/transformers/tree/master/examples/research_projects/distillation