Add new SentenceTransformer model

3d6d02b verified 3 months ago

27.8 kB

language:
  - en
license: apache-2.0
tags:
  - sentence-transformers
  - sentence-similarity
  - feature-extraction
  - dense
  - generated_from_trainer
  - dataset_size:900
  - loss:MatryoshkaLoss
  - loss:MultipleNegativesRankingLoss
base_model: jinaai/jina-code-embeddings-0.5b
widget:
  - source_sentence: >-
      Best practices for async
      test_similarity_search_with_relevance_score_with_threshold_and_filter
    sentences:
      - |-
        def test_tool_retry_custom_failure_formatter() -> None:
            """Test ToolRetryMiddlewarewith custom failure message formatter."""

            def custom_formatter(exc: Exception) -> str:
                return f"Custom error: {type(exc).__name__}"

            model = FakeToolCallingModel(
                tool_calls=[
                    [ToolCall(name="failing_tool", args={"value": "test"}, id="1")],
                    [],
                ]
            )

            retry = ToolRetryMiddleware(
                max_retries=1,
                initial_delay=0.01,
                jitter=False,
                on_failure=custom_formatter,
            )

            agent = create_agent(
                model=model,
                tools=[failing_tool],
                middleware=[retry],
                checkpointer=InMemorySaver(),
            )

            result = agent.invoke(
                {"messages": [HumanMessage("Use failing tool")]},
                {"configurable": {"thread_id": "test"}},
            )

            tool_messages = [m for m in result["messages"] if isinstance(m, ToolMessage)]
            assert len(tool_messages) == 1
            assert "Custom error: ValueError" in tool_messages[0].content
      - |-
        def test_parse_scores(answer: str) -> None:
            result = output_parser.parse(answer)

            assert result["answer"] == "foo bar answer."

            score = int(result["score"])
            assert score == 80
      - >-
        async def
        test_similarity_search_with_relevance_score_with_threshold_and_filter(
            vector_name: str | None,
            qdrant_location: str,
        ) -> None:
            """Test end to end construction and search."""
            texts = ["foo", "bar", "baz"]
            metadatas = [
                {"page": i, "metadata": {"page": i + 1, "pages": [i + 2, -1]}}
                for i in range(len(texts))
            ]
            docsearch = Qdrant.from_texts(
                texts,
                ConsistentFakeEmbeddings(),
                metadatas=metadatas,
                vector_name=vector_name,
                location=qdrant_location,
            )
            score_threshold = 0.99  # for almost exact match
            # test negative filter condition
            negative_filter = {"page": 1, "metadata": {"page": 2, "pages": [3]}}
            kwargs = {"filter": negative_filter, "score_threshold": score_threshold}
            output = docsearch.similarity_search_with_relevance_scores("foo", k=3, **kwargs)
            assert len(output) == 0
            # test positive filter condition
            positive_filter = {"page": 0, "metadata": {"page": 1, "pages": [2]}}
            kwargs = {"filter": positive_filter, "score_threshold": score_threshold}
            output = await docsearch.asimilarity_search_with_relevance_scores(
                "foo", k=3, **kwargs
            )
            assert len(output) == 1
            assert all(score >= score_threshold for _, score in output)
  - source_sentence: Explain the test_empty_token logic
    sentences:
      - |-
        def test_empty_token(self) -> None:
                assert len(_get_token_ids_default_method("")) == 0
      - |-
        def convert_to_openai_tool(
            tool: Mapping[str, Any] | type[BaseModel] | Callable | BaseTool,
            *,
            strict: bool | None = None,
        ) -> dict[str, Any]:
            """Convert a tool-like object to an OpenAI tool schema.

            [OpenAI tool schema reference](https://platform.openai.com/docs/api-reference/chat/create#chat-create-tools)

            Args:
                tool:
                    Either a dictionary, a `pydantic.BaseModel` class, Python function, or
                    `BaseTool`. If a dictionary is passed in, it is assumed to already be a
                    valid OpenAI function, a JSON schema with top-level `title` key specified,
                    an Anthropic format tool, or an Amazon Bedrock Converse format tool.
                strict:
                    If `True`, model output is guaranteed to exactly match the JSON Schema
                    provided in the function definition. If `None`, `strict` argument will not
                    be included in tool definition.

            Returns:
                A dict version of the passed in tool which is compatible with the
                OpenAI tool-calling API.

            !!! warning "Behavior changed in `langchain-core` 0.3.16"

                `description` and `parameters` keys are now optional. Only `name` is
                required and guaranteed to be part of the output.

