bge-small-matryoshka-fine-tuned

BGE-base-en-v1.5-Hotpotqa

This is a sentence-transformers model finetuned from BAAI/bge-small-en on the sentence-transformers/hotpotqa dataset. It maps sentences & paragraphs to a 384-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-small-en <!-- at revision 2275a7bdee235e9b4f01fa73aa60d3311983cfea -->
• Maximum Sequence Length: 512 tokens
• Output Dimensionality: 384 tokens
• Similarity Function: Cosine Similarity
• Training Dataset:
  • sentence-transformers/hotpotqa
• Language: en
• License: apache-2.0

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': 384, '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("sentence_transformers_model_id")
# Run inference
sentences = [
    'Red Velvet is a 2012 play by Lolita Chakrabarti, dealing with the biography of a 19th century actor born in which year ?',
    'Red Velvet (play) Red Velvet is a 2012 play by Lolita Chakrabarti, dealing with the biography of the 19th century actor Ira Aldridge and his taking the role of "Othello". It premiered at the Tricycle Theatre (directed by its new artistic director Indhu Rubasingham) from 11 October to 24 November 2012, with Aldridge played by Adrian Lester.',
    "Herbert Campbell Herbert Campbell (22 December 1844 – 19 July 1904) born Herbert Edward Story was an English comedian and actor who appeared in music hall, Victorian burlesques and musical comedies during the Victorian era. He was famous for starring, for forty years, in the Theatre Royal, Drury Lane's annual Christmas pantomimes, predominantly as a dame.",
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 384]

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

<!-- ### Direct Usage (Transformers) <details><summary>Click to see the direct usage in Transformers</summary> </details> --> <!-- ### Downstream Usage (Sentence Transformers) You can finetune this model on your own dataset. <details><summary>Click to expand</summary> </details> --> <!-- ### Out-of-Scope Use *List how the model may foreseeably be misused and address what users ought not to do with the model.* -->

Evaluation

Metrics

Triplet

• Dataset: dim_384
• Evaluated with <code>TripletEvaluator</code>

Triplet

• Dataset: dim_256
• Evaluated with <code>TripletEvaluator</code>

Triplet

• Dataset: dim_128
• Evaluated with <code>TripletEvaluator</code>

Triplet

• Dataset: dim_64
• Evaluated with <code>TripletEvaluator</code>

<!-- ## Bias, Risks and Limitations *What are the known or foreseeable issues stemming from this model? You could also flag here known failure cases or weaknesses of the model.* --> <!-- ### Recommendations *What are recommendations with respect to the foreseeable issues? For example, filtering explicit content.* -->

Training Details

Training Dataset

sentence-transformers/hotpotqa

• Dataset: sentence-transformers/hotpotqa at f07d3cd
• Size: 76,064 training samples
• Columns: <code>anchor</code>, <code>positive</code>, and <code>negative</code>
• Approximate statistics based on the first 1000 samples:

  	anchor	positive	negative
  type	string	string	string
  details	<ul><li>min: 8 tokens</li><li>mean: 24.08 tokens</li><li>max: 95 tokens</li></ul>	<ul><li>min: 23 tokens</li><li>mean: 100.12 tokens</li><li>max: 512 tokens</li></ul>	<ul><li>min: 15 tokens</li><li>mean: 88.02 tokens</li><li>max: 512 tokens</li></ul>

• Samples:

  anchor	positive	negative
  <code>What nationality was the player named MVP in 2017 World Baseball Classic – Pool C ?</code>	<code>2017 World Baseball Classic – Pool C Pool C of the First Round of the 2017 World Baseball Classic was held at Marlins Park, Miami, Florida, United States, from March 9 to 12, 2017, between Canada, Colombia, the Dominican Republic, and the United States. Pool C was a round-robin tournament. Each team played the other three teams once, with the top two teams – the Dominican Republic and the United States – advancing to Pool F, one of two second-round pools. Manny Machado of the Dominican Republic was named MVP for the first-round Pool C bracket of the WBC, after batting .357.</code>	<code>2017 World Baseball Classic – Qualifier 2 Qualifier 2 of the Qualifying Round of the 2017 World Baseball Classic was held at Estadio B'Air, Mexicali, Mexico from March 17 to 20, 2016.</code>
  <code>Karl Kraepelin specialized in the study of what predatory arachnids?</code>	<code>Karl Kraepelin Karl Matthias Friedrich Magnus Kraepelin (14 December 1848 Neustrelitz – 28 June 1915 Hamburg), was a German naturalist who specialised in the study of scorpions, centipedes, spiders and solfugids, and was noted for his monograph ""Scorpiones und Pedipalpi"" (Berlin) in 1899, which was an exhaustive survey of the taxonomy of the Order Scorpiones. From 1889–1914 he was Director of the "Naturhistorisches Museum Hamburg ", which was destroyed during World War II, and worked on myriapods from 1901–1916.</code>	<code>Teuthology Teuthology (from Greek "τεῦθος" , "cuttlefish, squid", and -λογία , "-logia") is the study of cephalopods.</code>
  <code>Who directed the 1990 American crime film in which Vito Pick me played a bit part?</code>	<code>Vito Picone Vito Picone (born March 20, 1941) is the lead singer of The Elegants, and along with Jimmy Mochella is a remaining original member. He has also played bit parts in "Goodfellas", "Analyze This", and "The Sopranos".</code>	<code>The Rookie (1990 film) The Rookie is a 1990 American buddy cop film directed by Clint Eastwood and produced by Howard G. Kazanjian, Steven Siebert and David Valdes. It was written from a screenplay conceived by Boaz Yakin and Scott Spiegel. The film stars Charlie Sheen, Clint Eastwood, Raúl Juliá, Sônia Braga, Lara Flynn Boyle, and Tom Skerritt. Eastwood plays a veteran police officer teamed up with a younger detective played by Sheen ("the rookie"), whose intent is to take down a German crime lord in downtown Los Angeles following months of investigation into an exotic car theft ring.</code>

• Loss: <code>MatryoshkaLoss</code> with these parameters:

  {
      "loss": "TripletLoss",
      "matryoshka_dims": [
          384,
          256,
          128,
          64
      ],
      "matryoshka_weights": [
          1,
          1,
          1,
          1
      ],
      "n_dims_per_step": -1
  }
  

Evaluation Dataset

sentence-transformers/hotpotqa

• Dataset: sentence-transformers/hotpotqa at f07d3cd
• Size: 8,452 evaluation samples
• Columns: <code>anchor</code>, <code>positive</code>, and <code>negative</code>
• Approximate statistics based on the first 1000 samples:

  	anchor	positive	negative
  type	string	string	string
  details	<ul><li>min: 10 tokens</li><li>mean: 24.53 tokens</li><li>max: 114 tokens</li></ul>	<ul><li>min: 19 tokens</li><li>mean: 103.87 tokens</li><li>max: 407 tokens</li></ul>	<ul><li>min: 17 tokens</li><li>mean: 88.32 tokens</li><li>max: 356 tokens</li></ul>

• Samples:

  anchor	positive	negative
  <code>Which actress, known for her role as Harper Munroe on the MTV comedy series "Happyland", starred alongside Laura Marano, Parker Mack, Michelle Clunie and Kathleen Wilhoite in the film A Sort of Homecoming?</code>	<code>A Sort of Homecoming (film) A Sort of Homecoming is an American drama directed by Maria Burton, her fifth feature film. The films stars Katherine McNamara, Laura Marano, Parker Mack, Michelle Clunie and Kathleen Wilhoite. The film premiered March 14, 2015 at the Omaha Film Festival.</code>	<code>Nellie Bellflower Nellie Bellflower (born May 1, 1946 in Phoenix, Arizona) is an American actress and voice artist who provided the voice of Princess Ariel in the Ruby-Spears animated television series "Thundarr the Barbarian". She has also been in "The Last Unicorn" (voice), Rankin/Bass "The Return of the King", "Americathon", the miniseries "East of Eden", and guest roles on various TV shows such as "Barnaby Jones", "Barney Miller", "Starsky and Hutch", and "Happy Days" as Fonzie's ex-fiancée Maureen Johnson, a.k.a. "The Lone Stripper", in the Season 2 episode of the series titled "Fonzie's Getting Married" (episode #13). Nellie has been involved in movie production with three projects: "The Girl in Melanie Klein" (2008), "Miss Pettigrew Lives for a Day" (2008) and "Finding Neverland" (2004), for which she was nominated for an Academy Award as Producer for Best Picture. She is married to Michael Mislove.</code>
  <code>Between Pizza Fusion and Pizzeria Venti, which restaurant emphasizes organic ingredients and green building methods?</code>	<code>Pizza Fusion Pizza Fusion is a Deerfield Beach, Florida-based pizza restaurant chain. Using mostly organic ingredients and emphasizing green building methods, the restaurants operate under the tagline Saving the Earth, One Pizza at a Time.</code>	<code>Pizza Schmizza Pizza Schmizza is an American pizza chain with 23 locations throughout the Portland, Oregon area, and two in southern Oregon. Pizza Schmizza, primarily selling thin crust pizza by-the-slice.</code>
  <code>What company distributed the stop motion spin-off special "The Year Without a Santa Claus," which aired on December 10, 1974?</code>	<code>A Miser Brothers' Christmas A Miser Brothers' Christmas is a stop motion spin-off special based on some of the characters from the 1974 Rankin-Bass special "The Year Without a Santa Claus". Distributed by Warner Bros. Animation under their Warner Premiere label (the rights holders of the post-1974 Rankin-Bass library) and Toronto-based Cuppa Coffee Studios, the one-hour special premiered on ABC Family on Saturday, December 13, 2008, during the network's annual The 25 Days of Christmas programming. Mickey Rooney and George S. Irving reprised their respective roles as Santa Claus and Heat Miser at ages 88 and 86. Snow Miser, originally portrayed by Dick Shawn who died in 1987, was voiced by Juan Chioran, while Mrs. Claus, voiced by Shirley Booth in the original, was portrayed by Catherine Disher (because Booth had died in 1992). The movie aimed to emulate the Rankin/Bass animation style. This is the last Christmas special to feature Mickey Rooney as Santa Claus, as he died in 2014, as well as the last time George Irving voiced Heat Miser, as he died in 2016.</code>	<code>Holidaze: The Christmas That Almost Didn't Happen Holidaze: The Christmas That Almost Didn't Happen is an American stop-motion animated Christmas television special directed by David H. Brooks, that originally aired in 2006 and produced by BixPix Entertainment, Once Upon a Frog and Madison Street Entertainment. The show's plot has Rusty Reindeer (Fred Savage) the brother of Rudolph the Red Nosed Reindeer joining a support group for depressed holiday icons, and he and the other characters search for the meaning of Christmas and help a young boy (Dylan and Cole Sprouse) to get on Santa's nice list. Rusty's cohorts include Candie, the Easter Bunny (Gladys Knight); Mr. C, the grouchy cherub (Paul Rodriguez); Albert, the Thanksgiving Turkey (Harland Williams); And Trick and Treat (Brenda Song and Emily Osment) the teenage Halloween Ghosts.</code>

• Loss: <code>MatryoshkaLoss</code> with these parameters:

  {
      "loss": "TripletLoss",
      "matryoshka_dims": [
          384,
          256,
          128,
          64
      ],
      "matryoshka_weights": [
          1,
          1,
          1,
          1
      ],
      "n_dims_per_step": -1
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 32
• per_device_eval_batch_size: 32
• gradient_accumulation_steps: 16
• learning_rate: 2e-05
• num_train_epochs: 50
• lr_scheduler_type: cosine
• warmup_ratio: 0.1
• bf16: True
• tf32: True
• load_best_model_at_end: True
• optim: adamw_torch_fused
• resume_from_checkpoint: bge-small-hotpotwa-matryoshka
• 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: 32
• per_device_eval_batch_size: 32
• per_gpu_train_batch_size: None
• per_gpu_eval_batch_size: None
• gradient_accumulation_steps: 16
• eval_accumulation_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: 50
• 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: 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: 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: bge-small-hotpotwa-matryoshka
• hub_model_id: None
• hub_strategy: every_save
• hub_private_repo: False
• hub_always_push: False
• gradient_checkpointing: False
• gradient_checkpointing_kwargs: None
• include_inputs_for_metrics: False
• 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
• batch_sampler: no_duplicates
• multi_dataset_batch_sampler: proportional

</details>

Training Logs

Framework Versions

• Python: 3.10.10
• Sentence Transformers: 3.0.1
• Transformers: 4.41.2
• PyTorch: 2.1.2+cu121
• Accelerate: 0.33.0
• Datasets: 2.19.1
• Tokenizers: 0.19.1

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",
}

MatryoshkaLoss

@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}
}

TripletLoss

@misc{hermans2017defense,
    title={In Defense of the Triplet Loss for Person Re-Identification}, 
    author={Alexander Hermans and Lucas Beyer and Bastian Leibe},
    year={2017},
    eprint={1703.07737},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

<!-- ## Glossary *Clearly define terms in order to be accessible across audiences.* --> <!-- ## Model Card Authors *Lists the people who create the model card, providing recognition and accountability for the detailed work that goes into its construction.* --> <!-- ## Model Card Contact *Provides a way for people who have updates to the Model Card, suggestions, or questions, to contact the Model Card authors.* -->