Thermostatic/qwen3-4b-embeddings-akkadian

SentenceTransformer

This model was finetuned with Unsloth.

based on unsloth/Qwen3-Embedding-4B

This is a sentence-transformers model finetuned from unsloth/Qwen3-Embedding-4B. It maps sentences & paragraphs to a 2560-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: unsloth/Qwen3-Embedding-4B
• Maximum Sequence Length: 1024 tokens
• Output Dimensionality: 2560 dimensions
• Similarity Function: Cosine Similarity

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': 1024, 'do_lower_case': False, 'architecture': 'PeftModelForFeatureExtraction'})
  (1): Pooling({'word_embedding_dimension': 2560, 'pooling_mode_cls_token': False, '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': True, '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("sentence_transformers_model_id")
# Run inference
sentences = [
    '37 ku-ta-ni 34 {túg}šu-ru-tum a-na lá-qé-pí-im áp-qí-id IGI i-dí-{d}IŠKUR IGI i-dí-a-šur DUMU sú-e-ta-ta 7 TÚG.HI.A ša li-wi-tim IGI a-šùr-SIPA a-dí-šu-um',
    "37 -textiles (and) 34 dark textiles I entrusted to Lā-qēpum in the presence of Iddin-Adad and of Iddin-Aššur, son of Suettata. 7 textiles for wrapping I gave him in the presence of Aššur-rē'ī.",
    'To Ešarra and Ab-šalim from Ennam-Aššur: 10 shekels of silver and an undergarment sealed by me is for Ešarra. 10 shekels of silver and 2 sashes are for Ab-šalim and the girl. 2 shekels of silver is for <big_gap> sister Ištar-lamassī <big_gap>',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 2560]

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings, embeddings)
print(similarities)
# tensor([[1.0000, 0.7541, 0.0110],
#         [0.7541, 1.0000, 0.0221],
#         [0.0110, 0.0221, 1.0000]])

Evaluation

Metrics

Semantic Similarity

• Dataset: akkadian_val
• Evaluated with EmbeddingSimilarityEvaluator

Metric	Value
pearson_cosine	0.9124
spearman_cosine	0.8625

Training Details

Training Dataset

Unnamed Dataset

• Size: 1,137 training samples
• Columns: anchor and positive
• Approximate statistics based on the first 1000 samples:

  	anchor	positive
  type	string	string
  details	min: 13 tokens mean: 229.79 tokens max: 579 tokens	min: 7 tokens mean: 137.38 tokens max: 442 tokens

• Samples:

  anchor	positive
  1 ma-na KÙ.BABBAR big_gap 4 {túg}ku-ta-ni big_gap ni-ik-na-x- big_gap KÙ.BABBAR a-ha-ma big_gap 2 GÍN KÙ.BABBAR big_gap ša áb-na-tim kà-ú-nam big_gap áb-na-tim big_gap uk-ta-in big_gap KÙ.BABBAR i-za-az big_gap ŠU.NÍGIN 1 ma-na 2 GÍN KÙ.BABBAR i li-bi big_gap IGI pì-lá-ah- big_gap IGI a-šur-na-da šu-ma KÙ.BABBAR a-na big_gap lá iš-ta-qá-al big_gap iš-tù ha-mu-uš-tim ša a-šur-be-el-a-wa-tim 1 ma-na-um 3 GÍN.TA ṣí-ib-tám ú-ṣa-áb i-na ITU.KAM a-ma-nu-šu-um	1 mina of silver 4 kutānu-textiles silver; further, 2 shekels of silver of the stones confirm He has confirmed the stones. The silver stands ready. In all: 1 mina 2 shekels of silver is owed by Witnessed by Pilah- , by Aššur-nādā. If he has not paid the silver in I shall count interest for him reckoned from the week of Aššur-bēl-awātim at the rate 3 shekels per mina per month.
  ŠU.NÍGIN KÙ.BABBAR-pì-kà 15 ma-na 10 GÍN lu ša AN.NA ú ṣú-ba-tí-kà ku-nu-ki-ni ṣí-li-a na-áš-a-ku-um	Total of your silver: 15 minas 10 shekels, Ṣilliya brings you under our seal - both that from the tin and that from your textiles.
  1 ma-na 7.5 GÍN KÙ.BABBAR ṣa-ru-pá-am i-ṣé-er a-mur-IŠTAR DUMU da-da e-la-ma i-šu iš-tù ha-muš-tim ša a-la-hi-im ú {d}MAR.TU-ba-ni a-na 11 ha-am-ša-tim i-ša-qal šu-ma lá iš-qú-ul 1½ GÍN.TA ṣí-ib-tám a-na ma-na-im i-na ITU.1.KAM ú-ṣa-áb ITU.KAM ša sà-ra-tim li-mu-um ša qá-té DINGIR-šu-GAL DUMU ba-zi-a IGI im-dí-lim DUMU šu-lá-ba-an IGI e-me-me-i DUMU a-zu-ta-a	1 mina 7.5 shekels of refined silver Āmur-Ištar, son of Dada, owes to Elamma. From the week of Ali-ahum and Amurrum-bāni he will pay in 11 weeks; if he does not pay he will add 1.5 shekel as interest per mina per month. Month II, eponymy of the successor of Ilšu-rabi, son of Baziya. In the presence of Imdī-ilum, son of Šu-Labān, of Ememe'i, son of Azutaya.

• Loss: MultipleNegativesRankingLoss with these parameters:

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

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 32
• learning_rate: 2e-05
• lr_scheduler_type: cosine
• warmup_ratio: 0.1
• bf16: True
• dataloader_pin_memory: False
• 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: 32
• 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: 2e-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: 3
• max_steps: -1
• lr_scheduler_type: cosine
• lr_scheduler_kwargs: None
• 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
• bf16: True
• 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: 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
• project: huggingface
• trackio_space_id: trackio
• ddp_find_unused_parameters: None
• ddp_bucket_cap_mb: None
• ddp_broadcast_buffers: False
• dataloader_pin_memory: False
• 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: 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: no
• 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: True
• prompts: None
• batch_sampler: no_duplicates
• multi_dataset_batch_sampler: proportional
• router_mapping: {}
• learning_rate_mapping: {}

Training Logs

Epoch	Step	Training Loss	akkadian_val_spearman_cosine
0.1389	5	2.402	-
0.2778	10	2.3992	-
0.4167	15	2.1648	-
0.5556	20	1.8975	-
0.6944	25	1.4115	0.7776
0.8333	30	1.0211	-
0.9722	35	0.6742	-
1.1111	40	0.4176	-
1.25	45	0.2966	-
1.3889	50	0.2419	0.8580
1.5278	55	0.2028	-
1.6667	60	0.1523	-
1.8056	65	0.1445	-
1.9444	70	0.106	-
2.0833	75	0.0906	0.8614
2.2222	80	0.1198	-
2.3611	85	0.0625	-
2.5	90	0.1019	-
2.6389	95	0.0474	-
2.7778	100	0.0945	0.8625
2.9167	105	0.1227	-

Framework Versions

• Python: 3.12.3
• Sentence Transformers: 5.2.0
• Transformers: 4.57.6
• PyTorch: 2.9.0+cu128
• Accelerate: 1.12.0
• Datasets: 4.3.0
• Tokenizers: 0.22.2

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}
}