            !!! warning "Behavior changed in `langchain-core` 0.3.44"

                Return OpenAI Responses API-style tools unchanged. This includes
                any dict with `"type"` in `"file_search"`, `"function"`,
                `"computer_use_preview"`, `"web_search_preview"`.

            !!! warning "Behavior changed in `langchain-core` 0.3.63"

                Added support for OpenAI's image generation built-in tool.
            """
            # Import locally to prevent circular import
            from langchain_core.tools import Tool  # noqa: PLC0415

            if isinstance(tool, dict):
                if tool.get("type") in _WellKnownOpenAITools:
                    return tool
                # As of 03.12.25 can be "web_search_preview" or "web_search_preview_2025_03_11"
                if (tool.get("type") or "").startswith("web_search_preview"):
                    return tool
            if isinstance(tool, Tool) and (tool.metadata or {}).get("type") == "custom_tool":
                oai_tool = {
                    "type": "custom",
                    "name": tool.name,
                    "description": tool.description,
                }
                if tool.metadata is not None and "format" in tool.metadata:
                    oai_tool["format"] = tool.metadata["format"]
                return oai_tool
            oai_function = convert_to_openai_function(tool, strict=strict)
            return {"type": "function", "function": oai_function}
      - |-
        def strip_think_tags(text: str) -> str:
            """Removes all <think>...</think> tags and their content from text.

            This function removes all occurrences of think tags, preserving text
            before, between, and after the tags. It also handles markdown code fences.

            Args:
                text: The input text that may contain think tags.

            Returns:
                The text with all `<think>...</think>` blocks removed.
            """
            # Remove all <think>...</think> blocks using regex
            # The pattern matches <think> followed by any content (non-greedy) until </think>
            result = re.sub(r"<think>.*?</think>", "", text, flags=re.DOTALL)

            # Remove markdown code fence markers if present
            result = result.strip()
            if result.startswith("```json"):
                result = result[len("```json") :].strip()
            elif result.startswith("```"):
                result = result[3:].strip()

            if result.endswith("```"):
                result = result[:-3].strip()

            return result
  - source_sentence: Explain the test_dict_int_op_invalid_types logic
    sentences:
      - |-
        def test_parse_empty_after_stripping_think_tags(self) -> None:
                """Test handling when only think tags remain."""
                parser: ReasoningStructuredOutputParser[MockPerson] = (
                    ReasoningStructuredOutputParser(pydantic_object=MockPerson)
                )
                text = "<think>Only reasoning here</think>"
                generation = Generation(text=text)
                with pytest.raises(OutputParserException):
                    parser.parse_result([generation])
      - |-
        def test_dict_int_op_invalid_types() -> None:
            left = {"a": 1, "b": "string"}
            right = {"a": 2, "b": 3}
            with pytest.raises(
                ValueError,
                match="Only dict and int values are supported",
            ):
                _dict_int_op(left, right, operator.add)
      - |-
        def _invocation_params(self) -> dict[str, Any]:
                params: dict = {"model": self.model, **self.model_kwargs}
                if self.dimensions is not None:
                    params["dimensions"] = self.dimensions
                return params
  - source_sentence: Explain the get_allowed_tools logic
    sentences:
      - |-
        def test_text_accessor() -> None:
            """Test that `message.text` property and `.text()` method return the same value."""
            human_msg = HumanMessage(content="Hello world")
            assert human_msg.text == "Hello world"
            assert human_msg.text == "Hello world"
            assert str(human_msg.text) == str(human_msg.text)

            system_msg = SystemMessage(content="You are a helpful assistant")
            assert system_msg.text == "You are a helpful assistant"
            assert system_msg.text == "You are a helpful assistant"
            assert str(system_msg.text) == str(system_msg.text)

            ai_msg = AIMessage(content="I can help you with that")
            assert ai_msg.text == "I can help you with that"
            assert ai_msg.text == "I can help you with that"
            assert str(ai_msg.text) == str(ai_msg.text)

            tool_msg = ToolMessage(content="Task completed", tool_call_id="tool_1")
            assert tool_msg.text == "Task completed"
            assert tool_msg.text == "Task completed"
            assert str(tool_msg.text) == str(tool_msg.text)

            complex_msg = HumanMessage(
                content=[{"type": "text", "text": "Hello "}, {"type": "text", "text": "world"}]
            )
            assert complex_msg.text == "Hello world"
            assert complex_msg.text == "Hello world"
            assert str(complex_msg.text) == str(complex_msg.text)

