Skip to main content
API docs by Redocly

SQuARE Model Inference API (0.1.0)

Download OpenAPI specification:Download

API reference for model inference.

health

Heartbeat

Authorizations:
Auth

Responses

Response samples

Content type
application/json
{
  • "is_alive": true
}

prediction

Sequence Classification

Authorizations:
Auth
path Parameters
identifier
required
string (Identifier)
hf_username
required
string (Hf Username)
Request Body schema: application/json
required
Array of Input (strings) or Array of Input (strings) or Input (object) (Input)

Input for the model. Supports Huggingface Transformer inputs (i.e., list of sentences, or list of pairs of sentences), a dictionary with Transformer inputs, or a dictionary containing numpy arrays (as lists). For the numpy arrays, also set is_preprocessed=True.

Transformer/ Adapter:
Task 'question_answering' expects the input to be in the (question, context) format.

is_preprocessed
boolean (Is Preprocessed)
Default: false

Flag indicating that the input contains already pre-processed numpy arrays as list and that it needs no further pre-processing.

Transformer/ Adapter/ SentenceTransformer: 'is_preprocessed' is not supported.

preprocessing_kwargs
object (Preprocessing Kwargs)
Default: {}

Optional dictionary containing additional parameters for the pre-processing step.

SentenceTransformer: This is ignored.
Transformer/ Adapter: See the Huggingface tokenizer for possible parameters.

model_kwargs
object (Model Kwargs)
Default: {}

Optional dictionary containing parameters that are passed to the model for the forward pass to control what additional tensors are returned.

SentenceTransformer: This is ignored.
Transformer/ Adapter: See the forward method of the Huggingface models for possible parametersFor example, set ‘output_attentions=True’ to receive the attention results in the output.For adapter models, the following options are also available:1. set average_adapters to True to average the adapter weights

task_kwargs
object (Task Kwargs)
Default: {}

Optional dictionary containing additional parameters for handling of the task and task-related post-processing.

SentenceTransformer: This is ignored.
Transformer/ Adapter:
'sentence_classification':
- 'is_regression': Flag to treat output of models with num_labels>1 as regression, too, i.e., no softmax and no labels are returned
'token_classification':
- 'is_regression': Flag to treat output of models with num_labels>1 as regression, too, i.e., no softmax and no labels are returned
'embedding':
- 'embedding_mode: One of 'mean', 'max', 'cls', 'pooler', 'token'. The pooling mode used (or not used for 'token'). 'pooler' uses the pooler_output of a Transformer, i.e. the processed CLS token. Default value 'mean'.
'question_answering':
- 'topk': Return the top-k most likely spans. Default 1.
- 'max_answer_len': Maximal token length of answers. Default 128.
'generation':
- 'clean_up_tokenization_spaces': See parameter in Huggingface tokenizer.decode(). Default False
- See Huggingface model.generate() for all possible parameters that can be used. Note, 'model_kwargs' and 'task_kwargs' are merged for generation.
'normalize',boolen, 'True' for using normalized embedding, default 'False'

explain_kwargs
object (Explain Kwargs)
Default: {}

Optional dictionary containing additional parameters for explaining predictions
- 'method': explanation method such as 'simple_grads, integrated_grads,smooth_grads, attention or scaled_attention':
- 'top_k': number of word attributions to return:
- 'mode: One of 'question', 'context', 'all'. Returns respective attributions.

attack_kwargs
object (Attack Kwargs)
Default: {}

Optional dictionary containing additional parameters for attacking models
- 'method': explanation method such as 'hotflip', 'input_reduction'
'saliency_method': simple_grads, integrated_grads, smooth_grads, attention or scaled_attention :
- 'max_flips': number of words to flip in hotflip
- 'include_answer: Whether to remove answer from context while attacking model.

