You can deploy this model and start earning money today!

Views

No views yet

dragonkue/colbert-ko-0.1b

Name: colbert-ko-0.1b
Availability: InStock

This is a Korean ColBERT model finetuned from skt/A.X-Encoder-base using PyLate. It maps sentences and paragraphs to sequences of 32, 64, 96, or 128-dimensional token embeddings, and uses the MaxSim operator for late-interaction retrieval.

Model Details

Model Description

Model Type: PyLate model
Base model: skt/A.X-Encoder-base
Document Length: 2048 tokens
Query Length: 32 tokens
Output Dimensionality: 128 (Matryoshka: 32 / 64 / 96 / 128)
Similarity Function: MaxSim
Language: Korean
License: Apache-2.0

Model Sources

Documentation: PyLate Documentation
Repository: PyLate on GitHub
Hugging Face: PyLate models on Hugging Face

Full Model Architecture

ColBERTWrapper(
  (0): Transformer({'max_seq_length': 2047, 'do_lower_case': False, 'architecture': 'ModernBertModel'})
  (1): Dense({'in_features': 768, 'out_features': 128, 'bias': False, 'activation_function': 'torch.nn.modules.linear.Identity', 'use_residual': False})
)

Usage

This model supports Matryoshka embeddings with multiple dimensions (32, 64, 96, 128) via separate projection heads (Jina-ColBERT-v2 style).

The model ships a bundled model.py so it can be loaded with the transformers library alone — no third-party retrieval package is required. The custom code is opt-in: pass trust_remote_code=True to from_pretrained.

Installation

pip install transformers torch

Quick Start

python

from transformers import AutoModel, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("dragonkue/colbert-ko-0.1b")
model = AutoModel.from_pretrained(
    "dragonkue/colbert-ko-0.1b",
    trust_remote_code=True,
)
model.eval()
# model.cuda()  # optional

# Pick embedding dimension (32, 64, 96, or 128). Default is 128.
model.set_active_dim(128)

queries = ["검색 쿼리"]
documents = ["문서 내용"]

q_embs = model.encode(tokenizer, queries, is_query=True)
d_embs = model.encode(tokenizer, documents, is_query=False)

print(f"Query shape: {q_embs[0].shape}")  # (num_tokens, 128)
print(f"Doc shape:   {d_embs[0].shape}")  # (num_tokens, 128)

# ColBERT late-interaction (MaxSim) score for one (query, doc) pair
score = model.maxsim(q_embs[0], d_embs[0])
print(f"MaxSim: {score:.4f}")

Reranking

python

query = "대한민국의 수도는 어디인가요?"
candidates = [
    "대한민국의 수도는 서울이며, 인구가 가장 많은 도시이다.",
    "파리는 프랑스의 수도이고 에펠탑으로 유명하다.",
    "오늘 날씨는 맑고 기온은 25도 정도이다.",
]

q_embs = model.encode(tokenizer, [query], is_query=True)
c_embs = model.encode(tokenizer, candidates, is_query=False)

ranked = sorted(
    [(i, model.maxsim(q_embs[0], c_embs[i])) for i in range(len(candidates))],
    key=lambda x: x[1],
    reverse=True,
)
for rank, (i, score) in enumerate(ranked, start=1):
    print(f"  rank {rank}: score={score:.4f}  | {candidates[i]}")

Approximate Retrieval (PyLate PLAID, optional)

For PLAID indexing over a large corpus you can additionally use PyLate. PyLate loads the model as a standard sentence-transformers checkpoint (no trust_remote_code needed). The 128-dim head saved at 1_Dense/model.safetensors matches the Matryoshka 128-dim head, so PyLate produces the same embeddings as the snippet above.

python

from pylate import indexes, models, retrieve

model = models.ColBERT(model_name_or_path="dragonkue/colbert-ko-0.1b")

index = indexes.PLAID(
    index_folder="pylate-index",
    index_name="index",
    override=True,
)

documents_ids = ["1", "2", "3"]
documents = ["첫번째 문서입니다", "두번째 문서입니다", "세번째 문서입니다"]

documents_embeddings = model.encode(documents, is_query=False)
index.add_documents(
    documents_ids=documents_ids,
    documents_embeddings=documents_embeddings,
)

retriever = retrieve.ColBERT(index=index)
queries_embeddings = model.encode(["첫번째 문서 검색"], is_query=True)
scores = retriever.retrieve(queries_embeddings=queries_embeddings, k=3)
print(scores)

The PyLate path is fixed to the 128-dim head; for Matryoshka dimension switching (32 / 64 / 96 / 128), use the trust_remote_code=True path shown above.

