pj-mathematician/JobBGE-m3

Job - Job matching finetuned BAAI/bge-m3

Top performing model on TalentCLEF 2025 Task A. Use it for multilingual job title matching

Model Details

Model Description

• Model Type: Sentence Transformer
• Base model: BAAI/bge-m3
• Maximum Sequence Length: 512 tokens
• Output Dimensionality: 1024 dimensions
• Similarity Function: Cosine Similarity
• Training Datasets:
  • full_en
  • full_de
  • full_es
  • full_zh
  • mix

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: XLMRobertaModel 
  (1): Pooling({'word_embedding_dimension': 1024, 'pooling_mode_cls_token': True, 'pooling_mode_mean_tokens': False, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False, 'pooling_mode_weightedmean_tokens': False, 'pooling_mode_lasttoken': False, 'include_prompt': True})
  (2): Normalize()
)

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("pj-mathematician/JobBGE-m3")
# Run inference
sentences = [
    'Volksvertreter',
    'Parlamentarier',
    'Oberbürgermeister',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 1024]

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

Evaluation

Metrics

Information Retrieval

• Datasets: full_en, full_es, full_de, full_zh, mix_es, mix_de and mix_zh
• Evaluated with InformationRetrievalEvaluator

Metric	full_en	full_es	full_de	full_zh	mix_es	mix_de	mix_zh
cosine_accuracy@1	0.6476	0.1135	0.2956	0.6796	0.7395	0.6927	0.1789
cosine_accuracy@20	0.9905	1.0	0.9852	0.9903	0.9636	0.9641	1.0
cosine_accuracy@50	0.9905	1.0	0.9901	0.9903	0.9828	0.9839	1.0
cosine_accuracy@100	0.9905	1.0	0.9901	0.9903	0.9927	0.9922	1.0
cosine_accuracy@150	0.9905	1.0	0.9901	0.9903	0.9948	0.9932	1.0
cosine_accuracy@200	0.9905	1.0	0.9901	0.9903	0.9964	0.9943	1.0
cosine_precision@1	0.6476	0.1135	0.2956	0.6796	0.7395	0.6927	0.1789
cosine_precision@20	0.5062	0.5668	0.5404	0.4709	0.1249	0.128	0.1544
cosine_precision@50	0.3065	0.3903	0.3828	0.2804	0.0517	0.0533	0.0618
cosine_precision@100	0.1858	0.2525	0.2503	0.1732	0.0263	0.0271	0.0309
cosine_precision@150	0.1325	0.1901	0.1878	0.1239	0.0176	0.0181	0.0206
cosine_precision@200	0.1025	0.1508	0.1503	0.0977	0.0133	0.0136	0.0154
cosine_recall@1	0.0669	0.0035	0.0111	0.0643	0.2854	0.2604	0.0577
cosine_recall@20	0.5392	0.3796	0.3433	0.5119	0.9226	0.9285	1.0
cosine_recall@50	0.72	0.5636	0.534	0.6727	0.9548	0.965	1.0
cosine_recall@100	0.8254	0.6727	0.6499	0.788	0.9705	0.9796	1.0
cosine_recall@150	0.872	0.736	0.7101	0.8329	0.9766	0.9837	1.0
cosine_recall@200	0.9006	0.7698	0.7513	0.8687	0.9811	0.9862	1.0
cosine_ndcg@1	0.6476	0.1135	0.2956	0.6796	0.7395	0.6927	0.1789
cosine_ndcg@20	0.6822	0.6136	0.5648	0.6515	0.8119	0.7967	0.5443
cosine_ndcg@50	0.6975	0.5908	0.5522	0.6599	0.8208	0.8069	0.5443
cosine_ndcg@100	0.752	0.6168	0.5796	0.7157	0.8243	0.8102	0.5443
cosine_ndcg@150	0.7725	0.6489	0.6112	0.7357	0.8255	0.811	0.5443
cosine_ndcg@200	0.7827	0.6653	0.6309	0.7501	0.8262	0.8114	0.5443
cosine_mrr@1	0.6476	0.1135	0.2956	0.6796	0.7395	0.6927	0.1789
cosine_mrr@20	0.8	0.5536	0.5164	0.8217	0.8059	0.7767	0.4002
cosine_mrr@50	0.8	0.5536	0.5166	0.8217	0.8066	0.7774	0.4002
cosine_mrr@100	0.8	0.5536	0.5166	0.8217	0.8067	0.7775	0.4002
cosine_mrr@150	0.8	0.5536	0.5166	0.8217	0.8067	0.7775	0.4002
cosine_mrr@200	0.8	0.5536	0.5166	0.8217	0.8067	0.7775	0.4002
cosine_map@1	0.6476	0.1135	0.2956	0.6796	0.7395	0.6927	0.1789
cosine_map@20	0.5392	0.481	0.4222	0.5012	0.744	0.721	0.3272
cosine_map@50	0.5258	0.4304	0.3791	0.4813	0.7465	0.7238	0.3272
cosine_map@100	0.558	0.4335	0.3829	0.5105	0.7469	0.7242	0.3272
cosine_map@150	0.5666	0.4485	0.3981	0.5184	0.747	0.7243	0.3272
cosine_map@200	0.5695	0.4551	0.4056	0.5228	0.7471	0.7244	0.3272
cosine_map@500	0.5744	0.4677	0.4189	0.5277	0.7472	0.7244	0.3272

