PyLate model based on LiquidAI/LFM2-ColBERT-350M

This is a PyLate model finetuned from LiquidAI/LFM2-ColBERT-350M. 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: LiquidAI/LFM2-ColBERT-350M
Document Length: 8192 tokens
Query Length: 64 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': 8191, 'do_lower_case': False, 'architecture': 'Lfm2Model'})
  (1): Dense({'in_features': 1024, 'out_features': 128, 'bias': False, 'activation_function': 'torch.nn.modules.linear.Identity', 'use_residual': False})
)

Usage

First install the PyLate library:

pip install -U pylate

Retrieval

Use this model with PyLate to index and retrieve documents. The index uses FastPLAID for efficient similarity search.

Indexing documents

Load the ColBERT model and initialize the PLAID index, then encode and index your documents:

from pylate import indexes, models, retrieve

# Step 1: Load the ColBERT model
model = models.ColBERT(
    model_name_or_path="yNilay/colbert-personal-context",
)

# Step 2: Initialize the PLAID index
index = indexes.PLAID(
    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:

# To load an index, simply instantiate it with the correct folder/name and without overriding it
index = indexes.PLAID(
    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:

# 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:

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="yNilay/colbert-personal-context",
)

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	0.9757

Training Details

Training Dataset

Unnamed Dataset

Size: 5,191 training samples
Columns: query, positive, and negative

Approximate statistics based on the first 1000 samples:

	query	positive	negative
type	string	string	string
details	min: 23 tokens mean: 54.59 tokens max: 64 tokens	min: 47 tokens mean: 63.91 tokens max: 64 tokens	min: 5 tokens mean: 39.07 tokens max: 64 tokens

Samples:

query	positive	negative
`Should I tell Sultan to accept Tom Barrack's inauguration invitation? I know Tom's the closest Trump confidant outside the family now, but from what I remember discussing with Michael about going public versus keeping my head down, I'm worried about the optics. Is it even worth it given what I told Sultan about the crowds? I need to decide tonight.`	`Subject: Re: Presidential inauguration --- From: Sultan Bin Sulayem Date: Fri, Jan 6, 2017 at 4:08 AM Should I accept the invitation sent by Tom barrack --- From: jeffrey E. [jeevacation@gmail.com] To: Sultan Bin Sulayem Date: 1/6/2017 2:49:57 PM http://www.cnn.com/2017/01/06/politics/tom-barrack-donald-trump-inauguration/index.html`	`Your reservation at The Four Seasons New York is confirmed for Friday.`
`What's the latest on my inroads to the Trump administration? Last I checked, Thiel was getting close to them and there was talk about him getting a role. Where do things stand now and who's my best access point at this point?`	Subject: Fwd: fact checking questions for New York Magazine story --- From: Yablon, Alex To: jeevacation@gmail.com Date: Tue, Mar 31, 2015 at 11:23 AM Hi Jeffrey, Sorry for the phone tag. In case it's easier to respond by email, I have put my questions below. I'm at my desk and should be here until 5, and should be free on my cell phone after 6:30. Best, Alex -what is the square footage of your Manhattan home? -do you work on a laptop from your dining room, with a large white board for notes and several pairs of reading glasses close at hand? Is the dining room windowless? -do paparazzi often camp outside your home? -did Michael once visit on the same day as a head of state who'd had a police escort? -did you tell Michael about a dinner you'd hosted for six tech entrepreneurs who had a combined worth of several hundred billion dollars? When was this meal? -Do you believe that there are now more people who possess Roosevelt or Carnegie-levels of wealth, the sort that can ri...	Subject: Re: --- From: jeffrey E. jeevacation@gmail.com Date: Sep 20, 2018, at 10:40 AM The Guardian has learned that Rubenfeld is currently the subject of an internal investigation at Yale. The investigation is focused on Rubenfeld's conduct, particularly with female law students. Students have also raised related concerns to Yale authorities about Chua's powerful influence in the clerkships process. The investigation was initiated before Kavanaugh was nominated by Donald Trump to serve on the high court. Rubenfeld said in a statement to the Guardian: "In June, Yale University informed me that it would conduct what it terms an 'informal review' of certain allegations, but that to preserve anonymity, I was not entitled to know any specifics. As a result, I do not know what I am alleged to have said or done. I was further advised that the allegations were not of the kind that would jeopardize my position as a long-tenured member of the faculty." --- From: Lawrence Krauss To: jef...
`What's the current status with Lawrence and those harassment allegations from BuzzFeed? Did he ever get around to writing that point by point refutation we talked about? And how's he doing with the Bulletin now that he took that leave?`	Subject: URGENT: BuzzFeed News inquiry re allegations of sexual harassment --- case, that the University would remove the allegation from my record after 5 years, which makes me surprised that someone violated that written agreement with you. Re item 6: You report on ASU’s response to item #6 , without including the fact that the University specifically stated there were never any allegations of sexual misconduct or harassment by me at the University, and the outside complaints were in fact related specifically to your item #6. Further you neglect to mention that this complaint was by an anonymous third party, not the individual who was allegedly harassed, who never lodged a complain, and that no specific evidence was provided of the alleged transgression. \| was surprised and dismayed that both ASU and ANU launched investigations on the basis of this but was told by both Universities that because of my high profile even such unsubstantiated third party complaints at private event...	Subject: Re: --- From: jeffrey E. jeevacation@gmail.com To: Joi Ito Date: Nov 22, 2017, at 10:44 all good? --- From: Joi Ito To: jeffrey E. jeevacation@gmail.com Date: Wed, Nov 22, 2017 at 10:51 AM Pretty good. Had a pinched nerve in my neck that screwed me up for awhile. How about you? Any plans to come to Boston? --- From: jeffrey E. jeevacation@gmail.com To: Joi Ito Date: Nov 22, 2017, at 10:57 maybe week of 3rd not seeing anything that exciting ornew ? you? with all these guys getting busted for harassment , 1 have moved slightly up on the repuation ladder and have been asked everday for advice etc. this morning I have Ken Starr coming to point out how if clinton cigar lewinsky were to be outed today the world would be a different place --- From: Joi To: jeffrey E. [jeevacation@gmail.com] Date: 11/22/2017 6:14:38 PM Lots of stuff going on that week but I’m in town. #metoo is quite amazing... Madars is doing well. His PhD paper won an award and he found a vulner...

Loss: pylate.losses.contrastive.Contrastive

Evaluation Dataset

Unnamed Dataset

Size: 577 evaluation samples
Columns: query, positive, and negative

Approximate statistics based on the first 577 samples:

	query	positive	negative
type	string	string	string
details	min: 23 tokens mean: 54.08 tokens max: 64 tokens	min: 47 tokens mean: 63.95 tokens max: 64 tokens	min: 5 tokens mean: 36.6 tokens max: 64 tokens

Samples:

query	positive	negative
`A reporter from BuzzFeed just reached out about those old allegations and they're publishing soon. What advice did I give that physicist friend Lawrence a few years back when he was dealing with something similar? I remember telling him something about how these reporters operate and what they do with your responses.`	Subject: Re: --- From: R. Couri Hay To: jeevacation@gmail.com Date: Friday, March 4 2011 07:03 PM Good morning Jeffery, The Newsweek story is being written without your input as you read this. Lloyd Grove is not writing the story, he made a few calls but he was taken off the project. This is for Newsweek, the magazine that is on the stands, not the website. It's still 1200 to 1500 words. Alexandra Wolfe needs to turn in a draft of the story by Monday. I've been subtly guiding Alexandra on your behalf, but would really like to formalize this job with your attorneys and then I would really need to give names and numbers of pro Jeffery power brokers for Alexandra to call. She has already called Donald Trump, Leon Black, and Les Wexner among others. I've given her Jonathans number, please let me focus on this before its too late to spin this in your direction. I'm in all afternoon my office number is My cell is: My email i _________ Cheers, Couri --- From: jeffrey epstein	Subject: Re: --- From: J [jeevacation@gmail.com] To: Michael Wolff Date: On Thu, May 30, 2019 at 5:29 PM is it a coincidence that the russian that bought the house in palm beach and knows all , is the same guy that sold a painting last year to mbs for 450 million dollars. that was only worth 1. 5m? --- From: Michael Wolff To: J jeevacation@gmail.com Date: On Thu, May 30, 2019 at 5:33 PM So MBS was paying him off? Why? Ideas? --- From: J jeevacation@gmail.com To: Michael Wolff Date: On Thu, May 30, 2019 at 5:35 PM reminder trump overuled congress on yemen. --- From: Michael Wolff To: J jeevacation@gmail.com Date: On Thu, May 30, 2019 at 5:37 PM Starting to smell sweet, the way you put it! --- From: J jeevacation@gmail.com To: Michael Wolff Date: On Thu, May 30, 2019 at 5:40 PM In addition my art guyd said the painting wasn't very good --- From: Michael Wolff To: J jeevacation@gmail.com Date: On Thu, May 30, 2019 at 5:41 PM You have an art guy? --- From: ...
`Michael Wolff wants me to do something with him again - he mentioned some big TV interview opportunity. Can you remind me what advice he gave me last time about going public and dealing with the media? I think it was something about going on Charlie Rose and becoming an anti-Trump voice to get political cover. And should I reach out to Kathy about this since she's been helping me understand Washington?`	Subject: Re: --- From: jeffrey E. jeevacation@gmail.com To: Thorbjørn Jagland Date: Feb 17, 2017 9:27 PM Im in paris until thurs? you? --- From: Thorbjørn Jagland To: jeffrey E. jeevacation@gmail.com Date: Feb 19, 2017 8:54 AM Is It possible for you to pass by Strasbourg, it would be great. I really need to understand more about Trump and what's going on in the American society. --- From: jeffrey E. jeevacation@gmail.com To: Thorbjørn Jagland Date: Feb 19, 2017 11:14 yes, that should be possible. remind me how long is the fast train? otherwise ill fly. what days are good for you --- From: Thorbjørn Jagland To: jeffrey E. [jeevacation@gmail.com] Date: 2/19/2017 9:11:12 PM Train, 1h45, I'll pick you up at the train station. Tuesday afternoon is ok. Also Wednesday, but only after 6	`Please find attached the Q3 board meeting minutes for your review.`
`Where do things stand with the media and legal situation now? I have that meeting next week and need to refresh my memory on what we've covered with reporters, our current search results status, and where Michael Wolff left off.`	Subject: Re: Shears Update --- From: Tyler Shears To: jeffrey E. [jeevacation@gmail.com] Date: 7/16/2014 5:56:14 PM yes agree result got worse with recent negative press, clinton, as we discussed we were down to 1 negative (forbes) before that - Christina can confirm this. It will get back to 1 and then 0 so long as negative things stop coming out. if new negative keeps coming out it really dismantles much of our effort... especially when it involves an ex-president no excuses here i'm not pleased with where it is at and am still working to make it happen --- From: jeffrey E. jeevacation@gmail.com Date: Wed, Jul 16, 2014 at 1:25 PM Results still very bad --- From: Christina Galbraith To: Tyler Shears Date: Wednesday, July 16, 2014 Hi Tyler, The social media sites are constantly updated: LinkedIn, Facebook, Twitter, google +. I'll be in touch later today re: feature article. Could you put a site map into the Net site.? This helps rankings. Also could you add it to google ana...	Subject: Re: --- From: jeffrey E. [jeevacation@gmail.com] To: Jonathan Farkas Date: 12/7/2016 12:25:22 P.M. Eastern Standard Time plenty left --- From: Jonathan Farkas Date: Wed, Dec 7, 2016 at 1:01 PM Hi jeffrey hope all is well \| think you are going to have a winderful life from now on in your opinion how much is left in this market it's been a trump triumph \| gave him some money through woody best jonathan --- From: jeffrey E. [jeevacation@gmail.com] To: Jonathan Farkas Date: 12/7/2016 5:30:18 PM oy

