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': 179, '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-5-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-5-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	0.5022

Training Details

Training Dataset

Unnamed Dataset

Size: 9,383,917 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: 13.3 tokens
max: 21 tokens
min: 16 tokens
mean: 31.77 tokens
max: 32 tokens
min: 14 tokens
mean: 31.54 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.`
`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.`	`Most of all, there's AJ, the infant son of Clementine's ally Rebecca, who Clementine promised to raise when Rebecca died back in Season Two. The Final Season rejoins Clementine and AJ, now around six years old, on the open road.`
`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.`	`Clementines — commonly known by the brand names Cuties or Halos — are a hybrid of mandarin and sweet oranges. These tiny fruits are bright orange, easy to peel, sweeter than most other citrus fruits, and typically seedless.`

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

Click to expand

Epoch	Step	Training Loss	Validation Loss	accuracy
0	0	-	-	0.4560
0.0002	1	22.6729	-	-
0.0307	200	16.3893	-	-
0.0614	400	7.1556	-	-
0.0921	600	4.4451	-	-
0.1228	800	1.8384	-	-
0.1535	1000	1.0792	-	-
0.1842	1200	0.8636	-	-
0.2149	1400	0.7355	-	-
0.2455	1600	0.6498	-	-
0.2762	1800	0.5801	-	-
0.3069	2000	0.5318	-	-
0.3376	2200	0.49	-	-
0.3683	2400	0.4515	-	-
0.3990	2600	0.4245	-	-
0.4297	2800	0.3929	-	-
0.4604	3000	0.3704	-	-
0.4911	3200	0.3505	-	-
0.5218	3400	0.3294	-	-
0.5525	3600	0.3114	-	-
0.5832	3800	0.297	-	-
0.6139	4000	0.281	-	-
0.6446	4200	0.2723	-	-
0.6753	4400	0.2589	-	-
0.7060	4600	0.2518	-	-
0.7366	4800	0.2437	-	-
0.7673	5000	0.2333	-	-
0.7980	5200	0.2285	-	-
0.8287	5400	0.2236	-	-
0.8594	5600	0.2144	-	-
0.8901	5800	0.2122	-	-
0.9208	6000	0.2093	-	-
0.9515	6200	0.2015	-	-
0.9822	6400	0.1984	-	-
1.0129	6600	0.1936	-	-
1.0436	6800	0.1885	-	-
1.0743	7000	0.1841	-	-
1.1050	7200	0.1818	-	-
1.1357	7400	0.1805	-	-
1.1664	7600	0.1774	-	-
1.1971	7800	0.1742	-	-
1.2277	8000	0.1722	-	-
1.2584	8200	0.1679	-	-
1.2891	8400	0.1671	-	-
1.3198	8600	0.1646	-	-
1.3505	8800	0.1639	-	-
1.3812	9000	0.161	-	-
1.4119	9200	0.1604	-	-
1.4426	9400	0.1585	-	-
1.4733	9600	0.1562	-	-
1.5040	9800	0.1548	-	-
1.5347	10000	0.1528	-	-
1.5654	10200	0.1519	-	-
1.5961	10400	0.1492	-	-
1.6268	10600	0.149	-	-
1.6575	10800	0.1481	-	-
1.6882	11000	0.1473	-	-
1.7188	11200	0.1467	-	-
1.7495	11400	0.1448	-	-
1.7802	11600	0.1413	-	-
1.8109	11800	0.142	-	-
1.8416	12000	0.1398	-	-
1.8723	12200	0.1385	-	-
