File size: 46,644 Bytes
fb3490a 7440eb8 3599ed0 7440eb8 3599ed0 befc207 7440eb8 9787a46 a1b0563 9787a46 7440eb8 570b218 3599ed0 8d8b41a 31ac810 8d8b41a 31ac810 8d8b41a 31ac810 3599ed0 31ac810 2994678 31ac810 e019293 31ac810 2994678 31ac810 2994678 31ac810 64ceef5 15f9fe7 64ceef5 e019293 7440eb8 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 |
---
license: eupl-1.2
language:
- de
- en
tags:
- sentence-transformers
- sentence-similarity
- feature-extraction
- generated_from_trainer
- dataset_size:16753490
- loss:MatryoshkaLoss
- loss:MultipleNegativesRankingLoss
datasets:
- avemio/German-RAG-EMBEDDING-TRIPLES-HESSIAN-AI
- MarcGrumpyOlejak/germanrag-scored
- MarcGrumpyOlejak/ultradistil-intel-orca-dpo-de-scored
- Short-Answer-Feedback/saf_legal_domain_german
- jfeil/GermanDefinitionGeneration-Distillation
- google/wmt24pp
- jphme/synthia_german_experimental
- google-research-datasets/paws-x
- jinaai/parallel-sentences
- Polyglot-or-Not/Fact-Completion
pipeline_tag: sentence-similarity
library_name: sentence-transformers
---
# A static embedding model tokenized with dbmdz/bert-base-german-uncased and mainly built on DE/EN-datasets as a base for further experiments.
This is a [sentence-transformers](https://www.SBERT.net) model trained on 74 datasets (full list at the bottom). It maps sentences & paragraphs to a 2048-dimensional dense vector space and can be used for semantic textual similarity, semantic search, paraphrase mining, text classification, clustering, and more.
Further explanations of how to build such a model, you can find in the [Static Embeddings blogpost](https://huggingface.co/blog/static-embeddings) by [Tom Aarsen](https://huggingface.co/tomaarsen) in January 2025. It took me until the end of May to start this tiny spare time experiment.
After some tests with different tokenizers I decided to pick one of the oldest as it has performed best by delivering the smallest size (~240MB) – [bert-base-german-uncased by the dbmdz-team](https://huggingface.co/dbmdz/bert-base-german-uncased).
* **99% performance:** Unexpectedly this model scored nearly 99% in comparison to [e5-base-sts-en-de](https://huggingface.co/danielheinz/e5-base-sts-en-de) during the [GermanGovServiceRetrieval](https://huggingface.co/datasets/mteb/GermanGovServiceRetrieval)-Task in MTEB by taking only a 80th of the time (40.3 seconds vs. 0.49).
* **Matryoshka:** This model was trained with a [Matryoshka loss](https://huggingface.co/blog/matryoshka), allowing you to truncate the embeddings for faster retrieval at minimal performance costs.
* **Evaluations:** See [Evaluations](#evaluation) for details on performance on German MTEB, special [GermanGovService retrieval](https://huggingface.co/datasets/mteb/GermanGovServiceRetrieval), embedding speed, and Matryoshka dimensionality truncation.
* **Training Script:** See [train_base.py](train_base.py) for the training script used to train this model from scratch (be warned - it is wildly commented).
## Model Details
### Model Description
- **Model Type:** Sentence Transformer
<!-- - **Base model:** [Unknown](https://huggingface.co/unknown) -->
- **Maximum Sequence Length:** inf tokens
- **Output Dimensionality:** 2048 dimensions
- **Similarity Function:** Cosine Similarity
- **Training Datasets:**
- [mmarco](https://huggingface.co/datasets/unicamp-dl/mmarco) - german only, filtered and with 3, 2, 1 hard negatives and (leave no sentence behind) the rest as 0 negatives:
- mmarco_3hn
- mmarco_2hn
- mmarco_1hn
- mmarco_0hn
- [deutsche-telekom/wikipedia-22-12-de-dpr](https://huggingface.co/datasets/deutsche-telekom/wikipedia-22-12-de-dpr) - deduplicated and recombined all different wordings of direct and indirect sentences. Built hard negatives but in the end reversed from hard negatives as it did not really work out.
- wp-22-12-de
- [nthakur/swim-ir-monolingual](https://huggingface.co/datasets/nthakur/swim-ir-monolingual) - german only, deduplicated and different combinations with and without 3 hard negatives.
- swim_ir_de
- swim_ir_de_3hn
- swim_ir_de_title_3hn
- swim_ir_de_title
- [avemio_triples](https://huggingface.co/datasets/avemio/German-RAG-EMBEDDING-TRIPLES-HESSIAN-AI) - thanks to Avemio for this release.
- [avemio_pairs](https://huggingface.co/datasets/avemio/German-RAG-EMBEDDING-PAIRS-HESSIAN-AI) - no pairs anymore - now with 3 hard negatives per pair and the rest as pairs:
- avemio_pairs_3hn
- avemio_pairs_0hn
- [oliverguhr/natural-questions-german](https://huggingface.co/datasets/oliverguhr/natural-questions-german) - combined english and german sentences with 3 and 1 hard negatives.