Loss: pylate.losses.contrastive.Contrastive

Training Hyperparameters

Non-Default Hyperparameters

eval_strategy: epoch
per_device_train_batch_size: 2
gradient_accumulation_steps: 32
learning_rate: 3e-06
warmup_ratio: 0.1
bf16: True
load_best_model_at_end: True
gradient_checkpointing: True

All Hyperparameters

Click to expand

overwrite_output_dir: False
do_predict: False
eval_strategy: epoch
prediction_loss_only: True
per_device_train_batch_size: 2
per_device_eval_batch_size: 8
per_gpu_train_batch_size: None
per_gpu_eval_batch_size: None
gradient_accumulation_steps: 32
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: 3
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: 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: None
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}
parallelism_config: 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
hub_revision: None
gradient_checkpointing: True
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
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
liger_kernel_config: None
eval_use_gather_object: False
average_tokens_across_devices: False
prompts: None
batch_sampler: batch_sampler
multi_dataset_batch_sampler: proportional
router_mapping: {}
learning_rate_mapping: {}

Training Logs

Epoch	Step	Training Loss	Validation Loss	accuracy
0.1233	10	3.3802	-	-
0.2465	20	1.1614	-	-
0.3698	30	0.5549	-	-
0.4931	40	0.4642	-	-
0.6163	50	0.3749	-	-
0.7396	60	0.3587	-	-
0.8629	70	0.349	-	-
0.9861	80	0.367	-	-
0	0	-	-	0.9723
1.0	82	-	1.0230	-
1.0986	90	0.2812	-	-
1.2219	100	0.3016	-	-
1.3451	110	0.2187	-	-
1.4684	120	0.2104	-	-
1.5917	130	0.2656	-	-
1.7149	140	0.2199	-	-
1.8382	150	0.2356	-	-
1.9615	160	0.2607	-	-
0	0	-	-	0.9757
2.0	164	-	0.9950	-
2.0740	170	0.1836	-	-
2.1972	180	0.1789	-	-
2.3205	190	0.2342	-	-
2.4438	200	0.1904	-	-
2.5670	210	0.1312	-	-
2.6903	220	0.1813	-	-
2.8136	230	0.1826	-	-
2.9368	240	0.1456	-	-
0	0	-	-	0.9757
3.0	246	-	0.873	-

The bold row denotes the saved checkpoint.

Framework Versions

Python: 3.11.12
Sentence Transformers: 5.1.1
PyLate: 1.4.0
Transformers: 4.56.2
PyTorch: 2.9.0+cu128
Accelerate: 1.13.0
Datasets: 4.8.4
Tokenizers: 0.22.2

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"
}

PyLate

@inproceedings{DBLP:conf/cikm/ChaffinS25,
  author       = {Antoine Chaffin and
                  Rapha{"{e}}l Sourty},
  editor       = {Meeyoung Cha and
                  Chanyoung Park and
                  Noseong Park and
                  Carl Yang and
                  Senjuti Basu Roy and
                  Jessie Li and
                  Jaap Kamps and
                  Kijung Shin and
                  Bryan Hooi and
                  Lifang He},
  title        = {PyLate: Flexible Training and Retrieval for Late Interaction Models},
  booktitle    = {Proceedings of the 34th {ACM} International Conference on Information
                  and Knowledge Management, {CIKM} 2025, Seoul, Republic of Korea, November
                  10-14, 2025},
  pages        = {6334--6339},
  publisher    = {{ACM}},
  year         = {2025},
  url          = {https://github.com/lightonai/pylate},
  doi          = {10.1145/3746252.3761608},
}

Downloads last month: -

Safetensors

Model size

0.4B params

Tensor type

F32

Model tree for yNilay/colbert-personal-context

Base model

LiquidAI/LFM2-ColBERT-350M

Finetuned

(1)

this model

Paper for yNilay/colbert-personal-context

Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks

Paper • 1908.10084 • Published Aug 27, 2019 • 12

Evaluation results

Accuracy on Unknown
self-reported

0.976