This is a PyLate model finetuned from answerdotai/ModernBERT-base. It maps sentences & paragraphs to sequences of 128-dimensional dense vectors and can be used for semantic textual similarity using the MaxSim operator.

Model Details

Model Description

Model Type: PyLate model
Base model: answerdotai/ModernBERT-base
Document Length: 180 tokens
Query Length: 32 tokens
Output Dimensionality: 128 tokens
Similarity Function: MaxSim

Model Sources

Documentation: PyLate Documentation
Repository: PyLate on GitHub
Hugging Face: PyLate models on Hugging Face

Full Model Architecture

ColBERT(
  (0): Transformer({'max_seq_length': 31, 'do_lower_case': False}) with Transformer model: ModernBertModel 
  (1): Dense({'in_features': 768, 'out_features': 128, 'bias': False, 'activation_function': 'torch.nn.modules.linear.Identity'})
)

Usage

First install the PyLate library:

pip install -U pylate

Retrieval

PyLate provides a streamlined interface to index and retrieve documents using ColBERT models. The index leverages the Voyager HNSW index to efficiently handle document embeddings and enable fast retrieval.

Indexing documents

First, load the ColBERT model and initialize the Voyager index, then encode and index your documents:

python

from pylate import indexes, models, retrieve

# Step 1: Load the ColBERT model
model = models.ColBERT(
    model_name_or_path=ayushexel/colbert-ModernBERT-base-1-neg-5-epoch-gooaq-1995000,
)

# Step 2: Initialize the Voyager index
index = indexes.Voyager(
    index_folder="pylate-index",
    index_name="index",
    override=True,  # This overwrites the existing index if any
)

# Step 3: Encode the documents
documents_ids = ["1", "2", "3"]
documents = ["document 1 text", "document 2 text", "document 3 text"]

documents_embeddings = model.encode(
    documents,
    batch_size=32,
    is_query=False,  # Ensure that it is set to False to indicate that these are documents, not queries
    show_progress_bar=True,
)

# Step 4: Add document embeddings to the index by providing embeddings and corresponding ids
index.add_documents(
    documents_ids=documents_ids,
    documents_embeddings=documents_embeddings,
)

Note that you do not have to recreate the index and encode the documents every time. Once you have created an index and added the documents, you can re-use the index later by loading it:

python

# To load an index, simply instantiate it with the correct folder/name and without overriding it
index = indexes.Voyager(
    index_folder="pylate-index",
    index_name="index",
)

Retrieving top-k documents for queries

Once the documents are indexed, you can retrieve the top-k most relevant documents for a given set of queries. To do so, initialize the ColBERT retriever with the index you want to search in, encode the queries and then retrieve the top-k documents to get the top matches ids and relevance scores:

python

# Step 1: Initialize the ColBERT retriever
retriever = retrieve.ColBERT(index=index)

# Step 2: Encode the queries
queries_embeddings = model.encode(
    ["query for document 3", "query for document 1"],
    batch_size=32,
    is_query=True,  #  # Ensure that it is set to False to indicate that these are queries
    show_progress_bar=True,
)

# Step 3: Retrieve top-k documents
scores = retriever.retrieve(
    queries_embeddings=queries_embeddings,
    k=10,  # Retrieve the top 10 matches for each query
)

Reranking

If you only want to use the ColBERT model to perform reranking on top of your first-stage retrieval pipeline without building an index, you can simply use rank function and pass the queries and documents to rerank:

python

from pylate import rank, models

queries = [
    "query A",
    "query B",
]

documents = [
    ["document A", "document B"],
    ["document 1", "document C", "document B"],
]

documents_ids = [
    [1, 2],
    [1, 3, 2],
]

model = models.ColBERT(
    model_name_or_path=ayushexel/colbert-ModernBERT-base-1-neg-5-epoch-gooaq-1995000,
)

queries_embeddings = model.encode(
    queries,
    is_query=True,
)

documents_embeddings = model.encode(
    documents,
    is_query=False,
)

reranked_documents = rank.rerank(
    documents_ids=documents_ids,
    queries_embeddings=queries_embeddings,
    documents_embeddings=documents_embeddings,
)