            mixed_msg = AIMessage(
                content=[
                    {"type": "text", "text": "The answer is "},
                    {"type": "tool_use", "name": "calculate", "input": {"x": 2}, "id": "1"},
                    {"type": "text", "text": "42"},
                ]
            )
            assert mixed_msg.text == "The answer is 42"
            assert mixed_msg.text == "The answer is 42"
            assert str(mixed_msg.text) == str(mixed_msg.text)

            empty_msg = HumanMessage(content=[])
            assert empty_msg.text == ""
            assert empty_msg.text == ""
            assert str(empty_msg.text) == str(empty_msg.text)
      - |-
        def __getattr__(name: str) -> Any:
            """Look up attributes dynamically."""
            return _import_attribute(name)
      - |-
        def get_allowed_tools(self) -> list[str] | None:
                """Get allowed tools.

                Returns:
                    Allowed tools.
                """
                return None
  - source_sentence: Explain the test_hashing_custom_key_encoder logic
    sentences:
      - |-
        def __getattr__(name: str) -> Any:
            """Look up attributes dynamically."""
            return _import_attribute(name)
      - |-
        def test_hashing_custom_key_encoder() -> None:
            """Test hashing with a custom key encoder."""

            def custom_key_encoder(doc: Document) -> str:
                return f"quack-{doc.metadata['key']}"

            document = Document(
                page_content="Lorem ipsum dolor sit amet", metadata={"key": "like a duck"}
            )
            hashed_document = _get_document_with_hash(document, key_encoder=custom_key_encoder)
            assert hashed_document.id == "quack-like a duck"
            assert isinstance(hashed_document.id, str)
      - |-
        def __getattr__(name: str) -> Any:
            """Look up attributes dynamically."""
            return _import_attribute(name)
pipeline_tag: sentence-similarity
library_name: sentence-transformers

jina2 Base

This is a sentence-transformers model finetuned from jinaai/jina-code-embeddings-0.5b. It maps sentences & paragraphs to a 896-dimensional dense vector space and can be used for semantic textual similarity, semantic search, paraphrase mining, text classification, clustering, and more.

Model Details

Model Description

Model Type: Sentence Transformer
Base model: jinaai/jina-code-embeddings-0.5b
Maximum Sequence Length: 512 tokens
Output Dimensionality: 896 dimensions
Similarity Function: Cosine Similarity
Language: en
License: apache-2.0

Model Sources

Documentation: Sentence Transformers Documentation
Repository: Sentence Transformers on GitHub
Hugging Face: Sentence Transformers on Hugging Face

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 512, 'do_lower_case': False, 'architecture': 'Qwen2Model'})
  (1): Pooling({'word_embedding_dimension': 896, 'pooling_mode_cls_token': False, 'pooling_mode_mean_tokens': False, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False, 'pooling_mode_weightedmean_tokens': False, 'pooling_mode_lasttoken': True, 'include_prompt': True})
  (2): Normalize()
)

Usage

Direct Usage (Sentence Transformers)

First install the Sentence Transformers library:

pip install -U sentence-transformers

Then you can load this model and run inference.

from sentence_transformers import SentenceTransformer

# Download from the 🤗 Hub
model = SentenceTransformer("ondayex/jina-embed-base-dense-retriever")
# Run inference
sentences = [
    'Explain the test_hashing_custom_key_encoder logic',
    'def test_hashing_custom_key_encoder() -> None:\n    """Test hashing with a custom key encoder."""\n\n    def custom_key_encoder(doc: Document) -> str:\n        return f"quack-{doc.metadata[\'key\']}"\n\n    document = Document(\n        page_content="Lorem ipsum dolor sit amet", metadata={"key": "like a duck"}\n    )\n    hashed_document = _get_document_with_hash(document, key_encoder=custom_key_encoder)\n    assert hashed_document.id == "quack-like a duck"\n    assert isinstance(hashed_document.id, str)',
    'def __getattr__(name: str) -> Any:\n    """Look up attributes dynamically."""\n    return _import_attribute(name)',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 896]

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings, embeddings)
print(similarities)
# tensor([[1.0000, 0.8949, 0.1085],
#         [0.8949, 1.0000, 0.1456],
#         [0.1085, 0.1456, 1.0000]])

Training Details

Training Dataset

Unnamed Dataset

Size: 900 training samples
Columns: anchor and positive
Approximate statistics based on the first 900 samples:
anchor positive
type string string
details
min: 4 tokens
mean: 8.5 tokens
max: 27 tokens

min: 12 tokens
mean: 153.14 tokens
max: 512 tokens

	anchor	positive
type	string	string
details	min: 4 tokens mean: 8.5 tokens max: 27 tokens	min: 12 tokens mean: 153.14 tokens max: 512 tokens