Adapter Name (string) or Array of Adapter Name (strings) (Adapter Name)
Default: ""

Only necessary for Adapter. The fully specified name of the to-be-used adapter from adapterhub.ml

Responses

Request samples

Content type
application/json
{
  • "input": [
    ],
  • "is_preprocessed": false,
  • "preprocessing_kwargs": { },
  • "model_kwargs": { },
  • "task_kwargs": { },
  • "explain_kwargs": { },
  • "attack_kwargs": { },
  • "adapter_name": ""
}

Response samples

Content type
application/json
{
  • "message": "string",
  • "task_id": "string"
}

Sequence Classification

Authorizations:
Auth
path Parameters
identifier
required
string (Identifier)
query Parameters
hf_username
string (Hf Username)
Request Body schema: application/json
required
Array of Input (strings) or Array of Input (strings) or Input (object) (Input)

Input for the model. Supports Huggingface Transformer inputs (i.e., list of sentences, or list of pairs of sentences), a dictionary with Transformer inputs, or a dictionary containing numpy arrays (as lists). For the numpy arrays, also set is_preprocessed=True.

Transformer/ Adapter:
Task 'question_answering' expects the input to be in the (question, context) format.

is_preprocessed
boolean (Is Preprocessed)
Default: false

Flag indicating that the input contains already pre-processed numpy arrays as list and that it needs no further pre-processing.

Transformer/ Adapter/ SentenceTransformer: 'is_preprocessed' is not supported.

preprocessing_kwargs
object (Preprocessing Kwargs)
Default: {}

Optional dictionary containing additional parameters for the pre-processing step.

SentenceTransformer: This is ignored.
Transformer/ Adapter: See the Huggingface tokenizer for possible parameters.

model_kwargs
object (Model Kwargs)
Default: {}

Optional dictionary containing parameters that are passed to the model for the forward pass to control what additional tensors are returned.

SentenceTransformer: This is ignored.
Transformer/ Adapter: See the forward method of the Huggingface models for possible parametersFor example, set ‘output_attentions=True’ to receive the attention results in the output.For adapter models, the following options are also available:1. set average_adapters to True to average the adapter weights

task_kwargs
object (Task Kwargs)
Default: {}

Optional dictionary containing additional parameters for handling of the task and task-related post-processing.

SentenceTransformer: This is ignored.
Transformer/ Adapter:
'sentence_classification':
- 'is_regression': Flag to treat output of models with num_labels>1 as regression, too, i.e., no softmax and no labels are returned
'token_classification':
- 'is_regression': Flag to treat output of models with num_labels>1 as regression, too, i.e., no softmax and no labels are returned
'embedding':
- 'embedding_mode: One of 'mean', 'max', 'cls', 'pooler', 'token'. The pooling mode used (or not used for 'token'). 'pooler' uses the pooler_output of a Transformer, i.e. the processed CLS token. Default value 'mean'.
'question_answering':
- 'topk': Return the top-k most likely spans. Default 1.
- 'max_answer_len': Maximal token length of answers. Default 128.
'generation':
- 'clean_up_tokenization_spaces': See parameter in Huggingface tokenizer.decode(). Default False
- See Huggingface model.generate() for all possible parameters that can be used. Note, 'model_kwargs' and 'task_kwargs' are merged for generation.
'normalize',boolen, 'True' for using normalized embedding, default 'False'

explain_kwargs
object (Explain Kwargs)
Default: {}

Optional dictionary containing additional parameters for explaining predictions
- 'method': explanation method such as 'simple_grads, integrated_grads,smooth_grads, attention or scaled_attention':
- 'top_k': number of word attributions to return:
- 'mode: One of 'question', 'context', 'all'. Returns respective attributions.

attack_kwargs
object (Attack Kwargs)
Default: {}

Optional dictionary containing additional parameters for attacking models
- 'method': explanation method such as 'hotflip', 'input_reduction'
'saliency_method': simple_grads, integrated_grads, smooth_grads, attention or scaled_attention :
- 'max_flips': number of words to flip in hotflip
- 'include_answer: Whether to remove answer from context while attacking model.

