KoMiniLM

Current language models usually consist of hundreds of millions of parameters which brings challenges for fine-tuning and online serving in real-life applications due to latency and capacity constraints. In this project, we release a light weight korean language model to address the aforementioned shortcomings of existing language models.

Quick tour

from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained("BM-K/KoMiniLM") # 23M model
model = AutoModel.from_pretrained("BM-K/KoMiniLM")

inputs = tokenizer("안녕 세상아!", return_tensors="pt")
outputs = model(**inputs)

Update history

** Updates on 2022.06.20 **

• Release KoMiniLM-bert-68M

** Updates on 2022.05.24 **

• Release KoMiniLM-bert-23M

Pre-training

Teacher Model: KLUE-BERT(base)

Object

Self-Attention Distribution and Self-Attention Value-Relation [Wang et al., 2020] were distilled from each discrete layer of the teacher model to the student model. Wang et al. distilled in the last layer of the transformer, but that was not the case in this project.

Data sets

Data	News comments	News article
size	10G	10G

Note

• Performance can be further improved by adding wiki data to training.
• The crawling and preprocessing code for the News article is here.

Config

• KoMiniLM-23M

{
  "architectures": [
    "BertForPreTraining"
  ],
  "attention_probs_dropout_prob": 0.1,
  "classifier_dropout": null,
  "hidden_act": "gelu",
  "hidden_dropout_prob": 0.1,
  "hidden_size": 384,
  "initializer_range": 0.02,
  "intermediate_size": 1536,
  "layer_norm_eps": 1e-12,
  "max_position_embeddings": 512,
  "model_type": "bert",
  "num_attention_heads": 12,
  "num_hidden_layers": 6,
  "output_attentions": true,
  "pad_token_id": 0,
  "position_embedding_type": "absolute",
  "return_dict": false,
  "torch_dtype": "float32",
  "transformers_version": "4.13.0",
  "type_vocab_size": 2,
  "use_cache": true,
  "vocab_size": 32000
}

• KoMiniLM-68M

{
  "architectures": [
    "BertForPreTraining"
  ],
  "attention_probs_dropout_prob": 0.1,
  "classifier_dropout": null,
  "hidden_act": "gelu",
  "hidden_dropout_prob": 0.1,
  "hidden_size": 768,
  "initializer_range": 0.02,
  "intermediate_size": 3072,
  "layer_norm_eps": 1e-12,
  "max_position_embeddings": 512,
  "model_type": "bert",
  "num_attention_heads": 12,
  "num_hidden_layers": 6,
  "output_attentions": true,
  "pad_token_id": 0,
  "position_embedding_type": "absolute",
  "return_dict": false,
  "torch_dtype": "float32",
  "transformers_version": "4.13.0",
  "type_vocab_size": 2,
  "use_cache": true,
  "vocab_size": 32000
}

Performance on subtasks

• The results of our fine-tuning experiments are an average of 3 runs for each task.

cd KoMiniLM-Finetune
bash scripts/run_all_kominilm.sh

	#Param	Average	NSMC (Acc)	Naver NER (F1)	PAWS (Acc)	KorNLI (Acc)	KorSTS (Spearman)	Question Pair (Acc)	KorQuaD (Dev) (EM/F1)
KoBERT(KLUE)	110M	86.84	90.20±0.07	87.11±0.05	81.36±0.21	81.06±0.33	82.47±0.14	95.03±0.44	84.43±0.18 / 93.05±0.04
KcBERT	108M	78.94	89.60±0.10	84.34±0.13	67.02±0.42	74.17±0.52	76.57±0.51	93.97±0.27	60.87±0.27 / 85.01±0.14
KoBERT(SKT)	92M	79.73	89.28±0.42	87.54±0.04	80.93±0.91	78.18±0.45	75.98±2.81	94.37±0.31	51.94±0.60 / 79.69±0.66
DistilKoBERT	28M	74.73	88.39±0.08	84.22±0.01	61.74±0.45	70.22±0.14	72.11±0.27	92.65±0.16	52.52±0.48 / 76.00±0.71
KoMiniLM†	68M	85.90	89.84±0.02	85.98±0.09	80.78±0.30	79.28±0.17	81.00±0.07	94.89±0.37	83.27±0.08 / 92.08±0.06
KoMiniLM†	23M	84.79	89.67±0.03	84.79±0.09	78.67±0.45	78.10±0.07	78.90±0.11	94.81±0.12	82.11±0.42 / 91.21±0.29

• NSMC (Naver Sentiment Movie Corpus)
• Naver NER (NER task on Naver NLP Challenge 2018)
• PAWS (Korean Paraphrase Adversaries from Word Scrambling)
• KorNLI/KorSTS (Korean Natural Language Understanding)
• Question Pair (Paired Question)
• KorQuAD (The Korean Question Answering Dataset)

License

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.

Reference

• KLUE BERT
• KcBERT
• SKT KoBERT
• DistilKoBERT
• lassl