End of training

72b1c49 verified 8 days ago

55.6 kB

tags:
  - ColBERT
  - PyLate
  - sentence-transformers
  - sentence-similarity
  - feature-extraction
  - generated_from_trainer
  - dataset_size:1188486
  - loss:Contrastive
base_model: jhu-clsp/ettin-encoder-17m
datasets:
  - benjamintli/code-retrieval-combined-v2-llm-negatives
pipeline_tag: sentence-similarity
library_name: PyLate

PyLate model based on jhu-clsp/ettin-encoder-17m

This is a PyLate model finetuned from jhu-clsp/ettin-encoder-17m on the code-retrieval-combined-v2-llm-negatives 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: jhu-clsp/ettin-encoder-17m
Document Length: 180 tokens
Query Length: 32 tokens
Output Dimensionality: 128 tokens
Similarity Function: MaxSim
Training Dataset:
- code-retrieval-combined-v2-llm-negatives

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, 'architecture': 'ModernBertModel'})
  (1): Dense({'in_features': 256, '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="colbert-code-17m",
)

# 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="colbert-code-17m",
)

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,
)

Training Details

Training Dataset

code-retrieval-combined-v2-llm-negatives

Dataset: code-retrieval-combined-v2-llm-negatives at 1917069
Size: 1,188,486 training samples
Columns: query, positive, source, hard_negatives, and negatives

Approximate statistics based on the first 1000 samples:

	query	positive	source	hard_negatives	negatives
type	string	string	string	list	string
details	min: 6 tokens mean: 23.61 tokens max: 32 tokens	min: 14 tokens mean: 31.43 tokens max: 32 tokens	min: 4 tokens mean: 7.46 tokens max: 10 tokens	size: 1 elements	min: 16 tokens mean: 31.15 tokens max: 32 tokens

Samples:

query	positive	source	hard_negatives	negatives
`wait for AWS PCA CSR propagation before issuing certificate`	`func (c ACMPCA) WaitUntilCertificateAuthorityCSRCreated(input GetCertificateAuthorityCsrInput) error { return c.WaitUntilCertificateAuthorityCSRCreatedWithContext(aws.BackgroundContext(), input) }`	`csn_syntethic`	`['func (c ACMPCA) WaitUntilAuditReportCreated(input DescribeCertificateAuthorityAuditReportInput) error {\n\treturn c.WaitUntilAuditReportCreatedWithContext(aws.BackgroundContext(), input)\n}']`	`func (c ACMPCA) WaitUntilAuditReportCreated(input DescribeCertificateAuthorityAuditReportInput) error { return c.WaitUntilAuditReportCreatedWithContext(aws.BackgroundContext(), input) }`
`func (f Filter) Gather() ([]dto.MetricFamily, error) { mfs, err := f.Gatherer.Gather() if err != nil {`	`return nil, err } return f.Matcher.Match(mfs), nil }`	`csn_ccr`	`['\n\tf.err = err\n\tif err != nil {\n\t\tfor _, node := range f.cells {\n\t\t\tnode.PropagateWatchError(err)\n\t\t}\n\t}\n}']`	`f.err = err if err != nil { for _, node := range f.cells { node.PropagateWatchError(err) } } }`
def latent_to_dist(name, x, hparams, output_channels=None): """Map latent to the mean and log-scale of a Gaussian. Args: name: variable scope. x: 4-D Tensor of shape (NHWC) hparams: HParams. latent_architecture - can be "single_conv", "glow_nn" or "glow_resnet", default = single_conv latent_encoder_depth - int, depth of architecture, valid if latent_architecture is "glow_nn" or "glow_resnet". latent_pre_output_channels - 512, valid only when latent_architecture is "glow_nn". latent_encoder_width - 512, maximum width of the network output_channels: int, number of output channels of the mean (and std). if not provided, set it to be the output channels of x. Returns: dist: instance of tfp.distributions.Normal Raises: ValueError: If architecture not in ["single_conv", "glow_nn"] """ architecture = hparams.get("latent_a...	mid_channels=mid_channels) mean_log_scale = conv("glow_nn_zeros", mean_log_scale, filter_size=[3, 3], stride=[1, 1], output_channels=2output_channels, apply_actnorm=False, conv_init="zeros") elif architecture == "glow_resnet": h = x for layer in range(depth): h3 = conv_stack("latent_resnet_%d" % layer, h, mid_channels=width, output_channels=x_shape[-1], dropout=hparams.coupling_dropout) h += h3 mean_log_scale = conv("glow_res_final", h, conv_init="zeros", output_channels=2output_channels, apply_actnorm=False) else: raise ValueError("expected architecture to be single_conv or glow_nn " "got %s" % architecture) mean = mean_log_scale[:, :, :, 0::2] log_scale = mean_log_scale[:, :, :, 1::2] return tfp.distribu...	`csn_ccr`	`[' first_relu=False,\n padding="SAME",\n strides=(2, 2),\n force2d=True,\n name="conv0")\n x = common_layers.conv_block(\n x, 64, [((1, 1), (3, 3))], padding="SAME", force2d=True, name="conv1")\n x = xnet_resblock(x, min(128, hidden_dim), True, "block0")\n x = xnet_resblock(x, min(256, hidden_dim), False, "block1")\n return xnet_resblock(x, hidden_dim, False, "block2")']`	`first_relu=False, padding="SAME", strides=(2, 2), force2d=True, name="conv0") x = common_layers.conv_block( x, 64, [((1, 1), (3, 3))], padding="SAME", force2d=True, name="conv1") x = xnet_resblock(x, min(128, hidden_dim), True, "block0") x = xnet_resblock(x, min(256, hidden_dim), False, "block1") return xnet_resblock(x, hidden_dim, False, "block2")`