Adapter Name (string) or Array of Adapter Name (strings) (Adapter Name)
Default: ""

Only necessary for Adapter. The fully specified name of the to-be-used adapter from adapterhub.ml

Responses

Request samples

Content type
application/json
{
  • "input": [
    ],
  • "is_preprocessed": false,
  • "preprocessing_kwargs": { },
  • "model_kwargs": { },
  • "task_kwargs": { },
  • "explain_kwargs": { },
  • "attack_kwargs": { },
  • "adapter_name": ""
}

Response samples

Content type
application/json
{
  • "message": "string",
  • "task_id": "string"
}

Token Classification

Authorizations:
Auth
path Parameters
identifier
required
string (Identifier)
hf_username
required
string (Hf Username)
Request Body schema: application/json
required
Array of Input (strings) or Array of Input (strings) or Input (object) (Input)

Input for the model. Supports Huggingface Transformer inputs (i.e., list of sentences, or list of pairs of sentences), a dictionary with Transformer inputs, or a dictionary containing numpy arrays (as lists). For the numpy arrays, also set is_preprocessed=True.

Transformer/ Adapter:
Task 'question_answering' expects the input to be in the (question, context) format.

is_preprocessed
boolean (Is Preprocessed)
Default: false

Flag indicating that the input contains already pre-processed numpy arrays as list and that it needs no further pre-processing.

Transformer/ Adapter/ SentenceTransformer: 'is_preprocessed' is not supported.

preprocessing_kwargs
object (Preprocessing Kwargs)
Default: {}

Optional dictionary containing additional parameters for the pre-processing step.

SentenceTransformer: This is ignored.
Transformer/ Adapter: See the Huggingface tokenizer for possible parameters.

model_kwargs
object (Model Kwargs)
Default: {}

Optional dictionary containing parameters that are passed to the model for the forward pass to control what additional tensors are returned.

SentenceTransformer: This is ignored.
Transformer/ Adapter: See the forward method of the Huggingface models for possible parametersFor example, set ‘output_attentions=True’ to receive the attention results in the output.For adapter models, the following options are also available:1. set average_adapters to True to average the adapter weights

task_kwargs
object (Task Kwargs)
Default: {}

Optional dictionary containing additional parameters for handling of the task and task-related post-processing.

SentenceTransformer: This is ignored.
Transformer/ Adapter:
'sentence_classification':
- 'is_regression': Flag to treat output of models with num_labels>1 as regression, too, i.e., no softmax and no labels are returned
'token_classification':
- 'is_regression': Flag to treat output of models with num_labels>1 as regression, too, i.e., no softmax and no labels are returned
'embedding':
- 'embedding_mode: One of 'mean', 'max', 'cls', 'pooler', 'token'. The pooling mode used (or not used for 'token'). 'pooler' uses the pooler_output of a Transformer, i.e. the processed CLS token. Default value 'mean'.
'question_answering':
- 'topk': Return the top-k most likely spans. Default 1.
- 'max_answer_len': Maximal token length of answers. Default 128.
'generation':
- 'clean_up_tokenization_spaces': See parameter in Huggingface tokenizer.decode(). Default False
- See Huggingface model.generate() for all possible parameters that can be used. Note, 'model_kwargs' and 'task_kwargs' are merged for generation.
'normalize',boolen, 'True' for using normalized embedding, default 'False'

explain_kwargs
object (Explain Kwargs)
Default: {}

Optional dictionary containing additional parameters for explaining predictions
- 'method': explanation method such as 'simple_grads, integrated_grads,smooth_grads, attention or scaled_attention':
- 'top_k': number of word attributions to return:
- 'mode: One of 'question', 'context', 'all'. Returns respective attributions.

attack_kwargs
object (Attack Kwargs)
Default: {}

Optional dictionary containing additional parameters for attacking models
- 'method': explanation method such as 'hotflip', 'input_reduction'
'saliency_method': simple_grads, integrated_grads, smooth_grads, attention or scaled_attention :
- 'max_flips': number of words to flip in hotflip
- 'include_answer: Whether to remove answer from context while attacking model.