Evaluation Results (NDCG@10)

Comparison with Other Models (dim128)

Model	AutoRAG	Ko-StrategyQA	NanoBEIR-Ko	Avg
dragonkue-colbert-ko-0.1b (149M)	0.970	0.735	0.503	0.737
BGE-M3-MultiVec (568M)	0.844	0.797	0.569	0.737
LFM2-ColBERT (353M)	0.833	0.757	0.528	0.706
colbert-ko-v1 (149M)	0.966	0.713	0.476	0.718

Performance by Embedding Dimension

Dimension	AutoRAG	Ko-StrategyQA	NanoBEIR-Ko
32	0.976	0.713	0.485
64	0.979	0.725	0.492
96	0.969	0.739	0.501
128	0.970	0.735	0.503

Differences across dimensions are small (within ≈0.01 on each benchmark) and can flip between runs on the more saturated benchmarks such as AutoRAG. The 128-dim head is the default; smaller dimensions trade a few points of NDCG for a 2–4× reduction in vector size.

Scores for dragonkue-colbert-ko-0.1b are measured here with raw MaxSim (no PLAID quantization) on transformers==5.3.0 and pytorch==2.9. Scores for the other models are reproduced from prior reports and may differ from a re-run on the same stack.

Loss: src.losses.MatryoshkaColBERTLoss with these parameters:
json
1{ 2 "dims": [ 3 32, 4 64, 5 96, 6 128 7 ], 8 "weights": [ 9 0.25, 10 0.25, 11 0.25, 12 0.25 13 ], 14 "temperature": 1.0 15}

Training Hyperparameters

Non-Default Hyperparameters

eval_strategy: steps
per_device_train_batch_size: 64
learning_rate: 1e-05
num_train_epochs: 2
warmup_ratio: 0.1
fp16: True
dataloader_drop_last: True
gradient_checkpointing: True
gradient_checkpointing_kwargs: {'use_reentrant': False}
router_mapping: {'anchor': 'query', 'positive': 'document', 'neg_0': 'document'}

All Hyperparameters

Click to expand

overwrite_output_dir: False
do_predict: False
eval_strategy: steps
prediction_loss_only: True
per_device_train_batch_size: 64
per_device_eval_batch_size: 8
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: 1e-05
weight_decay: 0.0
adam_beta1: 0.9
adam_beta2: 0.999
adam_epsilon: 1e-08
max_grad_norm: 1.0
num_train_epochs: 2
max_steps: -1
lr_scheduler_type: linear
lr_scheduler_kwargs: {}
warmup_ratio: 0.1
warmup_steps: 0
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: True
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: True
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: False
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}
parallelism_config: None
deepspeed: None
label_smoothing_factor: 0.0
optim: adamw_torch_fused
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
hub_revision: None
gradient_checkpointing: True
gradient_checkpointing_kwargs: {'use_reentrant': False}
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
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
liger_kernel_config: None
eval_use_gather_object: False
average_tokens_across_devices: False
prompts: None
batch_sampler: batch_sampler
multi_dataset_batch_sampler: proportional
router_mapping: {'anchor': 'query', 'positive': 'document', 'neg_0': 'document'}
learning_rate_mapping: {}