- nq_german_en_de_a_3hn
- nq_german_en_de_3hn
- nq_german_3hn
- nq_german_1hn
- [AgentWaller/german-oasst1-qa-format](https://huggingface.co/datasets/AgentWaller/german-oasst1-qa-format) - rebuilt with 3 hard negatives
- german_oasst1_hn
- [germanrag_short](https://huggingface.co/datasets/MarcGrumpyOlejak/germanrag-scored)
- [jphme/slimorca_dedup_german_experimental](https://huggingface.co/datasets/jphme/slimorca_dedup_german_experimental) - after scoring and filtering mined as many hard duplicates as possible and leave no sentence behind:
- slimorca_dedup_3hn
- slimorca_dedup_2hn
- slimorca_dedup_1hn
- slimorca_dedup_0hn
- [CausalLM/GPT-4-Self-Instruct-German](https://huggingface.co/datasets/CausalLM/GPT-4-Self-Instruct-German) - after scoring and filtering mined 3 hard negatives:
- german_gpt4_3hn
- [german_orca_dpo](https://huggingface.co/datasets/MarcGrumpyOlejak/ultradistil-intel-orca-dpo-de-scored)
- [mayflowergmbh/alpaca-gpt4_de](https://huggingface.co/datasets/mayflowergmbh/alpaca-gpt4_de) - after scoring and filtering mined 3 hard negatives and left no sentence behind (0hn):
- alpaca_gpt4_3hn
- alpaca_gpt4_0hn
- [argilla/databricks-dolly-15k-curated-multilingual](https://huggingface.co/datasets/argilla/databricks-dolly-15k-curated-multilingual) - after scoring and filtering mined 3 hard negatives and left no sentence behind (0hn) - but sometimes only 1 or 2 sentences were left:
- dolly_context_de_3hn
- dolly_context_ende_3hn
- dolly_instructions_de_3hn
- dolly_instructions_de_0hn
- dolly_instructions_ende_3hn
- dolly_responses_de_3hn
- dolly_responses_de_0hn
- dolly_responses_ende_3hn
- [saf_legal_de](https://huggingface.co/datasets/Short-Answer-Feedback/saf_legal_domain_german)
- [lavis-nlp/german_legal_sentences](https://huggingface.co/datasets/lavis-nlp/german_legal_sentences) - mined 3 hard negatives and left no sentences behind (0hn). Almost noone uses this dataset but for german law stuff, it's very helpful.
- gls_3hn
- gls_2hn
- gls_1hn
- gls_0hn
- [sentence-transformers/parallel-sentences-europarl](https://huggingface.co/datasets/sentence-transformers/parallel-sentences-europarl) - after scoring, filtering and mining 3 hard negatives the results were much better - there are many "bad" trnaslations and even empty fields:
- europarl_3hn
- europarl_0hn
- [sentence-transformers/parallel-sentences-tatoeba](https://huggingface.co/datasets/sentence-transformers/parallel-sentences-tatoeba) - mined 3 hard negatives and left no sentences behind (0hn):
- tatoeba_3hn
- tatoeba_0hn
- [sentence-transformers/parallel-sentences-wikimatrix](https://huggingface.co/datasets/sentence-transformers/parallel-sentences-wikimatrix) - mined 3 hard negatives but did not use the leftover sentence-pairs due low scores:
- wikimatrix_3hn
- [laion/Wikipedia-Abstract](https://huggingface.co/datasets/laion/Wikipedia-Abstract) - mined 3 hard negatives and left no sentences behind (0hn):
- wikipedia_abstract_3hn
- wikipedia_abstract_0hn
- [jfeil/GermanDefinitionGeneration-Distillation](https://huggingface.co/jfeil/GermanDefinitionGeneration-Distillation) - built multiple combinations for classifications of long to short, mined 3 hard negatives. Also built a short-word list without hard negatives:
- wiktionary_gdg_de_3hn
- wiktionary_gdg_de_short
- [wmt24pp](https://huggingface.co/datasets/google/wmt24pp) - filtered - not sure, if this is not in one of the benchmark datasets?
- [synthia_de](https://huggingface.co/datasets/jphme/synthia_german_experimental) - filtered for "scores".
- [deutsche-telekom/ger-backtrans-paraphrase](https://huggingface.co/datasets/deutsche-telekom/ger-backtrans-paraphrase) - combined german/english sentences, filtered and mined 3 hard negatives.