Adapter Name (string) or Array of Adapter Name (strings) (Adapter Name)
Default: ""

Only necessary for Adapter. The fully specified name of the to-be-used adapter from adapterhub.ml

Responses

Request samples

Content type
application/json
{
  • "input": [
    ],
  • "is_preprocessed": false,
  • "preprocessing_kwargs": { },
  • "model_kwargs": { },
  • "task_kwargs": { },
  • "explain_kwargs": { },
  • "attack_kwargs": { },
  • "adapter_name": ""
}

Response samples

Content type
application/json
{
  • "message": "string",
  • "task_id": "string"
}

Token Classification

Authorizations:
Auth
path Parameters
identifier
required
string (Identifier)
query Parameters
hf_username
string (Hf Username)
Request Body schema: application/json
required
Array of Input (strings) or Array of Input (strings) or Input (object) (Input)

Input for the model. Supports Huggingface Transformer inputs (i.e., list of sentences, or list of pairs of sentences), a dictionary with Transformer inputs, or a dictionary containing numpy arrays (as lists). For the numpy arrays, also set is_preprocessed=True.

Transformer/ Adapter:
Task 'question_answering' expects the input to be in the (question, context) format.

is_preprocessed
boolean (Is Preprocessed)
Default: false

Flag indicating that the input contains already pre-processed numpy arrays as list and that it needs no further pre-processing.

Transformer/ Adapter/ SentenceTransformer: 'is_preprocessed' is not supported.

preprocessing_kwargs
object (Preprocessing Kwargs)
Default: {}

Optional dictionary containing additional parameters for the pre-processing step.

SentenceTransformer: This is ignored.
Transformer/ Adapter: See the Huggingface tokenizer for possible parameters.

model_kwargs
object (Model Kwargs)
Default: {}

Optional dictionary containing parameters that are passed to the model for the forward pass to control what additional tensors are returned.

SentenceTransformer: This is ignored.
Transformer/ Adapter: See the forward method of the Huggingface models for possible parametersFor example, set ‘output_attentions=True’ to receive the attention results in the output.For adapter models, the following options are also available:1. set average_adapters to True to average the adapter weights

task_kwargs
object (Task Kwargs)
Default: {}

Optional dictionary containing additional parameters for handling of the task and task-related post-processing.

SentenceTransformer: This is ignored.
Transformer/ Adapter:
'sentence_classification':
- 'is_regression': Flag to treat output of models with num_labels>1 as regression, too, i.e., no softmax and no labels are returned
'token_classification':
- 'is_regression': Flag to treat output of models with num_labels>1 as regression, too, i.e., no softmax and no labels are returned
'embedding':
- 'embedding_mode: One of 'mean', 'max', 'cls', 'pooler', 'token'. The pooling mode used (or not used for 'token'). 'pooler' uses the pooler_output of a Transformer, i.e. the processed CLS token. Default value 'mean'.
'question_answering':
- 'topk': Return the top-k most likely spans. Default 1.
- 'max_answer_len': Maximal token length of answers. Default 128.
'generation':
- 'clean_up_tokenization_spaces': See parameter in Huggingface tokenizer.decode(). Default False
- See Huggingface model.generate() for all possible parameters that can be used. Note, 'model_kwargs' and 'task_kwargs' are merged for generation.
'normalize',boolen, 'True' for using normalized embedding, default 'False'

explain_kwargs
object (Explain Kwargs)
Default: {}

Optional dictionary containing additional parameters for explaining predictions
- 'method': explanation method such as 'simple_grads, integrated_grads,smooth_grads, attention or scaled_attention':
- 'top_k': number of word attributions to return:
- 'mode: One of 'question', 'context', 'all'. Returns respective attributions.

attack_kwargs
object (Attack Kwargs)
Default: {}

Optional dictionary containing additional parameters for attacking models
- 'method': explanation method such as 'hotflip', 'input_reduction'
'saliency_method': simple_grads, integrated_grads, smooth_grads, attention or scaled_attention :
- 'max_flips': number of words to flip in hotflip
- 'include_answer: Whether to remove answer from context while attacking model.

