kgmyh/klue_bert-base_finetuning

SentenceTransformer based on klue/bert-base

This is a sentence-transformers model finetuned from klue/bert-base on the klue dataset. It maps sentences & paragraphs to a 768-dimensional dense vector space and can be used for semantic textual similarity, semantic search, paraphrase mining, text classification, clustering, and more.

Model Details

Model Description

• Model Type: Sentence Transformer
• Base model: klue/bert-base
• Maximum Sequence Length: 512 tokens
• Output Dimensionality: 768 dimensions
• Similarity Function: Cosine Similarity
• Training Dataset:
  • klue
• Language: ko

Model Sources

• Documentation: Sentence Transformers Documentation
• Repository: Sentence Transformers on GitHub
• Hugging Face: Sentence Transformers on Hugging Face

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 512, 'do_lower_case': False}) with Transformer model: BertModel 
  (1): Pooling({'word_embedding_dimension': 768, 'pooling_mode_cls_token': False, 'pooling_mode_mean_tokens': True, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False, 'pooling_mode_weightedmean_tokens': False, 'pooling_mode_lasttoken': False, 'include_prompt': True})
)

Usage

Direct Usage (Sentence Transformers)

First install the Sentence Transformers library:

pip install -U sentence-transformers

Then you can load this model and run inference.

from sentence_transformers import SentenceTransformer

# Download from the 🤗 Hub
model = SentenceTransformer("kgmyh/klue_bert-base_finetuning")
# Run inference
sentences = [
    'klue-sts-v1_dev_00159',
    'airbnb-sampled',
    '2층에 얇은 벽 하나 사이로 방이 두 개 있습니다.',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 768]

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings, embeddings)
print(similarities.shape)
# [3, 3]

Training Details

Training Dataset

klue

• Dataset: klue at 349481e
• Size: 11,668 training samples
• Columns: guid, source, sentence1, sentence2, and labels
• Approximate statistics based on the first 1000 samples:

  	guid	source	sentence1	sentence2	labels
  type	string	string	string	string	dict
  details	min: 17 tokens mean: 17.91 tokens max: 18 tokens	min: 7 tokens mean: 10.01 tokens max: 11 tokens	min: 7 tokens mean: 19.55 tokens max: 61 tokens	min: 6 tokens mean: 19.35 tokens max: 60 tokens

• Samples:

  guid	source	sentence1	sentence2	labels
  klue-sts-v1_train_00000	airbnb-rtt	숙소 위치는 찾기 쉽고 일반적인 한국의 반지하 숙소입니다.	숙박시설의 위치는 쉽게 찾을 수 있고 한국의 대표적인 반지하 숙박시설입니다.	{'label': 3.7, 'real-label': 3.714285714285714, 'binary-label': 1}
  klue-sts-v1_train_00001	policy-sampled	위반행위 조사 등을 거부·방해·기피한 자는 500만원 이하 과태료 부과 대상이다.	시민들 스스로 자발적인 예방 노력을 한 것은 아산 뿐만이 아니었다.	{'label': 0.0, 'real-label': 0.0, 'binary-label': 0}
  klue-sts-v1_train_00002	paraKQC-sampled	회사가 보낸 메일은 이 지메일이 아니라 다른 지메일 계정으로 전달해줘.	사람들이 주로 네이버 메일을 쓰는 이유를 알려줘	{'label': 0.3, 'real-label': 0.3333333333333333, 'binary-label': 0}

• Loss: MultipleNegativesRankingLoss with these parameters:

  {
      "scale": 20.0,
      "similarity_fct": "cos_sim"
  }
  

Evaluation Dataset

klue

• Dataset: klue at 349481e
• Size: 519 evaluation samples
• Columns: guid, source, sentence1, sentence2, and labels
• Approximate statistics based on the first 519 samples:

  	guid	source	sentence1	sentence2	labels
  type	string	string	string	string	dict
  details	min: 17 tokens mean: 17.82 tokens max: 18 tokens	min: 7 tokens mean: 9.72 tokens max: 11 tokens	min: 6 tokens mean: 19.47 tokens max: 61 tokens	min: 7 tokens mean: 19.42 tokens max: 58 tokens