- gbp_3hn
- gbp_ende_3hn
- [PhilipMay/stsb_multi_mt](https://huggingface.co/datasets/PhilipMay/stsb_multi_mt) - mined 3 hard negatives per each language version (german/english):
- stbs_de_3hn
- stbs_en_3hn
- [google-research-datasets/paws-x](https://huggingface.co/datasets/google-research-datasets/paws-x)
- pawsx_de
- pawsx_en
- [MoritzLaurer/multilingual-NLI-26lang-2mil7](https://huggingface.co/datasets/MoritzLaurer/multilingual-NLI-26lang-2mil7) - with max. 3 hard negatives (german only)
- nli_anli_entail_3hn
- nli_fever_entail_3hn
- nli_ling_entail_3hn
- nli_mnli_entail_3hn
- nli_wanli_entail_3hn
- nli_anli_transl_3hn
- nli_fever_transl_3hn
- nli_ling_transl_3hn
- nli_mnli_transl_3hn
- nli_wanli_transl_3hn
- [jinaai/parallel-sentences](https://huggingface.co/datasets/jinaai/parallel-sentences) - with max. 3 hard negatives (with 3 german/english combinations)
- jina_ai_3en
- jina_ai_ende
- jina_ai_dede
- [Polyglot-or-Not/Fact-Completion](https://huggingface.co/datasets/Polyglot-or-Not/Fact-Completion)
- polyglot_de
- polyglot_en
- [Tilde Model - EESC](https://tilde-model.s3-eu-west-1.amazonaws.com/Tilde_MODEL_Corpus.html) - an almost forgotten corpus from document texts of European Economic and Social Committee document portal.
- [miracl/miracl-corpus](https://huggingface.co/datasets/miracl/miracl-corpus) - scored and filtered ('cos_sim_sts_de' > 0.5 and 'cos_sim_sts_de' < 0.85 and 'text_unique_tokens_de' > 6). Mined 3 hard negatives and left no sentence behind.
- miracl_de_3hn
- miracl_de_0hn
- **Languages:** de, en
- **License:** eupl-1.2
### Model Sources
- **Documentation:** [Sentence Transformers Documentation](https://sbert.net)
- **Repository:** [Sentence Transformers on GitHub](https://github.com/UKPLab/sentence-transformers)
- **Hugging Face:** [Sentence Transformers on Hugging Face](https://huggingface.co/models?library=sentence-transformers)
### Full Model Architecture
```
SentenceTransformer(
(0): StaticEmbedding(
(embedding): EmbeddingBag(31102, 2048, mode='mean')
)
)
```
## Usage
### Direct Usage (Sentence Transformers)
First install the Sentence Transformers library:
```bash
pip install -U sentence-transformers
```
Then you can load this model and run inference.
```python
from sentence_transformers import SentenceTransformer
# Download from the 🤗 Hub
model = SentenceTransformer("MarcGrumpyOlejak/sts-mrl-en-de-base-v1")
# Run inference
queries = [
"Im April 1928 beschrieb er in seinem Artikel On the Construction of Tables by Interpolation die Verwendung von Lochkartenger\u00e4ten zum Interpolieren von Datentabellen und verglich dies mit den weniger effizienten und fehleranf\u00e4lligeren Methoden mit mechanischen Ger\u00e4ten wie den Windradrechnern unter dem Markennamen Brunsviga.",
]
documents = [
'Im April 1928 beschrieb er in seinem Artikel „On the Construction of Tables by Interpolation“ („Über die Erstellung von Tabellen durch Interpolation“) die Interpolation von Daten in Tabellen mit Hilfe von Lochkarten und verglich diese Methode mit dem uneffizienteren und fehleranfälligeren Verfahren, das mechanische Rechner verwendet.',
'POLES liefert nicht die direkten makro-ökonomischen Auswirkungen der Minderungsmaßnahmen wie im Stern-Report vorgesehen, erlaubt jedoch eine detaillierte Abschätzung der Kosten im Zusammenhang mit Techniken mit wenig Energieverbrauch oder Nullenergietechniken.',
'Im Lehrbuch Maschinenelemente – Funktion, Gestaltung und Berechnung von Decker (bisher 19 Auflagen) wird anhand praktischer Anwendungen mit Z88 die Berechnung von Maschinenelementen mit der Finiten-Elemente-Analyse gelehrt.',
]
query_embeddings = model.encode_query(queries)
document_embeddings = model.encode_document(documents)
print(query_embeddings.shape, document_embeddings.shape)
# [1, 2048] [3, 2048]
# Get the similarity scores for the embeddings
similarities = model.similarity(query_embeddings, document_embeddings)
print(similarities)
# tensor([[0.7737, 0.1275, 0.1184]])
```
### Matryoshka Usage (Sentence Transformers)
This model was trained with Matryoshka loss, allowing this model to be used with lower dimensionalities with minimal performance loss. Notably, a lower dimensionality allows for much faster downstream tasks, such as clustering or classification. You can specify a lower dimensionality with the truncate_dim argument when initializing the Sentence Transformer model:
```python
from sentence_transformers import SentenceTransformer
model = SentenceTransformer("MarcGrumpyOlejak/sts-mrl-en-de-base-v1", truncate_dim=256)
embeddings = model.encode([
"Welches Kabel wird benötigt, um die PlayStation mit einer externen Festplatte zu verbinden?",
"Pepsi wurde in der Stadt ''New Bern'' in North Carolina erfunden."