Adapter Name (string) or Array of Adapter Name (strings) (Adapter Name)
Default: ""

Only necessary for Adapter. The fully specified name of the to-be-used adapter from adapterhub.ml

Responses

Request samples

Content type
application/json
{
  • "input": [
    ],
  • "is_preprocessed": false,
  • "preprocessing_kwargs": { },
  • "model_kwargs": { },
  • "task_kwargs": { },
  • "explain_kwargs": { },
  • "attack_kwargs": { },
  • "adapter_name": ""
}

Response samples

Content type
application/json
{
  • "message": "string",
  • "task_id": "string"
}

Embedding

Authorizations:
Auth
path Parameters
identifier
required
string (Identifier)
hf_username
required
string (Hf Username)
Request Body schema: application/json
required
Array of Input (strings) or Array of Input (strings) or Input (object) (Input)

Input for the model. Supports Huggingface Transformer inputs (i.e., list of sentences, or list of pairs of sentences), a dictionary with Transformer inputs, or a dictionary containing numpy arrays (as lists). For the numpy arrays, also set is_preprocessed=True.

Transformer/ Adapter:
Task 'question_answering' expects the input to be in the (question, context) format.

is_preprocessed
boolean (Is Preprocessed)
Default: false

Flag indicating that the input contains already pre-processed numpy arrays as list and that it needs no further pre-processing.

Transformer/ Adapter/ SentenceTransformer: 'is_preprocessed' is not supported.

preprocessing_kwargs
object (Preprocessing Kwargs)
Default: {}

Optional dictionary containing additional parameters for the pre-processing step.

SentenceTransformer: This is ignored.
Transformer/ Adapter: See the Huggingface tokenizer for possible parameters.

model_kwargs
object (Model Kwargs)
Default: {}

Optional dictionary containing parameters that are passed to the model for the forward pass to control what additional tensors are returned.

SentenceTransformer: This is ignored.
Transformer/ Adapter: See the forward method of the Huggingface models for possible parametersFor example, set ‘output_attentions=True’ to receive the attention results in the output.For adapter models, the following options are also available:1. set average_adapters to True to average the adapter weights

task_kwargs
object (Task Kwargs)
Default: {}

Optional dictionary containing additional parameters for handling of the task and task-related post-processing.

SentenceTransformer: This is ignored.
Transformer/ Adapter:
'sentence_classification':
- 'is_regression': Flag to treat output of models with num_labels>1 as regression, too, i.e., no softmax and no labels are returned
'token_classification':
- 'is_regression': Flag to treat output of models with num_labels>1 as regression, too, i.e., no softmax and no labels are returned
'embedding':
- 'embedding_mode: One of 'mean', 'max', 'cls', 'pooler', 'token'. The pooling mode used (or not used for 'token'). 'pooler' uses the pooler_output of a Transformer, i.e. the processed CLS token. Default value 'mean'.
'question_answering':
- 'topk': Return the top-k most likely spans. Default 1.
- 'max_answer_len': Maximal token length of answers. Default 128.
'generation':
- 'clean_up_tokenization_spaces': See parameter in Huggingface tokenizer.decode(). Default False
- See Huggingface model.generate() for all possible parameters that can be used. Note, 'model_kwargs' and 'task_kwargs' are merged for generation.
'normalize',boolen, 'True' for using normalized embedding, default 'False'

explain_kwargs
object (Explain Kwargs)
Default: {}

Optional dictionary containing additional parameters for explaining predictions
- 'method': explanation method such as 'simple_grads, integrated_grads,smooth_grads, attention or scaled_attention':
- 'top_k': number of word attributions to return:
- 'mode: One of 'question', 'context', 'all'. Returns respective attributions.

attack_kwargs
object (Attack Kwargs)
Default: {}

Optional dictionary containing additional parameters for attacking models
- 'method': explanation method such as 'hotflip', 'input_reduction'
'saliency_method': simple_grads, integrated_grads, smooth_grads, attention or scaled_attention :
- 'max_flips': number of words to flip in hotflip
- 'include_answer: Whether to remove answer from context while attacking model.