Training Logs

Click to expand

Epoch	Step	Training Loss
0.0017	10	4.1388
0.0034	20	4.1142
0.0051	30	3.9797
0.0067	40	3.8761
0.0084	50	3.6167
0.0101	60	3.424
0.0118	70	3.0256
0.0135	80	2.827
0.0152	90	2.5787
0.0169	100	2.2696
0.0185	110	2.0266
0.0202	120	1.6815
0.0219	130	1.4739
0.0236	140	1.2877
0.0253	150	1.1474
0.0270	160	1.0143
0.0286	170	0.9363
0.0303	180	0.9189
0.0320	190	0.7442
0.0337	200	0.6919
0.0354	210	0.6251
0.0371	220	0.6527
0.0388	230	0.5923
0.0404	240	0.572
0.0421	250	0.5255
0.0438	260	0.4407
0.0455	270	0.5038
0.0472	280	0.3939
0.0489	290	0.3938
0.0506	300	0.3253
0.0522	310	0.335
0.0539	320	0.2855
0.0556	330	0.2396
0.0573	340	0.252
0.0590	350	0.2299
0.0607	360	0.2133
0.0624	370	0.2186
0.0640	380	0.1935
0.0657	390	0.1743
0.0674	400	0.1462
0.0691	410	0.1552
0.0708	420	0.1491
0.0725	430	0.1581
0.0741	440	0.1635
0.0758	450	0.1383
0.0775	460	0.1377
0.0792	470	0.1155
0.0809	480	0.1184
0.0826	490	0.1333
0.0843	500	0.1341
0.0859	510	0.1259
0.0876	520	0.0748
0.0893	530	0.1342
0.0910	540	0.1058
0.0927	550	0.1024
0.0944	560	0.0921
0.0961	570	0.104
0.0977	580	0.1069
0.0994	590	0.0925
0.1011	600	0.1146
0.1028	610	0.0682
0.1045	620	0.0711
0.1062	630	0.1491
0.1079	640	0.0602
0.1095	650	0.0753
0.1112	660	0.0713
0.1129	670	0.0739
0.1146	680	0.0783
0.1163	690	0.0678
0.1180	700	0.0963
0.1196	710	0.0677
0.1213	720	0.0829
0.1230	730	0.0719
0.1247	740	0.0646
0.1264	750	0.0927
0.1281	760	0.0755
0.1298	770	0.0799
0.1314	780	0.0535
0.1331	790	0.0555
0.1348	800	0.0804
0.1365	810	0.0627
0.1382	820	0.0726
0.1399	830	0.0685
0.1416	840	0.0421
0.1432	850	0.0895
0.1449	860	0.0964
0.1466	870	0.0515
0.1483	880	0.0825
0.1500	890	0.0801
0.1517	900	0.0579
0.1534	910	0.0559
0.1550	920	0.0432
0.1567	930	0.0553
0.1584	940	0.0577
0.1601	950	0.0451
0.1618	960	0.049
0.1635	970	0.0459
0.1651	980	0.0684
0.1668	990	0.0449
0.1685	1000	0.0392
0.1702	1010	0.071
0.1719	1020	0.0511
0.1736	1030	0.0501
0.1753	1040	0.0464
0.1769	1050	0.0678
0.1786	1060	0.0597
0.1803	1070	0.0569
0.1820	1080	0.044
0.1837	1090	0.0452
0.1854	1100	0.0394
0.1871	1110	0.0496
0.1887	1120	0.0296
0.1904	1130	0.0321
0.1921	1140	0.0525
0.1938	1150	0.058
0.1955	1160	0.0552
0.1972	1170	0.035
0.1989	1180	0.0468
0.1999	1186	-
0.2005	1190	0.0383
0.2022	1200	0.0599
0.2039	1210	0.0572
0.2056	1220	0.0383
0.2073	1230	0.0486
0.2090	1240	0.0407
0.2107	1250	0.044
0.2123	1260	0.04
0.2140	1270	0.0338
0.2157	1280	0.036
0.2174	1290	0.0511
0.2191	1300	0.0472
0.2208	1310	0.031
0.2224	1320	0.0614
0.2241	1330	0.0388
0.2258	1340	0.0403
0.2275	1350	0.047
0.2292	1360	0.033
0.2309	1370	0.0524
0.2326	1380	0.0357
0.2342	1390	0.0463
0.2359	1400	0.0355
0.2376	1410	0.0411
0.2393	1420	0.028
0.2410	1430	0.0386