])
print(embeddings.shape)
# => (2, 256)
```
### Out-of-Scope Use
After several tests it is not really good in reranking. Also everything about "news" is really low due the fact, that there is no open licensed and commercially usable dataset available.
Maybe you have knowledge about an official free and open licensed news-based dataset. Feel free to contact me.
## Evaluation
All steps and evaluations have been made locally on my very small hardware using a Nvidia RTX 2070 SUPER (8 GB).
This model has been benchmarked using mainly the [GermanGovServiceRetrieval](https://huggingface.co/datasets/mteb/GermanGovServiceRetrieval) task, developed by the [Open Source team in the Munich city administration](https://huggingface.co/it-at-m). It associates questions with a textual context containing the answer. The idea is to train it further on based upon german based administraional classification datasets. After the first results the full german MTEB(deu, v1) has also been tested as the GermanGovServiceRetrieval test is not a part of the german MTEB-benchmark. While testing with NanoBEIR it has been shown to be a bit insufficient for testing bilingual german/english - but I accidentally outscored [static-similarity-mrl-multilingual-v1](https://huggingface.co/sentence-transformers/static-similarity-mrl-multilingual-v1) by 0.03 points ;)
As for the static embeddings being built using [Model2Vec](https://github.com/MinishLab/model2vec), with [alikia2x/jina-embedding-v3-m2v-1024](https://huggingface.co/alikia2x/jina-embedding-v3-m2v-1024) I picked the largest one I could find with ~1GB size.
This model is compared against the excellent [e5-base-sts-en-de](https://huggingface.co/danielheinz/e5-base-sts-en-de) model made by [Daniel Heinz](https://huggingface.co/danielheinz) back in 2024 (ca. 1.1GB). The second model for comparisons with dense embeddings is the optimized [granite-embedding-107m-multilingual](https://huggingface.co/ibm-granite/granite-embedding-107m-multilingual) model made by the [IBM-Granite](https://huggingface.co/ibm-granite)-team (ca. 770MB).
### Benchmark details
Oops - I forgot to NanoBEIR [granite-embedding-107m-multilingual](https://huggingface.co/ibm-granite/granite-embedding-107m-multilingual) - that's for the week-end.
| | **NanoBEIR** | **MTEB** | **MTEB(deu, v1) – avg** |
| ------------------------------------- | ---------------------------- | ------------------------- | ----------------------- |
| **Dense Embeddings** | NanoBEIR_mean_cosine_ndcg@10 | GermanGovServiceRetrieval | Naive (sum/num) |
| [e5-base-sts-en-de](https://huggingface.co/danielheinz/e5-base-sts-en-de) | _0,5320_ | _0,7931_ | _0,5194_ |
| [granite-embedding-107m-multilingual](https://huggingface.co/ibm-granite/granite-embedding-107m-multilingual) | _<br>_ | _0,7880_ | _0,4992_ |
| **Static Embeddings** | _<br>_ | _<br>_ | _<br>_ |
| [static-retrieval-mrl-en-v1](https://huggingface.co/sentence-transformers/static-retrieval-mrl-en-v1)(*) | _0,5035_ | _0,6630_ | _0,3716_ |
| [jina-embedding-v3-m2v-1024](https://huggingface.co/alikia2x/jina-embedding-v3-m2v-1024) | 0,3480 | 0,7260 | 0,4081 |
| [static-similarity-mrl-multilingual-v1](https://huggingface.co/sentence-transformers/static-similarity-mrl-multilingual-v1) | 0,4350 | 0,7281 | 0,4259 |
| **sts-mrl-en-de-base-v1** | **0,4680** | **0,7841** | **0,4566** |
_((*)'static-retrieval-mrl-en-v1' only for comparison to mainly english based NanoBEIR)_
#### MTEB - GermanGovServiceRetrieval Evaluation
As [e5-base-sts-en-de](https://huggingface.co/danielheinz/e5-base-sts-en-de) scores with 0.7931 in the [GermanGovServiceRetrieval](https://huggingface.co/datasets/mteb/GermanGovServiceRetrieval) task, that means [sts-mrl-en-de-base-v1](MarcGrumpyOlejak/sts-mrl-en-de-base-v1) with 0.7841 achieves **98.865%** for the same task by using only ~230MB RAM and a CPU.
So it is only **0,4949%** behind [granite-embedding-107m-multilingual](https://huggingface.co/ibm-granite/granite-embedding-107m-multilingual).
#### MTEB(deu, v1) – avg
For the german version of the MTEB benchmark MTEB(deu, v1) the results are not as significant as the GermanGovServiceRetrieval task - but with **87,909%** of quality in comparison to [e5-base-sts-en-de](https://huggingface.co/danielheinz/e5-base-sts-en-de) you can use [sts-mrl-en-de-base-v1](MarcGrumpyOlejak/sts-mrl-en-de-base-v1) for example to [mine hard negatives](https://sbert.net/docs/package_reference/util.html#sentence_transformers.util.mine_hard_negatives) in a really short time instead of burning money with a whole bunch of GPU.