Training Details

Training Datasets

full_en

full_en

• Dataset: full_en
• Size: 28,880 training samples
• Columns: anchor and positive
• Approximate statistics based on the first 1000 samples:

  	anchor	positive
  type	string	string
  details	min: 3 tokens mean: 5.68 tokens max: 11 tokens	min: 3 tokens mean: 5.76 tokens max: 12 tokens

• Samples:

  anchor	positive
  air commodore	flight lieutenant
  command and control officer	flight officer
  air commodore	command and control officer

• Loss: GISTEmbedLoss with these parameters:

  {'guide': SentenceTransformer(
    (0): Transformer({'max_seq_length': 128, 'do_lower_case': False}) with Transformer model: BertModel 
    (1): Pooling({'word_embedding_dimension': 384, '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})
    (2): Normalize()
  ), 'temperature': 0.01, 'margin_strategy': 'absolute', 'margin': 0.0}
  

full_de

full_de

• Dataset: full_de
• Size: 23,023 training samples
• Columns: anchor and positive
• Approximate statistics based on the first 1000 samples:

  	anchor	positive
  type	string	string
  details	min: 3 tokens mean: 7.99 tokens max: 30 tokens	min: 3 tokens mean: 8.19 tokens max: 30 tokens

• Samples:

  anchor	positive
  Staffelkommandantin	Kommodore
  Luftwaffenoffizierin	Luftwaffenoffizier/Luftwaffenoffizierin
  Staffelkommandantin	Luftwaffenoffizierin

• Loss: GISTEmbedLoss with these parameters:

  {'guide': SentenceTransformer(
    (0): Transformer({'max_seq_length': 128, 'do_lower_case': False}) with Transformer model: BertModel 
    (1): Pooling({'word_embedding_dimension': 384, '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})
    (2): Normalize()
  ), 'temperature': 0.01, 'margin_strategy': 'absolute', 'margin': 0.0}
  

full_es

full_es

• Dataset: full_es
• Size: 20,724 training samples
• Columns: anchor and positive
• Approximate statistics based on the first 1000 samples:

  	anchor	positive
  type	string	string
  details	min: 3 tokens mean: 9.13 tokens max: 32 tokens	min: 3 tokens mean: 8.84 tokens max: 32 tokens

• Samples:

  anchor	positive
  jefe de escuadrón	instructor
  comandante de aeronave	instructor de simulador
  instructor	oficial del Ejército del Aire

• Loss: GISTEmbedLoss with these parameters:

  {'guide': SentenceTransformer(
    (0): Transformer({'max_seq_length': 128, 'do_lower_case': False}) with Transformer model: BertModel 
    (1): Pooling({'word_embedding_dimension': 384, '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})
    (2): Normalize()
  ), 'temperature': 0.01, 'margin_strategy': 'absolute', 'margin': 0.0}
  

full_zh

full_zh

• Dataset: full_zh
• Size: 30,401 training samples
• Columns: anchor and positive
• Approximate statistics based on the first 1000 samples:

  	anchor	positive
  type	string	string
  details	min: 5 tokens mean: 7.15 tokens max: 14 tokens	min: 5 tokens mean: 7.46 tokens max: 21 tokens