Loss: pylate.losses.contrastive.Contrastive

Evaluation Dataset

code-retrieval-combined-v2-llm-negatives

Dataset: code-retrieval-combined-v2-llm-negatives at 1917069
Size: 12,005 evaluation samples
Columns: query, positive, source, hard_negatives, and negatives

Approximate statistics based on the first 1000 samples:

	query	positive	source	hard_negatives	negatives
type	string	string	string	list	string
details	min: 6 tokens mean: 23.78 tokens max: 32 tokens	min: 14 tokens mean: 31.45 tokens max: 32 tokens	min: 4 tokens mean: 7.55 tokens max: 10 tokens	size: 1 elements	min: 14 tokens mean: 31.15 tokens max: 32 tokens

Samples:

query	positive	source	hard_negatives	negatives
`public static function list_templates($competencyid, $onlyvisible) { global $DB; $sql = 'SELECT tpl.* FROM {' . template::TABLE . '} tpl JOIN {' . self::TABLE . '} tplcomp ON tplcomp.templateid = tpl.id WHERE tplcomp.competencyid = ? '; $params = array($competencyid); if ($onlyvisible) { $sql .= ' AND tpl.visible = ?'; $params[] = 1; }`	`$sql .= ' ORDER BY tpl.id ASC'; $results = $DB->get_records_sql($sql, $params); $instances = array(); foreach ($results as $result) { array_push($instances, new template(0, $result)); } return $instances; }`	`csn_ccr`	`[" FROM {assign_grades} g\n JOIN(' . $esql . ') e ON e.id = g.userid\n WHERE g.assignment = :assignid';\n\n return $DB->count_records_sql($sql, $params);\n }"]`	`FROM {assign_grades} g JOIN(' . $esql . ') e ON e.id = g.userid WHERE g.assignment = :assignid'; return $DB->count_records_sql($sql, $params); }`
`python multiprocessing sandboxed process isolation file system network`	def start(self): '''Create a process in which the isolated code will be run.''' assert self._client is None logger.debug('IsolationContext[%d] starting', id(self)) # Create the queues request_queue = multiprocessing.Queue() response_queue = multiprocessing.Queue() # Launch the server process server = Server(request_queue, response_queue) # Do not keep a reference to this object! server_process = multiprocessing.Process(target=server.loop) server_process.start() # Create a client to talk to the server self._client = Client(server_process, request_queue, response_queue)	`csn_syntethic`	['def start(cls, _init_logging=True):\n """\n Arrange for the subprocess to be started, if it is not already running.\n\n The parent process picks a UNIX socket path the child will use prior to\n fork, creates a socketpair used essentially as a semaphore, then blocks\n waiting for the child to indicate the UNIX socket is ready for use.\n\n :param bool _init_logging:\n For testing, if :data:False, don't initialize logging.\n """\n if cls.worker_sock is not None:\n return\n\n if faulthandler is not None:\n faulthandler.enable()\n\n mitogen.utils.setup_gil()\n cls.unix_listener_path = mitogen.unix.make_socket_path()\n cls.worker_sock, cls.child_sock = socket.socketpair()\n atexit.register(lambda: clean_shutdown(cls.worker_sock))\n mitogen.core.set_cloexec(cls.worker_sock.fileno())\n mitogen.core.set_cloexec(cls.child_sock.fileno())\n\n cls.profiling = os.environ.get('MITOGEN_PROFILING') is not None\n if cls.profiling:\n mitogen.core.enable_profiling()\n if _init_logging:\n ansible_mitogen.logging.setup()\n\n cls.original_env = dict(os.environ)\n cls.child_pid = os.fork()\n if cls.child_pid:\n save_pid('controller')\n ansible_mitogen.logging.set_process_name('top')\n ansible_mitogen.affinity.policy.assign_controller()\n cls.child_sock.close()\n cls.child_sock = None\n mitogen.core.io_op(cls.worker_sock.recv, 