Even with the really well speed optimised [granite-embedding-107m-multilingual](https://huggingface.co/ibm-granite/granite-embedding-107m-multilingual) being almost as fast as the static embeddings, you'll still need a GPU.
#### Matryoshka Evaluation
_(have to be checked twice - looks like almost everyone has a glitch in the results … the results are better with a first reduction from 2048 down to 1024 dimensions? That's the 2nd thing for the week-end.)_
### Training Datasets
Sadly all details of the datasets had to be saved in a seperate file [details_datasets.md](details_datasets.md) as this README.md has a limit.
### Training Hyperparameters
#### Non-Default Hyperparameters
- `eval_strategy`: steps
- `per_device_train_batch_size`: 4096
- `per_device_eval_batch_size`: 4096
- `learning_rate`: 0.2
- `num_train_epochs`: 1
- `lr_scheduler_type`: cosine
- `warmup_ratio`: 0.1
- `fp16`: True
- `batch_sampler`: no_duplicates
#### All Hyperparameters
<details><summary>Click to expand</summary>
- `overwrite_output_dir`: False
- `do_predict`: False
- `eval_strategy`: steps
- `prediction_loss_only`: True
- `per_device_train_batch_size`: 4096
- `per_device_eval_batch_size`: 4096
- `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`: 0.2
- `weight_decay`: 0.0
- `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`: cosine
- `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`: 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`: 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`: 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
- `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
- `router_mapping`: {}
- `learning_rate_mapping`: {}
</details>
### Training Logs
| Epoch | Step | Training Loss | mmarco 3hn loss | mmarco 2hn loss | mmarco 1hn loss | mmarco 0hn loss | wp-22-12-de loss | swim ir de loss | swim ir de 3hn loss | swim ir de title 3hn loss | swim ir de title loss | avemio triples loss | avemio pairs 3hn loss | avemio pairs 0hn loss | nq german en de a 3hn loss | nq german en de 3hn loss | nq german 3hn loss | nq german 1hn loss | german oasst1 hn loss | germanrag short loss | slimorca dedup 3hn loss | slimorca dedup 2hn loss | slimorca dedup 1hn loss | slimorca dedup 0hn loss | german gpt4 3hn loss | german orca dpo loss | alpaca gpt4 3hn loss | alpaca gpt4 0hn loss | dolly context de 3hn loss | dolly context ende 3hn loss | dolly instructions de 3hn loss | dolly instructions de 0hn loss | dolly instructions ende 3hn loss | dolly responses de 3hn loss | dolly responses de 0hn loss | dolly responses ende 3hn loss | saf legal de loss | gls 3hn loss | gls 2hn loss | gls 1hn loss | gls 0hn loss | europarl 3hn loss | europarl 0hn loss | tatoeba 3hn loss | tatoeba 0hn loss | wikimatrix 3hn loss | wikipedia abstract 3hn loss | wikipedia abstract 0hn loss | wiktionary gdg de 3hn loss | wiktionary gdg de short loss | wmt24pp loss | synthia de loss | gbp 3hn loss | gbp ende 3hn loss | stbs de 3hn loss | stbs en 3hn loss | pawsx de loss | pawsx en loss | nli anli entail 3hn loss | nli fever entail 3hn loss | nli ling entail 3hn loss | nli mnli entail 3hn loss | nli wanli entail 3hn loss | nli anli transl 3hn loss | nli fever transl 3hn loss | nli ling transl 3hn loss | nli mnli transl 3hn loss | nli wanli transl 3hn loss | jina ai 3en loss | jina ai ende loss | jina ai dede loss | polyglot de loss | polyglot en loss | tilde EESC loss | miracl de 3hn loss | miracl de 0hn loss |