Samples:

anchor	positive
`How does test_multiple_msg work in Python?`	def test_multiple_msg() -> None: human_msg = HumanMessage(content="human", additional_kwargs={"key": "value"}) ai_msg = AIMessage(content="ai") sys_msg = SystemMessage(content="sys") msgs = [ human_msg, ai_msg, sys_msg, ] assert messages_from_dict(messages_to_dict(msgs)) == msgs # Test with tool calls msgs = [ AIMessage( content="", tool_calls=[create_tool_call(name="a", args={"b": 1}, id=None)], ), AIMessage( content="", tool_calls=[create_tool_call(name="c", args={"c": 2}, id=None)], ), ] assert messages_from_dict(messages_to_dict(msgs)) == msgs
`How to implement getattr?`	`def getattr(name: str) -> Any: """Look up attributes dynamically.""" return _import_attribute(name)`
`Example usage of test_multiple_msg`	def test_multiple_msg() -> None: human_msg = HumanMessage(content="human", additional_kwargs={"key": "value"}) ai_msg = AIMessage(content="ai") sys_msg = SystemMessage(content="sys") msgs = [ human_msg, ai_msg, sys_msg, ] assert messages_from_dict(messages_to_dict(msgs)) == msgs # Test with tool calls msgs = [ AIMessage( content="", tool_calls=[create_tool_call(name="a", args={"b": 1}, id=None)], ), AIMessage( content="", tool_calls=[create_tool_call(name="c", args={"c": 2}, id=None)], ), ] assert messages_from_dict(messages_to_dict(msgs)) == msgs

Loss: MatryoshkaLoss with these parameters:

{
    "loss": "MultipleNegativesRankingLoss",
    "matryoshka_dims": [
        768,
        512,
        256,
        128,
        64
    ],
    "matryoshka_weights": [
        1,
        1,
        1,
        1,
        1
    ],
    "n_dims_per_step": -1
}

Training Hyperparameters

Non-Default Hyperparameters

per_device_train_batch_size: 4
per_device_eval_batch_size: 4
gradient_accumulation_steps: 16
learning_rate: 2e-05
num_train_epochs: 4
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: True
load_best_model_at_end: True
optim: adamw_torch
batch_sampler: no_duplicates

All Hyperparameters

Click to expand

overwrite_output_dir: False
do_predict: False
eval_strategy: no
prediction_loss_only: True
per_device_train_batch_size: 4
per_device_eval_batch_size: 4
per_gpu_train_batch_size: None
per_gpu_eval_batch_size: None
gradient_accumulation_steps: 16
eval_accumulation_steps: None
torch_empty_cache_steps: None
learning_rate: 2e-05
weight_decay: 0.0
adam_beta1: 0.9
adam_beta2: 0.999
adam_epsilon: 1e-08
max_grad_norm: 1.0
num_train_epochs: 4
max_steps: -1
lr_scheduler_type: cosine
lr_scheduler_kwargs: None
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
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
optim_args: None
adafactor: False
group_by_length: False
length_column_name: length
project: huggingface
trackio_space_id: trackio
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: 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: no
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: True
prompts: None
batch_sampler: no_duplicates
multi_dataset_batch_sampler: proportional
router_mapping: {}
learning_rate_mapping: {}

Training Logs

Epoch	Step	Training Loss
0.7111	10	0.0808
1.3556	20	0.022
2.0	30	0.0104
2.7111	40	0.0136
3.3556	50	0.0001
4.0	60	0.0442

Framework Versions

Python: 3.12.12
Sentence Transformers: 5.2.0
Transformers: 4.57.6
PyTorch: 2.9.0+cu126
Accelerate: 1.12.0
Datasets: 4.0.0
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",
}

MatryoshkaLoss

@misc{kusupati2024matryoshka,
    title={Matryoshka Representation Learning},
    author={Aditya Kusupati and Gantavya Bhatt and Aniket Rege and Matthew Wallingford and Aditya Sinha and Vivek Ramanujan and William Howard-Snyder and Kaifeng Chen and Sham Kakade and Prateek Jain and Ali Farhadi},
    year={2024},
    eprint={2205.13147},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}

MultipleNegativesRankingLoss

@misc{henderson2017efficient,
    title={Efficient Natural Language Response Suggestion for Smart Reply},
    author={Matthew Henderson and Rami Al-Rfou and Brian Strope and Yun-hsuan Sung and Laszlo Lukacs and Ruiqi Guo and Sanjiv Kumar and Balint Miklos and Ray Kurzweil},
    year={2017},
    eprint={1705.00652},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}