Evaluation

Metrics

Col BERTTriplet

Evaluated with pylate.evaluation.colbert_triplet.ColBERTTripletEvaluator

Metric	Value
accuracy	0.4588

Training Details

Training Dataset

Unnamed Dataset

Size: 1,867,662 training samples
Columns: question, answer, and negative
Approximate statistics based on the first 1000 samples:
question answer negative
type string string string
details
min: 9 tokens
mean: 13.12 tokens
max: 22 tokens
min: 16 tokens
mean: 31.73 tokens
max: 32 tokens
min: 15 tokens
mean: 31.64 tokens
max: 32 tokens

Samples:

question	answer	negative
`are mandarins same as clementines?`	`Mandarins… When it comes to Clementines vs. Mandarins, the Mandarin is the master orange of the family, and Clementines, tangerines, and satsumas all fall under this umbrella.`	`A: CUTIES® are actually two varieties of mandarins: Clementine mandarins, available November through January; and W. Murcott mandarins, available February through April. ... Unlike other mandarins or oranges, they are seedless, super sweet, easy to peel and kid-sized—only a select few achieve CUTIES® ' high standards.`
`why are snow leopards no longer endangered?`	`The snow leopard is no longer an endangered species, but its population in the wild is still at risk because of poaching and habitat loss, conservationists said this week. ... Conservationists warned that the risks are not over for the snow leopards, whose distinctive appearances make them attractive to poachers.`	`The term "big cat" is typically used to refer to any of the five living members of the genus Panthera, namely tiger, lion, jaguar, leopard, and snow leopard. Except the snow leopard, these species are able to roar.`
`are waves measured from the front or back?`	`In scientific terms and most used by the surfing community around the world, the wave height is measured vertically from the trough to the crest and is known by surfers as face scale. In Hawaii, local surfers use the back of the wave to measure wave height and is called Hawaiian scale or local scale.`	`Wavelength is the distance between sound waves while frequency is the number of times in which the sound wave occurs. 2. Wavelength is used to measure the length of sound waves while frequency is used to measure the recurrence of sound waves.`

Loss: pylate.losses.contrastive.Contrastive

Evaluation Dataset

Unnamed Dataset

Size: 5,000 evaluation samples
Columns: question, answer, and negative_1
Approximate statistics based on the first 1000 samples:
question answer negative_1
type string string string
details
min: 9 tokens
mean: 13.02 tokens
max: 25 tokens
min: 16 tokens
mean: 31.66 tokens
max: 32 tokens
min: 15 tokens
mean: 31.41 tokens
max: 32 tokens

Samples:

question	answer	negative_1
`what is the best shampoo for thin curly hair?`	`['Best For Daily Cleansing: Mizani True Textures Cream Cleansing Conditioner. ... ', 'Best For Coils: Ouidad VitalCurl Clear & Gentle Shampoo. ... ', 'Best For Restoring Shine: Shea Moisture Coconut & Hibiscus Curl & Shine Shampoo. ... ', 'Best For Fine Curls: Renee Furterer Sublime Curl Curl Activating Shampoo.']`	`Whether you have straight or curly hair, thin or thick, this is another option that you should not miss for the best OGX shampoo. The Australian tea tree oils in this shampoo are effective for repair of oily, damaged, and frizzy hair. ... It also makes a great choice of shampoo for people who have dry scalp.`
`how many days after my period do i start ovulating?`	`Many women typically ovulate around 12 to 14 days after the first day of their last period, but some have a naturally short cycle. They may ovulate as soon as six days or so after the first day of their last period.`	`If you have a short cycle, for example, 21 days, and you bleed for 7 days, then you could ovulate right after your period. This is because ovulation generally occurs 12-16 days before your next period begins, and this would estimate you ovulating at days 6-10 of your cycle.`
`are the apes in planet of the apes cgi?`	`Unlike in the original 1968 film, there are no monkey suits, heavy makeup jobs or wigs. All of the apes audiences see on-screen are motion-capture CGI apes, which lends them a more realistic effect as the CGI is based on the actors' actual movements.`	`Among the living primates, humans are most closely related to the apes, which include the lesser apes (gibbons) and the great apes (chimpanzees, gorillas and orangutans).`