|:------:|:----:|:-------------:|:---------------:|:---------------:|:---------------:|:---------------:|:----------------:|:---------------:|:-------------------:|:-------------------------:|:---------------------:|:-------------------:|:---------------------:|:---------------------:|:--------------------------:|:------------------------:|:------------------:|:------------------:|:---------------------:|:--------------------:|:-----------------------:|:-----------------------:|:-----------------------:|:-----------------------:|:--------------------:|:--------------------:|:--------------------:|:--------------------:|:-------------------------:|:---------------------------:|:------------------------------:|:------------------------------:|:--------------------------------:|:---------------------------:|:---------------------------:|:-----------------------------:|:-----------------:|:------------:|:------------:|:------------:|:------------:|:-----------------:|:-----------------:|:----------------:|:----------------:|:-------------------:|:---------------------------:|:---------------------------:|:--------------------------:|:----------------------------:|:------------:|:---------------:|:------------:|:-----------------:|:----------------:|:----------------:|:-------------:|:-------------:|:------------------------:|:-------------------------:|:------------------------:|:------------------------:|:-------------------------:|:------------------------:|:-------------------------:|:------------------------:|:------------------------:|:-------------------------:|:----------------:|:-----------------:|:-----------------:|:----------------:|:----------------:|:---------------:|:------------------:|:------------------:|
| 0.0002 | 1 | 32.2328 | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - |
| 0.1211 | 500 | 17.4935 | 5.9441 | 18.6286 | 15.4380 | 21.7452 | 11.5899 | 15.7739 | 2.0470 | 6.4545 | 28.7021 | 3.4327 | 3.0953 | 21.4473 | 0.6579 | 2.3081 | 8.3028 | 17.1118 | 8.6341 | 0.7353 | 2.4550 | 10.1110 | 20.2165 | 15.1944 | 11.2822 | 14.0772 | 9.3205 | 12.3671 | 3.6399 | 0.2185 | 2.5653 | 27.0853 | 2.1334 | 4.3423 | 2.3262 | 0.5350 | 20.1312 | 6.6543 | 11.5668 | 11.2751 | 15.0010 | 2.5165 | 46.8575 | 6.4837 | 17.8191 | 0.9617 | 7.7542 | 3.3035 | 17.7944 | 4.9850 | 0.5039 | 6.9794 | 0.4971 | 1.7211 | 7.5595 | 5.8076 | 2.1527 | 0.4983 | 9.9586 | 7.6724 | 4.5647 | 4.4193 | 4.3135 | 0.8089 | 2.2057 | 0.8494 | 1.5787 | 2.4122 | 9.0588 | 1.6716 | 5.7378 | 17.4829 | 17.4252 | 2.7128 | 2.3019 | 4.9855 |
| 0.2421 | 1000 | 9.8434 | 5.9548 | 16.1939 | 13.6828 | 19.8400 | 10.3624 | 13.5662 | 1.7398 | 4.7552 | 26.9780 | 2.7763 | 2.6297 | 19.2160 | 0.6367 | 2.2657 | 8.1566 | 15.8885 | 7.0793 | 0.7799 | 1.6238 | 9.2113 | 18.4966 | 14.8541 | 11.6090 | 17.8812 | 7.3860 | 7.7746 | 3.2721 | 0.1734 | 2.2635 | 27.3627 | 1.7248 | 4.0169 | 3.3867 | 0.4930 | 19.1067 | 7.0229 | 13.2283 | 13.9238 | 17.1221 | 2.0835 | 47.1417 | 6.2599 | 15.3082 | 0.7972 | 6.9853 | 2.7917 | 15.1196 | 4.4008 | 0.1748 | 6.5392 | 0.4433 | 1.3500 | 7.5248 | 5.8447 | 2.1663 | 0.4949 | 9.3473 | 6.2105 | 4.2394 | 4.1746 | 4.2383 | 0.6806 | 2.1903 | 0.6338 | 1.3037 | 2.0331 | 5.0726 | 1.0650 | 5.8712 | 17.3595 | 16.0869 | 1.9498 | 2.1635 | 4.1986 |
| 0.3632 | 1500 | 9.4195 | 6.0462 | 15.3733 | 13.4579 | 19.1822 | 10.1358 | 13.7938 | 2.0818 | 4.3716 | 26.0843 | 2.7380 | 2.6063 | 18.9278 | 0.6317 | 2.1179 | 8.5954 | 15.0949 | 6.2069 | 0.8866 | 1.5936 | 9.0869 | 18.6605 | 14.5752 | 12.3640 | 15.1111 | 7.5786 | 8.6830 | 2.9134 | 0.1539 | 2.3901 | 24.0635 | 1.5851 | 3.0859 | 2.8681 | 0.4823 | 20.1934 | 6.9440 | 11.9040 | 11.6429 | 13.5179 | 1.9956 | 46.0385 | 6.0581 | 15.7130 | 0.7430 | 6.2928 | 2.9993 | 14.2742 | 4.1868 | 0.1639 | 5.8340 | 0.4744 | 1.3372 | 7.7122 | 5.6745 | 2.1703 | 0.4930 | 9.6020 | 6.0473 | 3.5016 | 3.7158 | 4.2441 | 0.5784 | 2.1883 | 0.5912 | 1.2164 | 1.9767 | 7.0197 | 1.0216 | 4.4556 | 14.8992 | 15.8563 | 1.8581 | 2.1515 | 4.4043 |
