JianLiao/spectrum-doc-fine-tuned

SentenceTransformer based on BAAI/bge-large-en-v1.5

This is a sentence-transformers model finetuned from BAAI/bge-large-en-v1.5 on the spectrum-design-docs dataset. It maps sentences & paragraphs to a 1024-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: BAAI/bge-large-en-v1.5
• Maximum Sequence Length: 512 tokens
• Output Dimensionality: 1024 dimensions
• Similarity Function: Cosine Similarity
• Training Dataset:
  • spectrum-design-docs
• Language: en

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': True}) with Transformer model: BertModel 
  (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("JianLiao/spectrum-doc-fine-tuned")
# Run inference
sentences = [
    'Represent this sentence for searching relevant passages: How can a designer balance the need for clear text links and the need for emphasized text in a user interface?',
    "Typography\nUsage guidelines\nDon't use underlines for adding emphasis: Underlines are reserved for text links only. They should not be used as a way for adding emphasis to words.\n\n",
    'Meter\nOptions\nPositive variant: The positive variant has a green fill to show the value. This can be used to represent a positive semantic value, such as when there’s a lot of space remaining.',
]
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

• Dataset: sds
• Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.0075
cosine_accuracy@3	0.0156
cosine_accuracy@5	0.0475
cosine_accuracy@10	0.7815
cosine_precision@1	0.0075
cosine_precision@3	0.0052
cosine_precision@5	0.0095
cosine_precision@10	0.0782
cosine_recall@1	0.0075
cosine_recall@3	0.0156
cosine_recall@5	0.0475
cosine_recall@10	0.7815
cosine_ndcg@10	0.2544
cosine_mrr@10	0.1078
cosine_map@100	0.1164

Training Details

Training Dataset

spectrum-design-docs

• Dataset: spectrum-design-docs at 23f5565
• Size: 14,737 training samples
• Columns: anchor and positive
• Approximate statistics based on the first 1000 samples:

  	anchor	positive
  type	string	string
  details	min: 20 tokens mean: 30.87 tokens max: 47 tokens	min: 18 tokens mean: 97.17 tokens max: 512 tokens

• Samples:

  anchor	positive
  Represent this sentence for searching relevant passages: Are there any specific guidelines or best practices provided by the Spectrum team for integrating Spectrum CSS into a new or existing project?	Spectrum CSS: An open source CSS-only implementation of Spectrum, maintained by the Spectrum team. Dependency chain: Spectrum DNA → Spectrum CSS GitHub repository Website #spectrum_css
  Represent this sentence for searching relevant passages: How does the default setting for progress circles affect their behavior in a UI?	Progress circle Options Indeterminate: A progress circle can be either determinate or indeterminate. By default, progress circles are determinate. Use a determinate progress circle when progress can be calculated against a specific goal (e.g., downloading a file of a known size). Use an indeterminate progress circle when progress is happening but the time or effort to completion can’t be determined (e.g., attempting to reconnect to a server).
  Represent this sentence for searching relevant passages: What tools or methods can designers use to test the effectiveness of wrapped legends in their designs?	Legend Behaviors Wrapping: When there isn’t enough space, wrap legends to ensure that dimension values are shown.

• Loss: MultipleNegativesRankingLoss with these parameters:

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

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: epoch
• per_device_train_batch_size: 22
• per_device_eval_batch_size: 16
• gradient_accumulation_steps: 16
• learning_rate: 2e-05
• num_train_epochs: 100
• lr_scheduler_type: cosine
• warmup_ratio: 0.1
• bf16: True
• tf32: True
• load_best_model_at_end: True
• optim: adamw_torch_fused
• prompts: {'anchor': 'Represent this sentence for searching relevant passages: '}
• batch_sampler: no_duplicates

All Hyperparameters

Click to expand

• overwrite_output_dir: False
• do_predict: False
• eval_strategy: epoch
• prediction_loss_only: True
• per_device_train_batch_size: 22
• per_device_eval_batch_size: 16
• per_gpu_train_batch_size: None
• per_gpu_eval_batch_size: None
• gradient_accumulation_steps: 16
• 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: 100
• 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: True
• fp16: False
• fp16_opt_level: O1
• half_precision_backend: auto
• bf16_full_eval: False
• fp16_full_eval: False
• tf32: True
• 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: 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_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
• 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: {'anchor': 'Represent this sentence for searching relevant passages: '}
• batch_sampler: no_duplicates
• multi_dataset_batch_sampler: proportional

