distilrubert-tiny-cased-conversational-5k

Conversational DistilRuBERT-tiny-5k (Russian, cased, 3‑layers, 264‑hidden, 12‑heads, 3.6M parameters, 5k vocab) was trained on OpenSubtitles[1], Dirty, Pikabu, and a Social Media segment of Taiga corpus[2] (as Conversational RuBERT).

Our DistilRuBERT-tiny-5k is highly inspired by [3], [4] and architecture is very close to [5]. Namely, we use

MLM loss (between token labels and student output distribution)
KL loss (between averaged student and teacher hidden states)

The key feature is:

reduced vocabulary size (5K vs 30K in tiny vs. 100K in base and small)

Here is comparison between teacher model (Conversational RuBERT) and other distilled models.

Model name	# params, M	# vocab, K	Mem., MB
`rubert-base-cased-conversational`	177.9	120	679
`distilrubert-base-cased-conversational`	135.5	120	517
`distilrubert-small-cased-conversational`	107.1	120	409
`cointegrated/rubert-tiny`	11.8	30	46
`cointegrated/rubert-tiny2`	29.3	84	112
`distilrubert-tiny-cased-conversational-v1`	10.4	31	41
`distilrubert-tiny-cased-conversational-5k`	3.6	5	14

DistilRuBERT-tiny was trained for about 100 hrs. on 7 nVIDIA Tesla P100-SXM2.0 16Gb.

We used PyTorchBenchmark from transformers to evaluate model's performance and compare it with other pre-trained language models for Russian. All tests were performed on NVIDIA GeForce GTX 1080 Ti and Intel(R) Core(TM) i7-7700K CPU @ 4.20GHz

| Model name | Batch size | Seq len | Time, s || Mem, MB || |---|---|---|------||------|| | | | | CPU | GPU | CPU | GPU | | rubert-base-cased-conversational | 16 | 512 | 5.283 | 0.1866 | 1550 | 1938 | | distilrubert-base-cased-conversational | 16 | 512 | 2.335 | 0.0553 | 2177 | 2794 | | distilrubert-small-cased-conversational | 16 | 512 | 0.802 | 0.0015 | 1541 | 1810 | | cointegrated/rubert-tiny | 16 | 512 | 0.942 | 0.0022 | 1308 | 2088 | | cointegrated/rubert-tiny2 | 16 | 512 | 1.786 | 0.0023 | 3054 | 3848 | | distilrubert-tiny-cased-conversational-v1 | 16 | 512 | 0.374 | 0.002 | 714 | 1158 | | distilrubert-tiny-cased-conversational-5k | 16 | 512 | 0.354 | 0.0018 | 664 | 1126 |

To evaluate model quality, we fine-tuned DistilRuBERT-tiny-5k on classification (RuSentiment, ParaPhraser), NER and question answering data sets for Russian. The results could be found in the paper Table 4 as well as performance benchmarks and training details.

Citation

If you found the model useful for your research, we are kindly ask to cite this paper:

@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

[5]: https://habr.com/ru/post/562064/, https://huggingface.co/cointegrated/rubert-tiny