• Samples:

  anchor	positive
  技术总监	技术和运营总监
  技术总监	技术主管
  技术总监	技术艺术总监

• Loss: GISTEmbedLoss with these parameters:

  {'guide': SentenceTransformer(
    (0): Transformer({'max_seq_length': 128, 'do_lower_case': False}) with Transformer model: BertModel 
    (1): Pooling({'word_embedding_dimension': 384, '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})
    (2): Normalize()
  ), 'temperature': 0.01, 'margin_strategy': 'absolute', 'margin': 0.0}
  

mix

mix

• Dataset: mix
• Size: 21,760 training samples
• Columns: anchor and positive
• Approximate statistics based on the first 1000 samples:

  	anchor	positive
  type	string	string
  details	min: 2 tokens mean: 6.71 tokens max: 19 tokens	min: 2 tokens mean: 7.69 tokens max: 19 tokens

• Samples:

  anchor	positive
  technical manager	Technischer Direktor für Bühne, Film und Fernsehen
  head of technical	directora técnica
  head of technical department	技术艺术总监

• Loss: GISTEmbedLoss with these parameters:

  {'guide': SentenceTransformer(
    (0): Transformer({'max_seq_length': 128, 'do_lower_case': False}) with Transformer model: BertModel 
    (1): Pooling({'word_embedding_dimension': 384, '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})
    (2): Normalize()
  ), 'temperature': 0.01, 'margin_strategy': 'absolute', 'margin': 0.0}
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 64
• per_device_eval_batch_size: 128
• gradient_accumulation_steps: 2
• num_train_epochs: 5
• warmup_ratio: 0.05
• log_on_each_node: False
• fp16: True
• dataloader_num_workers: 4
• ddp_find_unused_parameters: True
• batch_sampler: no_duplicates