Training Logs

Click to expand

Epoch	Step	Training Loss	sds_cosine_ndcg@10
1.0	7	-	0.2255
1.48	10	0.2646	-
2.0	14	-	0.2282
2.96	20	0.1412	-
3.0	21	-	0.2358
4.0	28	-	0.2397
4.32	30	0.0638	-
5.0	35	-	0.2430
5.8	40	0.0425	-
6.0	42	-	0.2449
7.0	49	-	0.2462
7.16	50	0.0237	-
8.0	56	-	0.2428
8.64	60	0.015	-
9.0	63	-	0.2456
10.0	70	0.0082	0.2456
11.0	77	-	0.2498
11.48	80	0.0052	-
12.0	84	-	0.2474
12.96	90	0.0035	-
13.0	91	-	0.2455
14.0	98	-	0.2475
14.32	100	0.0022	-
15.0	105	-	0.2472
15.8	110	0.002	-
16.0	112	-	0.2486
17.0	119	-	0.2506
17.16	120	0.0015	-
18.0	126	-	0.2490
18.64	130	0.0013	-
19.0	133	-	0.2489
20.0	140	0.0012	0.2491
21.0	147	-	0.2493
21.48	150	0.0011	-
22.0	154	-	0.2487
22.96	160	0.001	-
23.0	161	-	0.2486
24.0	168	-	0.2490
24.32	170	0.0008	-
25.0	175	-	0.2502
25.8	180	0.0008	-
26.0	182	-	0.2505
27.0	189	-	0.2523
27.16	190	0.0008	-
28.0	196	-	0.2516
28.64	200	0.0007	-
29.0	203	-	0.2509
30.0	210	0.0007	0.2522
31.0	217	-	0.2522
31.48	220	0.0006	-
32.0	224	-	0.2534
32.96	230	0.0007	-
33.0	231	-	0.2523
34.0	238	-	0.2524
34.32	240	0.0006	-
35.0	245	-	0.2518
35.8	250	0.0006	-
36.0	252	-	0.2529
37.0	259	-	0.2524
37.16	260	0.0006	-
38.0	266	-	0.2530
38.64	270	0.0005	-
39.0	273	-	0.2526
40.0	280	0.0006	0.2539
41.0	287	-	0.2529
41.48	290	0.0005	-
42.0	294	-	0.2545
42.96	300	0.0006	-
43.0	301	-	0.2534
44.0	308	-	0.2536
44.32	310	0.0004	-
45.0	315	-	0.2521
45.8	320	0.0005	-
46.0	322	-	0.2532
47.0	329	-	0.2519
47.16	330	0.0005	-
48.0	336	-	0.2525
48.64	340	0.0004	-
49.0	343	-	0.2535
50.0	350	0.0005	0.2542
51.0	357	-	0.2540
51.48	360	0.0004	-
52.0	364	-	0.2542
52.96	370	0.0005	-
53.0	371	-	0.2538
54.0	378	-	0.2533
54.32	380	0.0004	-
55.0	385	-	0.2544
55.8	390	0.0004	-
56.0	392	-	0.2539
57.0	399	-	0.2541
57.16	400	0.0005	-
58.0	406	-	0.2532
58.64	410	0.0004	-
59.0	413	-	0.2543
60.0	420	0.0004	0.2532
61.0	427	-	0.2541
61.48	430	0.0004	-
62.0	434	-	0.2542
62.96	440	0.0005	-
63.0	441	-	0.2546
64.0	448	-	0.2549
64.32	450	0.0003	-
65.0	455	-	0.2557
65.8	460	0.0004	-
66.0	462	-	0.2557
67.0	469	-	0.2539
67.16	470	0.0004	-
68.0	476	-	0.2538
68.64	480	0.0004	-
69.0	483	-	0.2538
70.0	490	0.0004	0.2542
71.0	497	-	0.2532
71.48	500	0.0004	-
72.0	504	-	0.2538
72.96	510	0.0004	-
73.0	511	-	0.2545
74.0	518	-	0.2531
74.32	520	0.0003	-
75.0	525	-	0.2534
75.8	530	0.0004	-
76.0	532	-	0.2541
77.0	539	-	0.2545
77.16	540	0.0004	-
78.0	546	-	0.2536
78.64	550	0.0004	-
79.0	553	-	0.2545
80.0	560	0.0004	0.2540
81.0	567	-	0.2545
81.48	570	0.0004	-
82.0	574	-	0.2541
82.96	580	0.0004	-
83.0	581	-	0.2545
84.0	588	-	0.2538
84.32	590	0.0004	-
85.0	595	-	0.2546
85.8	600	0.0004	0.2544

• The bold row denotes the saved checkpoint.

Framework Versions

• Python: 3.12.8
• Sentence Transformers: 3.3.1
• Transformers: 4.47.1
• PyTorch: 2.5.1+cu124
• Accelerate: 1.2.1
• 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}
}