hi-di-hi-base

This is a sentence-transformers model finetuned from BAAI/bge-base-en-v1.5 on the json dataset. It maps sentences & paragraphs to a 768-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-base-en-v1.5
Maximum Sequence Length: 512 tokens
Output Dimensionality: 768 dimensions
Similarity Function: Cosine Similarity
Training Dataset:
- json
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': 768, '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("carsondial/slinger20241231-1")
# Run inference
sentences = [
    'What is article marketing and how does it work?\nHow can I use article marketing to increase traffic to my website?\nWhat are some effective strategies for an article marketing campaign?',
    "Advertising An Article Marketing\nAs every website owner knows, content is essential to raising your rankings in the search engines. There's another way to use content to bring even more visitors to your site. It's called article marketing. You can use one article in dozens of venues, to multiply your exposure exponentially. More traffic means more sales. Let's look at some strategies you can use in your article marketing campaign.Let's say you sell seeds, gardening supplies and gardening books on your site. You're facing stiff competition, as there are many websites marketing these products.\nHow do you get your share of attention? The answer is simple: exposure, as much as you can generate! Article marketing is a natural for this purpose. Write one article, and use it to advertise your site, over and over again.\nYou know your products and their uses best. Write an article on how to start seeds. Give detailed and useful information such as equipment required, temperature and seasonal concerns, thinning and transplanting. There are many topics you can write about.\nWrite an article telling your reader how to choose a good pair of garden gloves, and the advantages of a leather glove for one task, or a sturdy cotton glove for light work.\nOr, tell your reader how a quality bulb planter makes short work of a big planting project. I'm sure you get the idea. Do not write an advertisement!\nWhen your articles are prepared, your related subjectsarticle marketing campaign is ready to launch. Do a search using the term ”article directories gardening”. You'll get hundreds or even thousands of results. Visit some of these article directories and browse their gardening category. Take a look at the quality of their articles. Unfortunately, there are many article directories with poorly written articles, so be choosy when selecting those you want to be associated with, as quality-conscious directories get the most readers.\nHere's how it works: you submit your article to the directory. You are not paid money, but are rewarded with a byline and a link to your site. Every person who reads your article has the opportunity to click through to your site. Give your reader useful information in an engaging style, and chances are good they'll come to your site to see what else they can learn. You can submit the same article to other directories, and watch your traffic and sales grow.\nYour article marketing strategy should include contacts with other website owners, especially those with an e-newsletter. Offer your article as newsletter content in exchange for a byline. See if you can negotiate reciprocal links. Of course you don't want to contact direct competition. A florist, greenhouse or organic gardening site might be good choices.\nYour article marketing campaign's success is directly related to the amount of exposure you create through your effort. One indirect result of a good article marketing campaign is establishing yourself as an expert in your business. Success is measured by traffic and sales. Be prolific and diligent!\nAbout the Author: InsightsOnMarketing provides readers with the latest reviews, articles, commentaries and write-ups on all article marketing, article directories, article reader related subjects",
    'Alandari Gray Loveseat\nSporting a pleasing roll arm for classic flair, the Alandari loveseat in soft gray exudes an easy elegance sure to please. Feel-good upholstery with linen texturing is made for everyday luxury. Muted-tone toss pillows are a soothing, sophisticated complement.\nFrame constructions have been rigorously tested to simulate the home and transportation environments for improved durability. Frame components are secured with combinations of glue, blocks, interlocking panels and staples. Seats and back spring rails are cut from mixed hardwood and engineered lumber. Stripes and patterns are match cut. All fabrics are pre-approved for wearability and durability against AHFA standards. Cushions are constructed of low melt fiber wrapped over high quality foam. Constructed with a platform seat foundation.',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 768]

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

Evaluation

Metrics

Information Retrieval

Datasets: dim_768, dim_512, dim_256, dim_128 and dim_64
Evaluated with InformationRetrievalEvaluator