1.9030	12400	0.1398	-	-
1.9337	12600	0.1375	-	-
1.9644	12800	0.1376	-	-
1.9951	13000	0.1369	-	-
2.0258	13200	0.1303	-	-
2.0565	13400	0.1305	-	-
2.0872	13600	0.1286	-	-
2.1179	13800	0.1266	-	-
2.1486	14000	0.1273	-	-
2.1793	14200	0.1269	-	-
2.2099	14400	0.1253	-	-
2.2406	14600	0.1263	-	-
2.2713	14800	0.1249	-	-
2.3020	15000	0.1248	-	-
2.3327	15200	0.1227	-	-
2.3634	15400	0.1239	-	-
2.3941	15600	0.1233	-	-
2.4248	15800	0.1211	-	-
2.4555	16000	0.1208	-	-
2.4862	16200	0.1206	-	-
2.5169	16400	0.1211	-	-
2.5476	16600	0.1209	-	-
2.5783	16800	0.1195	-	-
2.6090	17000	0.1192	-	-
2.6397	17200	0.1176	-	-
2.6703	17400	0.1177	-	-
2.7010	17600	0.1168	-	-
2.7317	17800	0.1163	-	-
2.7624	18000	0.116	-	-
2.7931	18200	0.1165	-	-
2.8238	18400	0.1157	-	-
2.8545	18600	0.1145	-	-
2.8852	18800	0.1154	-	-
2.9159	19000	0.1153	-	-
2.9466	19200	0.1132	-	-
2.9773	19400	0.1128	-	-
3.0080	19600	0.1121	-	-
3.0387	19800	0.1099	-	-
3.0694	20000	0.1087	-	-
0	0	-	-	0.5022
3.0694	20000	-	1.1151	-
3.1001	20200	0.1086	-	-
3.1308	20400	0.108	-	-
3.1614	20600	0.1087	-	-
3.1921	20800	0.1084	-	-
3.2228	21000	0.1072	-	-
3.2535	21200	0.1087	-	-
3.2842	21400	0.1067	-	-
3.3149	21600	0.1073	-	-
3.3456	21800	0.1067	-	-
3.3763	22000	0.1045	-	-
3.4070	22200	0.105	-	-
3.4377	22400	0.1046	-	-
3.4684	22600	0.1061	-	-
3.4991	22800	0.1043	-	-
3.5298	23000	0.105	-	-
3.5605	23200	0.105	-	-
3.5912	23400	0.1047	-	-
3.6219	23600	0.1034	-	-
3.6525	23800	0.1037	-	-
3.6832	24000	0.1042	-	-
3.7139	24200	0.1038	-	-
3.7446	24400	0.1039	-	-
3.7753	24600	0.1031	-	-
3.8060	24800	0.1019	-	-
3.8367	25000	0.1023	-	-
3.8674	25200	0.1036	-	-
3.8981	25400	0.1022	-	-
3.9288	25600	0.102	-	-
3.9595	25800	0.1022	-	-
3.9902	26000	0.1017	-	-
4.0209	26200	0.0997	-	-
4.0516	26400	0.0992	-	-
4.0823	26600	0.0993	-	-
4.1130	26800	0.099	-	-
4.1436	27000	0.098	-	-
4.1743	27200	0.0986	-	-
4.2050	27400	0.0987	-	-
4.2357	27600	0.0993	-	-
4.2664	27800	0.0991	-	-
4.2971	28000	0.0993	-	-
4.3278	28200	0.098	-	-
4.3585	28400	0.0979	-	-
4.3892	28600	0.0967	-	-
4.4199	28800	0.0983	-	-
4.4506	29000	0.0976	-	-
4.4813	29200	0.0975	-	-
4.5120	29400	0.0979	-	-
4.5427	29600	0.0971	-	-
4.5734	29800	0.0972	-	-
4.6041	30000	0.0969	-	-
4.6347	30200	0.0972	-	-
4.6654	30400	0.0975	-	-
4.6961	30600	0.0987	-	-
4.7268	30800	0.0964	-	-
4.7575	31000	0.0974	-	-
4.7882	31200	0.0964	-	-
4.8189	31400	0.0974	-	-
4.8496	31600	0.0974	-	-
4.8803	31800	0.0975	-	-
4.9110	32000	0.097	-	-
4.9417	32200	0.0973	-	-
4.9724	32400	0.0973	-	-

The bold row denotes the saved checkpoint.

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: 13.3 tokens max: 21 tokens	min: 16 tokens mean: 31.77 tokens max: 32 tokens	min: 14 tokens mean: 31.54 tokens max: 32 tokens