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: 128
• per_gpu_train_batch_size: None
• per_gpu_eval_batch_size: None
• gradient_accumulation_steps: 2
• eval_accumulation_steps: None
• torch_empty_cache_steps: None
• learning_rate: 5e-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: 5
• max_steps: -1
• lr_scheduler_type: linear
• lr_scheduler_kwargs: {}
• warmup_ratio: 0.05
• warmup_steps: 0
• log_level: passive
• log_level_replica: warning
• log_on_each_node: False
• 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: 4
• 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}
• tp_size: 0
• 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: True
• 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
• 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: no_duplicates
• multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	Training Loss	full_en_cosine_ndcg@200	full_es_cosine_ndcg@200	full_de_cosine_ndcg@200	full_zh_cosine_ndcg@200	mix_es_cosine_ndcg@200	mix_de_cosine_ndcg@200	mix_zh_cosine_ndcg@200
-1	-1	-	0.6856	0.5207	0.4655	0.6713	0.6224	0.5604	0.5548
0.0010	1	5.3354	-	-	-	-	-	-	-
0.1027	100	2.665	-	-	-	-	-	-	-
0.2053	200	1.3375	0.7691	0.6530	0.6298	0.7517	0.7513	0.7393	0.5490
0.3080	300	1.1101	-	-	-	-	-	-	-
0.4107	400	0.9453	0.7802	0.6643	0.6246	0.7531	0.7610	0.7441	0.5493
0.5133	500	0.9202	-	-	-	-	-	-	-
0.6160	600	0.7887	0.7741	0.6549	0.6171	0.7542	0.7672	0.7540	0.5482
0.7187	700	0.7604	-	-	-	-	-	-	-
0.8214	800	0.7219	0.7846	0.6674	0.6244	0.7648	0.7741	0.7592	0.5497
0.9240	900	0.6965	-	-	-	-	-	-	-
1.0267	1000	0.6253	0.7646	0.6391	0.6122	0.7503	0.7825	0.7704	0.5463
1.1294	1100	0.4737	-	-	-	-	-	-	-
1.2320	1200	0.5055	0.7758	0.6582	0.6178	0.7514	0.7857	0.7764	0.5501
1.3347	1300	0.5042	-	-	-	-	-	-	-
1.4374	1400	0.5073	0.7613	0.6578	0.6178	0.7505	0.7829	0.7762	0.5452
1.5400	1500	0.4975	-	-	-	-	-	-	-
1.6427	1600	0.5242	0.7736	0.6673	0.6279	0.7555	0.7940	0.7859	0.5477
1.7454	1700	0.4713	-	-	-	-	-	-	-
1.8480	1800	0.4814	0.7845	0.6733	0.6285	0.7642	0.7992	0.7904	0.5449
1.9507	1900	0.4526	-	-	-	-	-	-	-
2.0544	2000	0.36	0.7790	0.6639	0.6252	0.7500	0.8032	0.7888	0.5499
2.1571	2100	0.3744	-	-	-	-	-	-	-
2.2598	2200	0.3031	0.7787	0.6614	0.6190	0.7537	0.7993	0.7811	0.5476
2.3624	2300	0.3638	-	-	-	-	-	-	-
2.4651	2400	0.358	0.7798	0.6615	0.6258	0.7497	0.8018	0.7828	0.5481
2.5678	2500	0.3247	-	-	-	-	-	-	-
2.6704	2600	0.3247	0.7854	0.6663	0.6248	0.7560	0.8081	0.7835	0.5452
2.7731	2700	0.3263	-	-	-	-	-	-	-
2.8758	2800	0.3212	0.7761	0.6681	0.6250	0.7517	0.8121	0.7927	0.5458
2.9784	2900	0.3291	-	-	-	-	-	-	-
3.0821	3000	0.2816	0.7727	0.6604	0.6163	0.7370	0.8163	0.7985	0.5473
3.1848	3100	0.2698	-	-	-	-	-	-	-
3.2875	3200	0.2657	0.7757	0.6615	0.6247	0.7417	0.8117	0.8004	0.5436
3.3901	3300	0.2724	-	-	-	-	-	-	-
3.4928	3400	0.2584	0.7850	0.6583	0.6320	0.7458	0.8120	0.7980	0.5454
3.5955	3500	0.2573	-	-	-	-	-	-	-
3.6982	3600	0.2744	0.7796	0.6552	0.6237	0.7409	0.8193	0.8018	0.5466
3.8008	3700	0.3054	-	-	-	-	-	-	-
3.9035	3800	0.2727	0.7825	0.6642	0.6293	0.7504	0.8213	0.8058	0.5463
4.0062	3900	0.2353	-	-	-	-	-	-	-
4.1088	4000	0.2353	0.7747	0.6628	0.6263	0.7384	0.8239	0.8065	0.5447
4.2115	4100	0.2385	-	-	-	-	-	-	-
4.3142	4200	0.231	0.7811	0.6608	0.6254	0.7463	0.8226	0.8051	0.5442
4.4168	4300	0.2115	-	-	-	-	-	-	-
4.5195	4400	0.2151	0.7815	0.6634	0.6301	0.7489	0.8251	0.8101	0.5450
4.6222	4500	0.2496	-	-	-	-	-	-	-
4.7248	4600	0.2146	0.7814	0.6654	0.6294	0.7523	0.8258	0.8104	0.5436
4.8275	4700	0.2535	-	-	-	-	-	-	-
4.9302	4800	0.2058	0.7827	0.6653	0.6309	0.7501	0.8262	0.8114	0.5443

Framework Versions

• Python: 3.11.11
• Sentence Transformers: 4.1.0
• Transformers: 4.51.2
• PyTorch: 2.6.0+cu124
• Accelerate: 1.6.0
• Datasets: 3.5.0
• Tokenizers: 0.21.1

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",
}

GISTEmbedLoss

@misc{solatorio2024gistembed,
    title={GISTEmbed: Guided In-sample Selection of Training Negatives for Text Embedding Fine-tuning},
    author={Aivin V. Solatorio},
    year={2024},
    eprint={2402.16829},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}