• Samples:

  guid	source	sentence1	sentence2	labels
  klue-sts-v1_dev_00000	airbnb-rtt	무엇보다도 호스트분들이 너무 친절하셨습니다.	무엇보다도, 호스트들은 매우 친절했습니다.	{'label': 4.9, 'real-label': 4.857142857142857, 'binary-label': 1}
  klue-sts-v1_dev_00001	airbnb-sampled	주요 관광지 모두 걸어서 이동가능합니다.	위치는 피렌체 중심가까지 걸어서 이동 가능합니다.	{'label': 1.4, 'real-label': 1.428571428571429, 'binary-label': 0}
  klue-sts-v1_dev_00002	policy-sampled	학생들의 균형 있는 영어능력을 향상시킬 수 있는 학교 수업을 유도하기 위해 2018학년도 수능부터 도입된 영어 영역 절대평가는 올해도 유지한다.	영어 영역의 경우 학생들이 한글 해석본을 암기하는 문제를 해소하기 위해 2016학년도부터 적용했던 EBS 연계 방식을 올해도 유지한다.	{'label': 1.3, 'real-label': 1.285714285714286, 'binary-label': 0}

• Loss: MultipleNegativesRankingLoss with these parameters:

  {
      "scale": 20.0,
      "similarity_fct": "cos_sim"
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 16
• per_device_eval_batch_size: 64
• weight_decay: 0.01
• num_train_epochs: 1
• warmup_steps: 100
• load_best_model_at_end: True