| 0.4843 | 2000 | 8.2114 | 5.8039 | 14.9131 | 12.9781 | 18.3934 | 9.9055 | 13.5402 | 2.0944 | 4.4961 | 26.2583 | 2.6002 | 2.5542 | 18.3124 | 0.5504 | 1.7278 | 8.4266 | 12.8837 | 5.5970 | 0.7967 | 1.5002 | 8.8843 | 18.2636 | 15.5366 | 12.1376 | 13.7508 | 6.1530 | 6.6779 | 2.2906 | 0.1435 | 1.8996 | 21.9520 | 1.5331 | 2.7177 | 3.0663 | 0.4214 | 19.7372 | 6.1346 | 10.9578 | 10.5089 | 13.6577 | 1.8838 | 46.2217 | 4.1247 | 12.9807 | 0.6397 | 6.3777 | 2.5970 | 13.7871 | 4.1784 | 0.1893 | 4.4490 | 0.4018 | 1.1374 | 7.1980 | 5.6566 | 2.1517 | 0.4921 | 9.2049 | 6.0599 | 3.4091 | 3.6662 | 4.0776 | 0.4841 | 2.0716 | 0.4860 | 0.9970 | 1.7709 | 7.5693 | 0.6321 | 4.9397 | 14.5334 | 15.4385 | 1.7821 | 1.9614 | 4.2582 |
| 0.6053 | 2500 | 8.038 | 5.5500 | 14.8000 | 12.8634 | 18.2342 | 9.7964 | 13.2195 | 1.9088 | 4.2172 | 25.7571 | 2.4768 | 2.4510 | 17.9053 | 0.4689 | 1.8237 | 8.1981 | 12.5957 | 6.0768 | 0.6939 | 1.5240 | 9.6936 | 18.5641 | 16.5833 | 12.5368 | 13.6839 | 6.6175 | 7.2916 | 2.3097 | 0.1377 | 1.9064 | 22.0331 | 1.5278 | 2.5185 | 4.8549 | 0.3997 | 20.1505 | 6.0001 | 10.3536 | 9.9127 | 12.7608 | 1.7728 | 46.1264 | 3.4876 | 13.2839 | 0.6246 | 6.0571 | 2.5264 | 13.6899 | 4.1796 | 0.1133 | 5.5862 | 0.3973 | 1.1315 | 7.0625 | 5.7281 | 2.1597 | 0.4939 | 9.3306 | 5.8505 | 3.0920 | 3.6364 | 4.2557 | 0.4513 | 1.9419 | 0.4341 | 0.7909 | 1.6440 | 7.5517 | 0.6997 | 4.9564 | 14.5145 | 15.7047 | 1.6838 | 1.9027 | 4.2791 |
| 0.7264 | 3000 | 8.4735 | 5.4690 | 14.0184 | 12.4418 | 17.2256 | 9.5584 | 12.8587 | 1.8026 | 4.2292 | 25.0699 | 2.4180 | 2.3386 | 17.5121 | 0.4924 | 1.7512 | 8.6264 | 12.9932 | 5.7242 | 0.7519 | 1.4209 | 8.7996 | 17.9024 | 15.0738 | 10.3888 | 12.8886 | 6.9268 | 7.5737 | 2.4082 | 0.1446 | 1.9202 | 22.0949 | 1.4499 | 2.7943 | 3.8219 | 0.4096 | 20.1391 | 5.9977 | 10.2577 | 9.9893 | 12.8969 | 1.8217 | 45.9583 | 3.6835 | 14.0661 | 0.6401 | 5.8992 | 2.4225 | 13.6148 | 4.0275 | 0.1058 | 4.2324 | 0.4046 | 1.1448 | 7.2012 | 5.7275 | 2.1669 | 0.4947 | 8.9883 | 5.8919 | 3.4086 | 3.5578 | 3.8109 | 0.4713 | 2.0382 | 0.4806 | 0.9071 | 1.7479 | 7.4633 | 0.6957 | 5.1938 | 14.2104 | 15.6664 | 1.7301 | 1.9228 | 4.1841 |
| 0.8475 | 3500 | 7.7352 | 5.3754 | 14.0426 | 12.5198 | 17.3227 | 9.4857 | 12.9446 | 1.8784 | 4.2447 | 25.1068 | 2.3991 | 2.3495 | 17.5300 | 0.4642 | 1.6235 | 8.4671 | 12.8252 | 5.3035 | 0.7126 | 1.4499 | 8.4552 | 16.9827 | 14.6279 | 10.8074 | 12.8392 | 6.5745 | 7.2679 | 2.4318 | 0.1319 | 1.8556 | 22.2088 | 1.3227 | 2.6365 | 4.3796 | 0.3783 | 20.1810 | 5.9464 | 10.2856 | 9.9382 | 12.6812 | 1.6933 | 46.2977 | 3.6286 | 13.8749 | 0.5844 | 5.8990 | 2.4661 | 13.3314 | 4.0382 | 0.1148 | 4.3655 | 0.4017 | 1.0360 | 7.1329 | 5.7121 | 2.1640 | 0.4945 | 8.9242 | 5.6470 | 3.2758 | 3.5739 | 4.0207 | 0.4303 | 1.9566 | 0.4515 | 0.8112 | 1.6914 | 7.4063 | 0.6659 | 5.2429 | 13.9946 | 15.6856 | 1.5650 | 1.8613 | 4.3350 |
| 0.9685 | 4000 | 7.4739 | 5.3820 | 13.9713 | 12.4551 | 17.2949 | 9.4687 | 12.9339 | 1.9303 | 4.2006 | 25.0763 | 2.3880 | 2.3362 | 17.4705 | 0.4638 | 1.6235 | 8.3594 | 12.6393 | 5.3609 | 0.7168 | 1.4452 | 8.3913 | 16.8145 | 14.9649 | 10.7862 | 12.5774 | 6.6076 | 7.1481 | 2.3770 | 0.1320 | 1.8618 | 22.2842 | 1.3191 | 2.6045 | 4.6015 | 0.3718 | 14.6598 | 5.9303 | 10.1947 | 9.8502 | 12.5003 | 1.6814 | 46.1385 | 3.6696 | 13.8947 | 0.5799 | 5.8546 | 2.4445 | 13.3022 | 4.0359 | 0.1090 | 4.4493 | 0.3932 | 1.0395 | 7.1369 | 5.6920 | 2.1641 | 0.4943 | 8.9089 | 5.6356 | 3.2438 | 3.5664 | 4.0016 | 0.4297 | 1.9810 | 0.4511 | 0.8123 | 1.6705 | 7.4795 | 0.6834 | 5.2668 | 13.9481 | 15.6508 | 1.5442 | 1.8556 | 4.3036 |