Metric	dim_768	dim_512	dim_256	dim_128	dim_64
cosine_accuracy@1	0.8376	0.837	0.8256	0.814	0.768
cosine_accuracy@3	0.9022	0.8996	0.8938	0.879	0.8472
cosine_accuracy@5	0.9204	0.9208	0.9166	0.9082	0.879
cosine_accuracy@10	0.9438	0.9436	0.9402	0.9344	0.9108
cosine_precision@1	0.8376	0.837	0.8256	0.814	0.768
cosine_precision@3	0.3007	0.2999	0.2979	0.293	0.2824
cosine_precision@5	0.1841	0.1842	0.1833	0.1816	0.1758
cosine_precision@10	0.0944	0.0944	0.094	0.0934	0.0911
cosine_recall@1	0.8376	0.837	0.8256	0.814	0.768
cosine_recall@3	0.9022	0.8996	0.8938	0.879	0.8472
cosine_recall@5	0.9204	0.9208	0.9166	0.9082	0.879
cosine_recall@10	0.9438	0.9436	0.9402	0.9344	0.9108
cosine_ndcg@10	0.8911	0.8901	0.8828	0.8728	0.8382
cosine_mrr@10	0.8742	0.873	0.8644	0.8532	0.8151
cosine_map@100	0.8762	0.875	0.8664	0.8553	0.8181

Training Details

Training Dataset

json

Dataset: json
Size: 45,000 training samples
Columns: anchor and positive
Approximate statistics based on the first 1000 samples:
anchor positive
type string string
details
min: 12 tokens
mean: 30.74 tokens
max: 112 tokens

min: 51 tokens
mean: 382.61 tokens
max: 512 tokens

	anchor	positive
type	string	string
details	min: 12 tokens mean: 30.74 tokens max: 112 tokens	min: 51 tokens mean: 382.61 tokens max: 512 tokens

Samples:

anchor	positive
`How to split the Join form into multiple parts Adding HTML and text to the Join form page Customizing the login logout message`	This tutorial will show you how to split the Join form/page into two or three parts. Whether you want a cleaner looking join form or you want to reduce the number of spam bots this should help. This Tutorial will show you how to add HTML and text to the Join form page. If you have ever wanted to add a few more details or spruce up the Join form then this might be what you have been looking for. This tutorial will show you how to spruce up the login - logout message people see. If you want to add a special message or just change the standard default phrase this one might be for you. It's very simple and only takes a few moments to apply this one. This tutorial will show you how to take advantage and optimize your site using the Page Block feature available in your Administration panel. With a little exploring and tweaking it's possible to boost your site's performance. If your Forum RSS feed links are redirecting to your homepage when clicked on you might need to update the forum module...
`Failed mobile technologies of 2011 What happened to Siri Disappointing mobile innovations`	Last year we all got excited about mobile-computing products that failed to deliver It's an awesome time to be a gadget-happy consumer electronics freak. Multi-touch user interfaces. Huge advances in miniaturization and battery life. Cloud-based storage. Mobile computing has never been better. But sometimes, when companies announce incredible new products or technologies, and everybody proclaims that a new era has dawned, and that culture-shifting transformations are about to take place -- nothing happens. Here are five mobile technologies from last year that were supposed to change the world, but didn't. Apple seemed to do everything right with its voice assistant strategy. The company acquired the leading app maker with the best technology. It spent two years perfecting and integrating the technology, and bulking up on servers to handle the number-crunching required to deliver human-like voice interaction. Siri was then launched to huge fanfare. Overnight, people changed how they int...
`What are the different types of web usage for voice over campaigns? What are the standard types of web usage for voice over? What are the different ways to use voice over in web campaigns?`	Different Types of Web Usage Clearly, nothing has upended the voice over industry quite like the Internet. On all ends of the spectrum, from the basic sourcing of voice over talent to right on down to hyper-targeted, increasingly personalized messaging, it can feel a little “Wild West” to navigate if you’re not going directly through an agent. And one of the biggest stumblers when sourcing for voice talent? The concept of “web usage.” Usage — a.k.a., how the finished voice over is being used — will cause the greatest fluctuation in a voice talent’s quote. As we all know, there are myriad ways any material can be used and distributed online… and not all distribution is created equally. Below, we’ve got a handy little guide detailing all the different, standard types of web usage we tend to come across here at Blue Wave Voiceover (Heck, the way things are going, we might have to update this once more before you finish reading it). Feel free to refer to this when putting together any kind...

