initial-colbert-ir

PyLate model based on colbert-ir/colbertv2.0

This is a PyLate model finetuned from colbert-ir/colbertv2.0 on the reasonir-data dataset. 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: colbert-ir/colbertv2.0 <!-- at revision c1e84128e85ef755c096a95bdb06b47793b13acf -->
• Document Length: 180 tokens
• Query Length: 32 tokens
• Output Dimensionality: 128 tokens
• Similarity Function: MaxSim
• Training Dataset:
  • reasonir-data
• Language: en

<!-- - **License:** Unknown -->

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: BertModel 
  (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:

from pylate import indexes, models, retrieve

# Step 1: Load the ColBERT model
model = models.ColBERT(
    model_name_or_path=pylate_model_id,
)

# 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:

# 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:

# 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=pylate_model_id,
)

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,
)

<!-- ### Direct Usage (Transformers) <details><summary>Click to see the direct usage in Transformers</summary> </details> --> <!-- ### Downstream Usage (Sentence Transformers) You can finetune this model on your own dataset. <details><summary>Click to expand</summary> </details> --> <!-- ### Out-of-Scope Use *List how the model may foreseeably be misused and address what users ought not to do with the model.* -->

Evaluation

Metrics

Col BERTTriplet

• Evaluated with <code>pylate.evaluation.colbert_triplet.ColBERTTripletEvaluator</code>

<!-- ## Bias, Risks and Limitations *What are the known or foreseeable issues stemming from this model? You could also flag here known failure cases or weaknesses of the model.* --> <!-- ### Recommendations *What are recommendations with respect to the foreseeable issues? For example, filtering explicit content.* -->

Training Details

Training Dataset

reasonir-data

• Dataset: reasonir-data at 0275f82
• Size: 242,513 training samples
• Columns: <code>query</code>, <code>positive</code>, and <code>negative</code>
• Approximate statistics based on the first 1000 samples:

  	query	positive	negative
  type	string	string	string
  details	<ul><li>min: 6 tokens</li><li>mean: 26.76 tokens</li><li>max: 32 tokens</li></ul>	<ul><li>min: 4 tokens</li><li>mean: 19.89 tokens</li><li>max: 32 tokens</li></ul>	<ul><li>min: 4 tokens</li><li>mean: 20.32 tokens</li><li>max: 32 tokens</li></ul>

• Samples:

  query	positive	negative
  <code>Market analysis suggests that the ongoing trend of increased adoption of renewable energy sources will continue to drive the demand for solar panels in the coming years. According to various reports, the global solar panel market is projected to witness significant growth over the next decade, with some analysts predicting a compound annual growth rate (CAGR) of up to 20%. This growth is expected to be fueled by declining costs of production, government incentives, and growing environmental concerns. However, some experts also caution that the market may experience fluctuations due to trade policies, technological advancements, and changes in government regulations. Despite these challenges, the overall outlook for the solar panel market remains optimistic, with many companies investing heavily in research and development to improve efficiency and reduce costs. As the demand for renewable energy continues to rise, it is likely that the solar panel market will play a significant role in...</code>	<code>Contrary to the market analysis suggesting a compound annual growth rate (CAGR) of up to 20% for the global solar panel market, our financial reports indicate a more modest growth rate of 12% over the next decade. While we agree that declining production costs, government incentives, and growing environmental concerns will drive demand for solar panels, we also believe that trade policies, technological advancements, and changes in government regulations will have a more significant impact on the market than previously anticipated. Our projections suggest that the market will experience fluctuations, with some years experiencing higher growth rates than others. However, we do not anticipate the market to experience the same level of growth as predicted by other analysts. Our research indicates that the market will reach a saturation point, beyond which growth will slow down. Additionally, we believe that the impact of advancements in energy storage technologies on the solar panel marke...</code>	<code>The demand for solar panels has been on the rise in recent years, driven by an increase in environmental awareness and the need for sustainable energy sources. One of the key factors contributing to this growth is the decline in production costs. As technology advances, the cost of manufacturing solar panels has decreased, making them more affordable for consumers. Additionally, governments around the world have implemented policies and incentives to encourage the adoption of renewable energy sources, which has further boosted demand for solar panels. However, the solar panel market is not without its challenges. Trade policies and technological advancements can impact the market, and changes in government regulations can create uncertainty. Despite these challenges, the outlook for the solar panel market remains positive, with many companies investing heavily in research and development to improve efficiency and reduce costs. The development of new technologies, such as bifacial panel...</code>
  <code>As the sun set over the vast savannah, a sense of tranquility washed over the pride of lions. Their tawny coats glistened in the fading light, and the sound of crickets provided a soothing background hum. Nearby, a group of humans, armed with cameras and curiosity, observed the wild animals from a safe distance.</code>	<code>The lions lazed in the shade of a nearby tree, their tawny coats a blur as they basked in the warmth. The visitors watched in awe, clicking away at their cameras to capture the majesty of the wild animals. Crickets provided a constant, soothing background noise as the humans took care to keep a safe distance from the pride.</code>	<code>The city's tree planting initiative has been a huge success, providing a serene oasis in the midst of the bustling metropolis. The sounds of the city – car horns, chatter and crickets – blend together to create a symphony of noise. While many humans have been drawn to the tranquility of the park, others have raised concerns about the integration of urban wildlife.</code>
  <code>Recent advancements in the field of artificial intelligence have led to significant breakthroughs in natural language processing. This has far-reaching implications for various industries, including education, where AI-powered chatbots can enhance student learning experiences by providing personalization and real-time feedback. Moreover, the integration of AI in educational settings can help address issues of accessibility and equity.</code>	<code>The rapid expansion of AI research has yielded substantial progress in natural language processing, allowing for the development of more sophisticated AI-powered tools. In the education sector, AI-driven chatbots can facilitate individualized learning and offer instantaneous feedback, thereby enriching the overall learning environment. However, it is crucial to address concerns surrounding the digital divide to ensure that these technological advancements are accessible to all.</code>	<code>One of the primary challenges facing archaeologists today is the authentication of ancient artifacts, which often involves meticulous analysis of relics and literary texts. The discovery of a previously unknown scroll, buried deep within the labyrinthine passages of an Egyptian tomb, shed new light on the role of language in ancient cultures. Interestingly, the sophisticated syntax and nuanced vocabulary of the ancient Egyptian language have some similarities with modern-day linguistic structures.</code>

• Loss: <code>pylate.losses.contrastive.Contrastive</code>

Evaluation Dataset

reasonir-data

• Dataset: reasonir-data at 0275f82
• Size: 2,450 evaluation samples
• Columns: <code>query</code>, <code>positive</code>, and <code>negative</code>
• Approximate statistics based on the first 1000 samples:

  	query	positive	negative
  type	string	string	string
  details	<ul><li>min: 6 tokens</li><li>mean: 26.92 tokens</li><li>max: 32 tokens</li></ul>	<ul><li>min: 4 tokens</li><li>mean: 19.98 tokens</li><li>max: 32 tokens</li></ul>	<ul><li>min: 4 tokens</li><li>mean: 20.5 tokens</li><li>max: 32 tokens</li></ul>

• Samples:

  query	positive	negative
  <code>In the recent speech, the politician claimed that the current tax reform will benefit the middle class and lead to a significant increase in economic growth. The politician stated, 'Our plan is to cut taxes across the board, simplifying the tax code and making it fairer for all Americans. We're taking the money from the unfair and complex system and putting it back in the pockets of the hardworking people.' The politician also emphasized the effects of the cut taxes, 'When we cut taxes, we're putting more money in the hands of our small business owners and workers, who are the backbone of our economy. This means new jobs will be created, wages will rise, and the people who actually create the jobs, the entrepreneurs, will have the funds needed to invest more into their businesses.' Moreover, the politician asserted, 'Our country will witness a major boost in job creation and economic growth which in turn will positively affect local communities all around the country.' Furthermore, the...</code>	<code>According to reputable sources in economics research, the tax reform that the administration is trying to implement closely resembles that of the 2001 and the 2003 cuts under former President George Bush and that of 1981 under President Ronald Reagan who reduced tax rates 23% and 25% respectively. Research carried out by a major university indicated that these reforms only yielded an estimated 10% increase in tax revenue, since the decrease in tax income could result in compensated revenues through economic stimulation. Some studies were actually pointing to the idea that no trickle-down economics apply as more funds were being placed in the already wealthy communities. This change could shift the economical inequalities to an extreme and showed a direct relationship between a tax reduction and a large national deficit increase. Employer demand for the borderline employee may not actually increase from the creation of the new jobs, and economists believed. The variation of wages for jo...</code>	<code>The concept of a universal basic income has been a topic of discussion among economists and policymakers in recent years. While some see it as a viable solution to poverty and economic inequality, others argue that it is not feasible due to the financial constraints it would impose on governments. One of the main concerns is that implementing a universal basic income would require significant funding, which would likely come from increased taxes or redistribution of existing social welfare funds. Critics argue that this could lead to a decrease in economic growth, as people may be less incentivized to work if they are receiving a guaranteed income. On the other hand, proponents argue that a universal basic income would provide a safety net for the most vulnerable members of society and allow people to pursue meaningful work rather than just taking any job for the sake of a paycheck. Some countries have experimented with universal basic income pilots, but the results have been mixed. Fi...</code>
  <code>Recent advances in super-resolution techniques have led to a greater understanding of many sub-cellular structures and have opened up new avenues for exploring cellular behavior at the nanoscale. Fluorescence imaging, in particular, has greatly benefited from these advances and has enabled researchers to visualize the distribution and dynamics of proteins in real time. However, further developments in fluorescence imaging rely on a better comprehension of the interactions between imaging probes and their molecular environment. A crucial factor in these interactions is the size and shape of the probes, which must be optimized to minimize disruption of the native dynamics of the system while also achieving high fluorescence yields. The DNA-based probes have emerged as a promising solution, offering the opportunity to tune the size and shape of the probes to optimize performance.</code>	<code>Microscopy</code>	<code>Biophysics</code>
  <code>I recently purchased this top-of-the-line smartwatch for my birthday, and I must say that it has been a revelation in terms of keeping track of my vital signs and daily activity levels. The watch has an elegant design that doesn't clash with my other wearable accessories, and I love how the touchscreen display lets me access a wealth of health metrics at a glance. Although I've encountered several instances where the heart rate monitoring system was delayed in capturing accurate readings, this minor shortcoming hardly detracts from my overall satisfaction with the product. The value proposition it presents in terms of quality, accuracy, and ascendancy over competing offerings makes it a compelling option in this class of devices. Despite never having owned one before, I found the smartwatch straightforward to use, and the companion app did an excellent job of simplifying the tracking and analysis of my fitness journey. Nothing in particular distinguishes this product's methodology in c...</code>	<code>authentic</code>	<code>fake</code>

• Loss: <code>pylate.losses.contrastive.Contrastive</code>

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 16
• per_device_eval_batch_size: 32
• gradient_accumulation_steps: 2
• learning_rate: 2e-05
• weight_decay: 0.01
• warmup_steps: 100
• fp16: True
• remove_unused_columns: False

All Hyperparameters

<details><summary>Click to expand</summary>

• overwrite_output_dir: False
• do_predict: False
• eval_strategy: steps
• prediction_loss_only: True
• per_device_train_batch_size: 16
• per_device_eval_batch_size: 32
• per_gpu_train_batch_size: None
• per_gpu_eval_batch_size: None
• gradient_accumulation_steps: 2
• eval_accumulation_steps: None
• torch_empty_cache_steps: None
• learning_rate: 2e-05
• weight_decay: 0.01
• 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: 0.0
• warmup_steps: 100
• 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: False
• fp16: True
• 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: False
• label_names: None
• load_best_model_at_end: False
• 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

</details>

Training Logs

<details><summary>Click to expand</summary> </details>

Framework Versions

• Python: 3.12.4
• Sentence Transformers: 4.0.2
• PyLate: 1.2.0
• Transformers: 4.48.2
• PyTorch: 2.6.0+cu124
• Accelerate: 1.7.0
• Datasets: 3.6.0
• Tokenizers: 0.21.1

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

@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}
}

<!-- ## Glossary *Clearly define terms in order to be accessible across audiences.* --> <!-- ## Model Card Authors *Lists the people who create the model card, providing recognition and accountability for the detailed work that goes into its construction.* --> <!-- ## Model Card Contact *Provides a way for people who have updates to the Model Card, suggestions, or questions, to contact the Model Card authors.* -->