1)\n else:\n save_pid('mux')\n ansible_mitogen.logging.set_process_name('mux')\n ansible_mitogen.affinity.policy.assign_muxprocess()\n cls.worker_sock.close()\n cls.worker_sock = None\n self = cls()\n self.worker_main()']	def start(cls, _init_logging=True): """ Arrange for the subprocess to be started, if it is not already running. The parent process picks a UNIX socket path the child will use prior to fork, creates a socketpair used essentially as a semaphore, then blocks waiting for the child to indicate the UNIX socket is ready for use. :param bool _init_logging: For testing, if :data:False, don't initialize logging. """ if cls.worker_sock is not None: return if faulthandler is not None: faulthandler.enable() mitogen.utils.setup_gil() cls.unix_listener_path = mitogen.unix.make_socket_path() cls.worker_sock, cls.child_sock = socket.socketpair() atexit.register(lambda: clean_shutdown(cls.worker_sock)) mitogen.core.set_cloexec(cls.worker_sock.fileno()) mitogen.core.set_cloexec(cls.child_sock.fileno()) cls.profiling = os.environ.get('MI...
`updates the current cookies with a new set @param array $cookies new cookies with which to update current ones @return boolean always return true @access private`	`function UpdateCookies($cookies)`
{
if (sizeof($this->cookies) == 0) {
// no existing cookies: take whatever is new
if (sizeof($cookies) > 0) {
$this->debug('Setting new cookie(s)');
$this->cookies = $cookies;
}

return TRUE;
}
if (sizeof($cookies) == 0) {
// no new cookies: keep what we've got
return TRUE;
}
// merge
foreach ($cookies as $newCookie) {
if (!is_array($newCookie)) {
continue;
}
if ((!isset($newCookie['name'])) \|\| (!isset($newCookie['value']))) {
continue;
}
$newName = $newCookie['name'];

$found = FALSE;
for ($i = 0; $i < count($this->cookies); $i++) {
$cookie = $this->cookies[$i];
if (!is_array($cookie)) {
continue;
...	`csn`	`["public function setCookies(array $cookies) : Request\n {\n $query = http_build_query($cookies);\n parse_str($query, $this->cookies);\n\n if ($cookies) {\n $cookie = str_replace('&', '; ', $query);\n $this->setHeader('Cookie', $cookie);\n } else {\n $this->removeHeader('Cookie');\n }\n\n return $this;\n }"]`	`public function setCookies(array $cookies) : Request { $query = http_build_query($cookies); parse_str($query, $this->cookies); if ($cookies) { $cookie = str_replace('&', '; ', $query); $this->setHeader('Cookie', $cookie); } else { $this->removeHeader('Cookie'); } return $this; }`

Loss: pylate.losses.contrastive.Contrastive

Training Hyperparameters

Non-Default Hyperparameters

eval_strategy: steps
per_device_train_batch_size: 256
per_device_eval_batch_size: 256
learning_rate: 3e-06
num_train_epochs: 1
fp16: True
hub_model_id: colbert-code-17m

All Hyperparameters

Click to expand

overwrite_output_dir: False
do_predict: False
eval_strategy: steps
prediction_loss_only: True
per_device_train_batch_size: 256
per_device_eval_batch_size: 256
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: 0.0
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: 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: True
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}
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: colbert-code-17m
hub_strategy: every_save
hub_private_repo: None
hub_always_push: False
hub_revision: None
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
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: {}

Framework Versions

Python: 3.12.3
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},
}