0.2427	1440	0.0553
0.2444	1450	0.0353
0.2460	1460	0.0462
0.2477	1470	0.0399
0.2494	1480	0.0319
0.2511	1490	0.0456
0.2528	1500	0.0302
0.2545	1510	0.0366
0.2562	1520	0.0409
0.2578	1530	0.0337
0.2595	1540	0.0362
0.2612	1550	0.0318
0.2629	1560	0.0433
0.2646	1570	0.0379
0.2663	1580	0.0419
0.2679	1590	0.0225
0.2696	1600	0.0269
0.2713	1610	0.0295
0.2730	1620	0.048
0.2747	1630	0.0382
0.2764	1640	0.0341
0.2781	1650	0.0334
0.2797	1660	0.0534
0.2814	1670	0.0445
0.2831	1680	0.0284
0.2848	1690	0.0327
0.2865	1700	0.0309
0.2882	1710	0.0372
0.2899	1720	0.0384
0.2915	1730	0.022
0.2932	1740	0.0266
0.2949	1750	0.0399
0.2966	1760	0.0342
0.2983	1770	0.0391
0.3000	1780	0.0349
0.3017	1790	0.0365
0.3033	1800	0.0322
0.3050	1810	0.0414
0.3067	1820	0.0297
0.3084	1830	0.0446
0.3101	1840	0.0312
0.3118	1850	0.0379
0.3134	1860	0.0252
0.3151	1870	0.0424
0.3168	1880	0.0367
0.3185	1890	0.0226
0.3202	1900	0.0319
0.3219	1910	0.0189
0.3236	1920	0.0219
0.3252	1930	0.0341
0.3269	1940	0.0505
0.3286	1950	0.0176
0.3303	1960	0.0328
0.3320	1970	0.0276
0.3337	1980	0.0251
0.3354	1990	0.0603
0.3370	2000	0.0243
0.3387	2010	0.0316
0.3404	2020	0.0294
0.3421	2030	0.025
0.3438	2040	0.0255
0.3455	2050	0.0318
0.3472	2060	0.025
0.3488	2070	0.0273
0.3505	2080	0.0338
0.3522	2090	0.0299
0.3539	2100	0.0275
0.3556	2110	0.0184
0.3573	2120	0.0244
0.3589	2130	0.0432
0.3606	2140	0.0325
0.3623	2150	0.0525
0.3640	2160	0.0329
0.3657	2170	0.0236
0.3674	2180	0.0309
0.3691	2190	0.0195
0.3707	2200	0.0318
0.3724	2210	0.0229
0.3741	2220	0.0312
0.3758	2230	0.0186
0.3775	2240	0.0231
0.3792	2250	0.0262
0.3809	2260	0.0287
0.3825	2270	0.0299
0.3842	2280	0.0302
0.3859	2290	0.0281
0.3876	2300	0.0252
0.3893	2310	0.0362
0.3910	2320	0.0266
0.3927	2330	0.0304
0.3943	2340	0.0259
0.3960	2350	0.0276
0.3977	2360	0.0219
0.3994	2370	0.0361
0.3997	2372	-

Framework Versions

Python: 3.11.13
Sentence Transformers: 5.1.1
PyLate: 1.3.4
Transformers: 4.56.2
PyTorch: 2.8.0+cu128
Accelerate: 1.12.0
Datasets: 4.4.1
Tokenizers: 0.22.2-rc0

Citation

BibTeX

Sentence Transformers

bibtex

@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"
}

PyLate

bibtex

@misc{PyLate,
title={PyLate: Flexible Training and Retrieval for Late Interaction Models},
author={Chaffin, Antoine and Sourty, Raphaël},
url={https://github.com/lightonai/pylate},
year={2024}
}

Jina ColBERT v2

bibtex

@article{jina-colbert-v2,
    title={Jina-ColBERT-v2: A General-Purpose Multilingual Late Interaction Retriever},
    author={Rohan Jha and Bo Wang and Michael Günther and Saba Sturua and Mohammad Kalim Akram and Han Xiao},
    year={2024},
    journal={arXiv preprint arXiv:2408.16672},
    url={https://arxiv.org/abs/2408.16672}
}