language:
- ru
distilrubert-tiny-cased-conversational
Conversational DistilRuBERT-tiny (Russian, cased, 3‑layers, 264‑hidden, 12‑heads, 11.8M parameters) was trained on OpenSubtitles[1], Dirty, Pikabu, and a Social Media segment of Taiga corpus[2] (as Conversational RuBERT). It can be considered as tiny copy of Conversational DistilRuBERT-small.
Our DistilRuBERT-tiny 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)
- MSE loss (between averaged student and teacher hidden states)
The key features are:
- unlike most of distilled language models, we didn't use KL loss during pre-training
- reduced vocabulary size (30K in tiny vs. 100K in base and small )
- two separate inputs for student: tokens obtained using student tokenizer (for MLM) and teacher tokens greedily splitted by student tokens (for MSE)
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 |
| distilrubert-tiny-cased-conversational | 10.4 | 31 | 41 |
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 Intel(R) Xeon(R) CPU E5-2698 v4 @ 2.20GHz and nVIDIA Tesla P100-SXM2.0 16Gb.
| Model name | Batch size | Seq len | Time, s || Mem, MB ||
|---|---|---|------||------||
| | | | CPU | GPU | CPU | GPU |
| rubert-base-cased-conversational | 1 | 512 | 0.147 | 0.014 | 897 | 1531 |
| distilrubert-base-cased-conversational | 1 | 512 | 0.083 | 0.006 | 766 | 1423 |
| distilrubert-small-cased-conversational | 1 | 512 | 0.03 | 0.002 | 600 | 1243 |
| cointegrated/rubert-tiny | 1 | 512 | 0.041 | 0.003 | 272 | 919 |
| distilrubert-tiny-cased-conversational | 1 | 512 | 0.023 | 0.003 | 206 | 855 |
| rubert-base-cased-conversational | 16 | 512 | 2.839 | 0.182 | 1499 | 2071 |
| distilrubert-base-cased-conversational | 16 | 512 | 1.065 | 0.055 | 2541 | 2927 |
| distilrubert-small-cased-conversational | 16 | 512 | 0.373 | 0.003 | 1360 | 1943 |
| cointegrated/rubert-tiny | 16 | 512 | 0.628 | 0.004 | 1293 | 2221 |
| distilrubert-tiny-cased-conversational | 16 | 512 | 0.219 | 0.003 | 633 | 1291 |
To evaluate model quality, we fine-tuned DistilRuBERT-small on classification (RuSentiment, ParaPhraser), NER and question answering data sets for Russian and obtained scores very similar to the Conversational DistilRuBERT-small.
[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