This is a PyLate model finetuned from answerdotai/answerai-colbert-small-v1. It maps sentences & paragraphs to sequences of 96-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/answerai-colbert-small-v1
Document Length: 300 tokens
Query Length: 32 tokens
Output Dimensionality: 96 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: BertModel 
  (1): Dense({'in_features': 384, 'out_features': 96, '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-answerai-colbert-small-v1-1-neg-1-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-answerai-colbert-small-v1-1-neg-1-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	0.5194

Training Details

Training Dataset

Unnamed Dataset

Size: 1,893,947 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: 12.91 tokens
max: 24 tokens
min: 16 tokens
mean: 31.72 tokens
max: 32 tokens
min: 12 tokens
mean: 31.75 tokens
max: 32 tokens

Samples:

question	answer	negative
`what are girl names that start with g?`	`['Grace. Grace is one of those classic girl names that never seems to go out of style. ... ', 'Gabriella. Hebrew Origin. ... ', 'Gabrielle. French Origin. ... ', 'Genevieve. French: White wave. ... ', 'Giselle. French: Pledge. ... ', 'Gloria. Latin: Glory, renown, and respect. ... ', 'Gina. English: Abbreviation of names ending in "gina" ... ', 'Gabriela.']`
`is wheat healthier than white bread?`	`Whole wheat is processed to include all three nutritious parts, but white flour uses only the endosperm. When put head-to-head with whole wheat bread, white is a nutritional lightweight. Whole wheat is much higher in fiber, vitamins B6 and E, magnesium, zinc, folic acid and chromium.`	`Very Strong White. Made from a blend of premium wheat, Allinson Very Strong White Bread Flour has a higher protein content and gluten strength than our strong flour. ... You can use this flour whenever your recipes call for strong white bread flour and it will produce a higher rise and a better texture.`
`are the mtv vmas live?`	`The annual VMA ceremony occurs before the end of summer and held either in late August or mid-September, and broadcast live on MTV, along with a "roadblock" simulcast across MTV's sister networks since 2014, which is utilized to maximize the ceremony's ratings.`	`TIL MTV and VH1 were always a part of the same company and are in fact SISTER CHANNELS. Viacom is huge, they also own Comedy Central. Viacom bought MTV, VH1, BET, Nickelodeon (MTV Networks) and Showtime from Warner in 1985.`

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: 12.89 tokens
max: 24 tokens
min: 17 tokens
mean: 31.76 tokens
max: 32 tokens
min: 14 tokens
mean: 31.41 tokens
max: 32 tokens

Samples:

question	answer	negative_1
`what season and episode does jon snow come back to life?`	`This is the mother of all Game of Thrones spoilers, so if you haven't watched season 6 episode 2 yet, please click away now. Jon Snow is alive.`	`Then, of course, there is Jon Snow's resurrection in Episode 2 of Season 6, an event that changed everything about Thrones.`
`do the rich get richer?`	`It's true, the rich do get richer—here's why... ... The growing wealth of the rich, and the relative stagnation of the middle class, are due in large part to diverging incomes, but investments also play an increasingly important role. And it's not just that the wealthy have more investments.`	`It's true, the rich do get richer—here's why... ... The growing wealth of the rich, and the relative stagnation of the middle class, are due in large part to diverging incomes, but investments also play an increasingly important role. And it's not just that the wealthy have more investments.`
`how long can you eat leftover lamb?`	`Store it safely You can safely store cooked lamb for up to three days in the fridge, or for up to two months in the freezer. Make sure it's fully defrosted before using and, if it's been frozen once, don't re-freeze.`	`Lamb. ... The meat of sheep 6 to 10 weeks old is usually sold as baby lamb, and spring lamb is from sheep of five to six months.`

Loss: pylate.losses.contrastive.Contrastive

Training Hyperparameters

Non-Default Hyperparameters

eval_strategy: steps
per_device_train_batch_size: 128
per_device_eval_batch_size: 128
learning_rate: 3e-06
num_train_epochs: 1
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: 128
per_device_eval_batch_size: 128
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: 1
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: False
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	-	0.5194
0.0001	1	3.319	-
0.0135	200	2.6741	-
0.0270	400	0.6117	-
0.0405	600	0.3176	-
0.0541	800	0.2894	-
0.0676	1000	0.2731	-
0.0811	1200	0.2621	-
0.0946	1400	0.2539	-
0.1081	1600	0.2399	-
0.1216	1800	0.238	-
0.1352	2000	0.2309	-
0.1487	2200	0.2345	-
0.1622	2400	0.2192	-
0.1757	2600	0.2155	-
0.1892	2800	0.2218	-
0.2027	3000	0.2214	-
0.2163	3200	0.2136	-
0.2298	3400	0.2136	-
0.2433	3600	0.2132	-
0.2568	3800	0.2104	-
0.2703	4000	0.212	-
0.2838	4200	0.2156	-
0.2974	4400	0.2034	-
0.3109	4600	0.2083	-
0.3244	4800	0.2082	-
0.3379	5000	0.2126	-
0.3514	5200	0.2098	-
0.3649	5400	0.2055	-
0.3785	5600	0.2024	-
0.3920	5800	0.1974	-
0.4055	6000	0.2005	-
0.4190	6200	0.205	-
0.4325	6400	0.2002	-
0.4460	6600	0.1989	-
0.4596	6800	0.1997	-
0.4731	7000	0.1954	-
0.4866	7200	0.192	-
0.5001	7400	0.2003	-
0.5136	7600	0.1942	-
0.5271	7800	0.1998	-
0.5407	8000	0.1985	-
0.5542	8200	0.1934	-
0.5677	8400	0.1951	-
0.5812	8600	0.1946	-
0.5947	8800	0.1919	-
0.6082	9000	0.2025	-
0.6217	9200	0.1924	-
0.6353	9400	0.1964	-
0.6488	9600	0.192	-
0.6623	9800	0.1884	-
0.6758	10000	0.1921	-
0.6893	10200	0.1922	-
0.7028	10400	0.1973	-
0.7164	10600	0.1911	-
0.7299	10800	0.191	-
0.7434	11000	0.1907	-
0.7569	11200	0.1866	-
0.7704	11400	0.1905	-
0.7839	11600	0.1936	-
0.7975	11800	0.193	-
0.8110	12000	0.1869	-
0.8245	12200	0.192	-
0.8380	12400	0.1889	-
0.8515	12600	0.1923	-
0.8650	12800	0.1902	-
0.8786	13000	0.1879	-
0.8921	13200	0.1918	-
0.9056	13400	0.1938	-
0.9191	13600	0.1927	-
0.9326	13800	0.1926	-
0.9461	14000	0.1939	-
0.9597	14200	0.1909	-
0.9732	14400	0.1878	-
0.9867	14600	0.1919	-

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: 12.91 tokens max: 24 tokens	min: 16 tokens mean: 31.72 tokens max: 32 tokens	min: 12 tokens mean: 31.75 tokens max: 32 tokens