### Framework Versions
- Python: 3.10.15
- Sentence Transformers: 5.0.0
- Transformers: 4.51.3
- PyTorch: 2.1.0+cu121
- Accelerate: 1.3.0
- Datasets: 2.21.0
- Tokenizers: 0.21.1
## 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",
}
```
#### MatryoshkaLoss
```bibtex
@misc{kusupati2024matryoshka,
title={Matryoshka Representation Learning},
author={Aditya Kusupati and Gantavya Bhatt and Aniket Rege and Matthew Wallingford and Aditya Sinha and Vivek Ramanujan and William Howard-Snyder and Kaifeng Chen and Sham Kakade and Prateek Jain and Ali Farhadi},
year={2024},
eprint={2205.13147},
archivePrefix={arXiv},
primaryClass={cs.LG}
}
```
#### MultipleNegativesRankingLoss
```bibtex
@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}
}
```
#### GermanGovServiceRetrieval
```bibtex
@software{lhm-dienstleistungen-qa,
author = {Schröder, Leon Marius and
Gutknecht, Clemens and
Alkiddeh, Oubada and
Susanne Weiß,
Lukas, Leon},
month = nov,
publisher = {it@M},
title = {LHM-Dienstleistungen-QA - german public domain question-answering dataset},
url = {https://huggingface.co/datasets/it-at-m/LHM-Dienstleistungen-QA},
year = {2022},
}
```
#### MMTEB
```bibtex
@article{enevoldsen2025mmtebmassivemultilingualtext,
title={MMTEB: Massive Multilingual Text Embedding Benchmark},
author={Kenneth Enevoldsen and Isaac Chung and Imene Kerboua and Márton Kardos and Ashwin Mathur and David Stap and Jay Gala and Wissam Siblini and Dominik Krzemiński and Genta Indra Winata and Saba Sturua and Saiteja Utpala and Mathieu Ciancone and Marion Schaeffer and Gabriel Sequeira and Diganta Misra and Shreeya Dhakal and Jonathan Rystrøm and Roman Solomatin and Ömer Çağatan and Akash Kundu and Martin Bernstorff and Shitao Xiao and Akshita Sukhlecha and Bhavish Pahwa and Rafał Poświata and Kranthi Kiran GV and Shawon Ashraf and Daniel Auras and Björn Plüster and Jan Philipp Harries and Loïc Magne and Isabelle Mohr and Mariya Hendriksen and Dawei Zhu and Hippolyte Gisserot-Boukhlef and Tom Aarsen and Jan Kostkan and Konrad Wojtasik and Taemin Lee and Marek Šuppa and Crystina Zhang and Roberta Rocca and Mohammed Hamdy and Andrianos Michail and John Yang and Manuel Faysse and Aleksei Vatolin and Nandan Thakur and Manan Dey and Dipam Vasani and Pranjal Chitale and Simone Tedeschi and Nguyen Tai and Artem Snegirev and Michael Günther and Mengzhou Xia and Weijia Shi and Xing Han Lù and Jordan Clive and Gayatri Krishnakumar and Anna Maksimova and Silvan Wehrli and Maria Tikhonova and Henil Panchal and Aleksandr Abramov and Malte Ostendorff and Zheng Liu and Simon Clematide and Lester James Miranda and Alena Fenogenova and Guangyu Song and Ruqiya Bin Safi and Wen-Ding Li and Alessia Borghini and Federico Cassano and Hongjin Su and Jimmy Lin and Howard Yen and Lasse Hansen and Sara Hooker and Chenghao Xiao and Vaibhav Adlakha and Orion Weller and Siva Reddy and Niklas Muennighoff},
publisher = {arXiv},
journal={arXiv preprint arXiv:2502.13595},
year={2025},
url={https://arxiv.org/abs/2502.13595},
doi = {10.48550/arXiv.2502.13595},
}
```
#### MTEB
```bibtex
@article{muennighoff2022mteb,
author = {Muennighoff, Niklas and Tazi, Nouamane and Magne, Lo{\"\i}c and Reimers, Nils},
title = {MTEB: Massive Text Embedding Benchmark},
publisher = {arXiv},
journal={arXiv preprint arXiv:2210.07316},
year = {2022}
url = {https://arxiv.org/abs/2210.07316},
doi = {10.48550/ARXIV.2210.07316},
}
```
|