Loss: pylate.losses.contrastive.Contrastive

Training Hyperparameters

Non-Default Hyperparameters

eval_strategy: steps
per_device_train_batch_size: 180
per_device_eval_batch_size: 180
learning_rate: 3e-06
num_train_epochs: 5
warmup_ratio: 0.1
seed: 12
bf16: True
dataloader_num_workers: 12
load_best_model_at_end: True

All Hyperparameters

Click to expand

overwrite_output_dir: False
do_predict: False
eval_strategy: steps
prediction_loss_only: True
per_device_train_batch_size: 180
per_device_eval_batch_size: 180
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: 3e-06
weight_decay: 0.0
adam_beta1: 0.9
adam_beta2: 0.999
adam_epsilon: 1e-08
max_grad_norm: 1.0
num_train_epochs: 5
max_steps: -1
lr_scheduler_type: linear
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: 12
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: None
local_rank: 0
ddp_backend: None
tpu_num_cores: None
tpu_metrics_debug: False
debug: []
dataloader_drop_last: True
dataloader_num_workers: 12
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
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: batch_sampler
multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	Training Loss	accuracy
0	0	-	0.4588
0.0004	1	19.095	-
0.0771	200	12.1999	-
0.1543	400	5.3887	-
0.2314	600	1.9588	-
0.3085	800	1.0338	-
0.3857	1000	0.8202	-
0.4628	1200	0.6943	-
0.5399	1400	0.6087	-
0.6170	1600	0.5484	-
0.6942	1800	0.5049	-
0.7713	2000	0.4754	-
0.8484	2200	0.4497	-
0.9256	2400	0.4298	-
1.0027	2600	0.4102	-
1.0798	2800	0.3861	-
1.1570	3000	0.3727	-
1.2341	3200	0.3632	-
1.3112	3400	0.3507	-
1.3884	3600	0.3431	-
1.4655	3800	0.3345	-
1.5426	4000	0.3282	-
1.6197	4200	0.3232	-
1.6969	4400	0.3148	-
1.7740	4600	0.3053	-
1.8511	4800	0.3003	-
1.9283	5000	0.2981	-
2.0054	5200	0.2913	-
2.0825	5400	0.2787	-
2.1597	5600	0.2764	-
2.2368	5800	0.2742	-
2.3139	6000	0.2693	-
2.3911	6200	0.2698	-
2.4682	6400	0.2635	-
2.5453	6600	0.2586	-
2.6224	6800	0.2577	-
2.6996	7000	0.257	-
2.7767	7200	0.2541	-
2.8538	7400	0.2539	-
2.9310	7600	0.25	-
3.0081	7800	0.2498	-
3.0852	8000	0.2377	-
3.1624	8200	0.2388	-
3.2395	8400	0.2377	-
3.3166	8600	0.2358	-
3.3938	8800	0.2363	-
3.4709	9000	0.2335	-
3.5480	9200	0.2329	-
3.6251	9400	0.2301	-
3.7023	9600	0.2334	-
3.7794	9800	0.2301	-
3.8565	10000	0.2309	-
3.9337	10200	0.2291	-
4.0108	10400	0.2268	-
4.0879	10600	0.2212	-
4.1651	10800	0.2224	-
4.2422	11000	0.2224	-
4.3193	11200	0.2211	-
4.3965	11400	0.2194	-
4.4736	11600	0.2192	-
4.5507	11800	0.2183	-
4.6278	12000	0.222	-
4.7050	12200	0.2199	-
4.7821	12400	0.22	-
4.8592	12600	0.2201	-
4.9364	12800	0.2198	-

Framework Versions

Python: 3.11.0
Sentence Transformers: 4.0.1
PyLate: 1.1.7
Transformers: 4.48.2
PyTorch: 2.6.0+cu124
Accelerate: 1.6.0
Datasets: 3.5.0
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"
}

PyLate

bibtex

@misc{PyLate,
title={PyLate: Flexible Training and Retrieval for Late Interaction Models},
author={Chaffin, Antoine and Sourty, Raphaël},
url={https://github.com/lightonai/pylate},
year={2024}
}

	question	answer	negative
type	string	string	string
details	min: 9 tokens mean: 13.12 tokens max: 22 tokens	min: 16 tokens mean: 31.73 tokens max: 32 tokens	min: 15 tokens mean: 31.64 tokens max: 32 tokens