All Hyperparameters

Click to expand

• overwrite_output_dir: False
• do_predict: False
• eval_strategy: steps
• prediction_loss_only: True
• per_device_train_batch_size: 16
• per_device_eval_batch_size: 64
• per_gpu_train_batch_size: None
• per_gpu_eval_batch_size: None
• gradient_accumulation_steps: 1
• eval_accumulation_steps: None
• torch_empty_cache_steps: None
• learning_rate: 5e-05
• weight_decay: 0.01
• adam_beta1: 0.9
• adam_beta2: 0.999
• adam_epsilon: 1e-08
• max_grad_norm: 1.0
• num_train_epochs: 1
• max_steps: -1
• lr_scheduler_type: linear
• lr_scheduler_kwargs: {}
• warmup_ratio: 0.0
• warmup_steps: 100
• log_level: passive
• log_level_replica: warning
• log_on_each_node: True
• logging_nan_inf_filter: True
• save_safetensors: True
• save_on_each_node: False
• save_only_model: False
• restore_callback_states_from_checkpoint: False
• no_cuda: False
• use_cpu: False
• use_mps_device: False
• seed: 42
• data_seed: None
• jit_mode_eval: False
• use_ipex: False
• bf16: False
• fp16: False
• fp16_opt_level: O1
• half_precision_backend: auto
• bf16_full_eval: False
• fp16_full_eval: False
• tf32: None
• local_rank: 0
• ddp_backend: None
• tpu_num_cores: None
• tpu_metrics_debug: False
• debug: []
• dataloader_drop_last: False
• dataloader_num_workers: 0
• dataloader_prefetch_factor: None
• past_index: -1
• disable_tqdm: False
• remove_unused_columns: True
• label_names: None
• load_best_model_at_end: True
• ignore_data_skip: False
• fsdp: []
• fsdp_min_num_params: 0
• fsdp_config: {'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': False}
• fsdp_transformer_layer_cls_to_wrap: None
• accelerator_config: {'split_batches': False, 'dispatch_batches': None, 'even_batches': True, 'use_seedable_sampler': True, 'non_blocking': False, 'gradient_accumulation_kwargs': None}
• deepspeed: None
• label_smoothing_factor: 0.0
• optim: adamw_torch
• optim_args: None
• adafactor: False
• group_by_length: False
• length_column_name: length
• ddp_find_unused_parameters: None
• ddp_bucket_cap_mb: None
• ddp_broadcast_buffers: False
• dataloader_pin_memory: True
• dataloader_persistent_workers: False
• skip_memory_metrics: True
• use_legacy_prediction_loop: False
• push_to_hub: False
• resume_from_checkpoint: None
• hub_model_id: None
• hub_strategy: every_save
• hub_private_repo: None
• hub_always_push: False
• gradient_checkpointing: False
• gradient_checkpointing_kwargs: None
• include_inputs_for_metrics: False
• include_for_metrics: []
• eval_do_concat_batches: True
• fp16_backend: auto
• push_to_hub_model_id: None
• push_to_hub_organization: None
• mp_parameters:
• auto_find_batch_size: False
• full_determinism: False
• torchdynamo: None
• ray_scope: last
• ddp_timeout: 1800
• torch_compile: False
• torch_compile_backend: None
• torch_compile_mode: None
• dispatch_batches: None
• split_batches: None
• include_tokens_per_second: False
• include_num_input_tokens_seen: False
• neftune_noise_alpha: None
• optim_target_modules: None
• batch_eval_metrics: False
• eval_on_start: False
• use_liger_kernel: False
• eval_use_gather_object: False
• average_tokens_across_devices: False
• prompts: None
• batch_sampler: batch_sampler
• multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	Training Loss	Validation Loss
0.0137	10	2.9128	-
0.0274	20	2.8336	-
0.0411	30	2.8053	-
0.0548	40	2.7919	-
0.0685	50	2.7815	-
0.0822	60	2.7722	-
0.0959	70	2.7779	-
0.1096	80	2.7768	-
0.1233	90	2.7846	-
0.1370	100	2.7747	-
0.1507	110	2.7786	-
0.1644	120	2.7719	-
0.1781	130	2.7745	-
0.1918	140	2.7747	-
0.2055	150	2.7749	-
0.2192	160	2.7715	-
0.2329	170	2.7863	-
0.2466	180	2.7732	-
0.2603	190	2.7744	-
0.2740	200	2.7754	-
0.2877	210	2.7726	-
0.3014	220	2.7718	-
0.3151	230	2.774	-
0.3288	240	2.7748	-
0.3425	250	2.7708	-
0.3562	260	2.7728	-
0.3699	270	2.7746	-
0.3836	280	2.7739	-
0.3973	290	2.7721	-
0.4110	300	2.7747	-
0.4247	310	2.7746	-
0.4384	320	2.7732	-
0.4521	330	2.7739	-
0.4658	340	2.7724	-
0.4795	350	2.7736	-
0.4932	360	2.7736	-
0.5068	370	2.7735	-
0.5205	380	2.7734	-
0.5342	390	2.7726	-
0.5479	400	2.7734	-
0.5616	410	2.7726	-
0.5753	420	2.7731	-
0.5890	430	2.7735	-
0.6027	440	2.7734	-
0.6164	450	2.7741	-
0.6301	460	2.7737	-
0.6438	470	2.7717	-
0.6575	480	2.7739	-
0.6712	490	2.7727	-
0.6849	500	2.7729	4.129
0.6986	510	2.7723	-
0.7123	520	2.7729	-
0.7260	530	2.7736	-
0.7397	540	2.7725	-
0.7534	550	2.7735	-
0.7671	560	2.7737	-
0.7808	570	2.7731	-
0.7945	580	2.7733	-
0.8082	590	2.7725	-
0.8219	600	2.773	-
0.8356	610	2.7729	-
0.8493	620	2.7724	-
0.8630	630	2.7719	-
0.8767	640	2.7719	-
0.8904	650	2.7735	-
0.9041	660	2.7731	-
0.9178	670	2.7716	-
0.9315	680	2.7736	-
0.9452	690	2.7734	-
0.9589	700	2.7728	-
0.9726	710	2.7721	-
0.9863	720	2.7726	-
1.0	730	2.6339	-

• The bold row denotes the saved checkpoint.

Framework Versions

• Python: 3.12.7
• Sentence Transformers: 3.3.1
• Transformers: 4.48.0
• PyTorch: 2.5.1+cpu
• Accelerate: 1.1.1
• Datasets: 3.2.0
• Tokenizers: 0.21.0

Citation

BibTeX

Sentence Transformers

@inproceedings{reimers-2019-sentence-bert,
    title = "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
    author = "Reimers, Nils and Gurevych, Iryna",
    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
    month = "11",
    year = "2019",
    publisher = "Association for Computational Linguistics",
    url = "https://arxiv.org/abs/1908.10084",
}

MultipleNegativesRankingLoss

@misc{henderson2017efficient,
    title={Efficient Natural Language Response Suggestion for Smart Reply},
    author={Matthew Henderson and Rami Al-Rfou and Brian Strope and Yun-hsuan Sung and Laszlo Lukacs and Ruiqi Guo and Sanjiv Kumar and Balint Miklos and Ray Kurzweil},
    year={2017},
    eprint={1705.00652},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}