Loss: MatryoshkaLoss with these parameters:

{
    "loss": "MultipleNegativesRankingLoss",
    "matryoshka_dims": [
        768,
        512,
        256,
        128,
        64
    ],
    "matryoshka_weights": [
        1,
        1,
        1,
        1,
        1
    ],
    "n_dims_per_step": -1
}

Training Hyperparameters

Non-Default Hyperparameters

eval_strategy: epoch
per_device_train_batch_size: 32
per_device_eval_batch_size: 16
gradient_accumulation_steps: 16
learning_rate: 2e-05
num_train_epochs: 4
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: True
tf32: True
load_best_model_at_end: True
optim: adamw_torch_fused
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: 32
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: 4
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: 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: None
batch_sampler: no_duplicates
multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	Training Loss	dim_768_cosine_ndcg@10	dim_512_cosine_ndcg@10	dim_256_cosine_ndcg@10	dim_128_cosine_ndcg@10	dim_64_cosine_ndcg@10
0.1137	10	1.0843	-	-	-	-	-
0.2274	20	0.6597	-	-	-	-	-
0.3412	30	0.3466	-	-	-	-	-
0.4549	40	0.253	-	-	-	-	-
0.5686	50	0.2286	-	-	-	-	-
0.6823	60	0.2007	-	-	-	-	-
0.7960	70	0.1576	-	-	-	-	-
0.9097	80	0.1652	-	-	-	-	-
0.9893	87	-	0.8880	0.8861	0.8788	0.8657	0.8267
1.0341	90	0.1563	-	-	-	-	-
1.1478	100	0.1485	-	-	-	-	-
1.2615	110	0.108	-	-	-	-	-
1.3753	120	0.0874	-	-	-	-	-
1.4890	130	0.0851	-	-	-	-	-
1.6027	140	0.0897	-	-	-	-	-
1.7164	150	0.0803	-	-	-	-	-
1.8301	160	0.0645	-	-	-	-	-
1.9439	170	0.0806	-	-	-	-	-
1.9893	174	-	0.8906	0.8893	0.8815	0.8712	0.8351
2.0682	180	0.0812	-	-	-	-	-
2.1819	190	0.0743	-	-	-	-	-
2.2957	200	0.0531	-	-	-	-	-
2.4094	210	0.0448	-	-	-	-	-
2.5231	220	0.0465	-	-	-	-	-
2.6368	230	0.0486	-	-	-	-	-
2.7505	240	0.0509	-	-	-	-	-
2.8643	250	0.0395	-	-	-	-	-
2.9780	260	0.0521	-	-	-	-	-
2.9893	261	-	0.8912	0.8897	0.8823	0.8720	0.8375
3.1023	270	0.0551	-	-	-	-	-
3.2161	280	0.0412	-	-	-	-	-
3.3298	290	0.0373	-	-	-	-	-
3.4435	300	0.0387	-	-	-	-	-
3.5572	310	0.0438	-	-	-	-	-
3.6709	320	0.0433	-	-	-	-	-
3.7846	330	0.0368	-	-	-	-	-
3.8984	340	0.0418	-	-	-	-	-
3.9893	348	-	0.8911	0.8901	0.8828	0.8728	0.8382

The bold row denotes the saved checkpoint.

Framework Versions

Python: 3.10.12
Sentence Transformers: 3.3.1
Transformers: 4.47.1
PyTorch: 2.5.1+cu121
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",
}

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}
}

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}
}

carsondial
/

slinger20241231-1