Adapter Name (string) or Array of Adapter Name (strings) (Adapter Name)
Default: ""

Only necessary for Adapter. The fully specified name of the to-be-used adapter from adapterhub.ml

Responses

Request samples

Content type
application/json
{
  • "input": [
    ],
  • "is_preprocessed": false,
  • "preprocessing_kwargs": { },
  • "model_kwargs": { },
  • "task_kwargs": { },
  • "explain_kwargs": { },
  • "attack_kwargs": { },
  • "adapter_name": ""
}

Response samples

Content type
application/json
{
  • "message": "string",
  • "task_id": "string"
}

Embedding

Authorizations:
Auth
path Parameters
identifier
required
string (Identifier)
query Parameters
hf_username
string (Hf Username)
Request Body schema: application/json
required
Array of Input (strings) or Array of Input (strings) or Input (object) (Input)

Input for the model. Supports Huggingface Transformer inputs (i.e., list of sentences, or list of pairs of sentences), a dictionary with Transformer inputs, or a dictionary containing numpy arrays (as lists). For the numpy arrays, also set is_preprocessed=True.

Transformer/ Adapter:
Task 'question_answering' expects the input to be in the (question, context) format.

is_preprocessed
boolean (Is Preprocessed)
Default: false

Flag indicating that the input contains already pre-processed numpy arrays as list and that it needs no further pre-processing.

Transformer/ Adapter/ SentenceTransformer: 'is_preprocessed' is not supported.

preprocessing_kwargs
object (Preprocessing Kwargs)
Default: {}

Optional dictionary containing additional parameters for the pre-processing step.

SentenceTransformer: This is ignored.
Transformer/ Adapter: See the Huggingface tokenizer for possible parameters.

model_kwargs
object (Model Kwargs)
Default: {}

Optional dictionary containing parameters that are passed to the model for the forward pass to control what additional tensors are returned.

SentenceTransformer: This is ignored.
Transformer/ Adapter: See the forward method of the Huggingface models for possible parametersFor example, set ‘output_attentions=True’ to receive the attention results in the output.For adapter models, the following options are also available:1. set average_adapters to True to average the adapter weights

task_kwargs
object (Task Kwargs)
Default: {}

Optional dictionary containing additional parameters for handling of the task and task-related post-processing.

SentenceTransformer: This is ignored.
Transformer/ Adapter:
'sentence_classification':
- 'is_regression': Flag to treat output of models with num_labels>1 as regression, too, i.e., no softmax and no labels are returned
'token_classification':
- 'is_regression': Flag to treat output of models with num_labels>1 as regression, too, i.e., no softmax and no labels are returned
'embedding':
- 'embedding_mode: One of 'mean', 'max', 'cls', 'pooler', 'token'. The pooling mode used (or not used for 'token'). 'pooler' uses the pooler_output of a Transformer, i.e. the processed CLS token. Default value 'mean'.
'question_answering':
- 'topk': Return the top-k most likely spans. Default 1.
- 'max_answer_len': Maximal token length of answers. Default 128.
'generation':
- 'clean_up_tokenization_spaces': See parameter in Huggingface tokenizer.decode(). Default False
- See Huggingface model.generate() for all possible parameters that can be used. Note, 'model_kwargs' and 'task_kwargs' are merged for generation.
'normalize',boolen, 'True' for using normalized embedding, default 'False'

explain_kwargs
object (Explain Kwargs)
Default: {}

Optional dictionary containing additional parameters for explaining predictions
- 'method': explanation method such as 'simple_grads, integrated_grads,smooth_grads, attention or scaled_attention':
- 'top_k': number of word attributions to return:
- 'mode: One of 'question', 'context', 'all'. Returns respective attributions.

attack_kwargs
object (Attack Kwargs)
Default: {}

Optional dictionary containing additional parameters for attacking models
- 'method': explanation method such as 'hotflip', 'input_reduction'
'saliency_method': simple_grads, integrated_grads, smooth_grads, attention or scaled_attention :
- 'max_flips': number of words to flip in hotflip
- 'include_answer: Whether to remove answer from context while attacking model.

Adapter Name (string) or Array of Adapter Name (strings) (Adapter Name)
Default: ""

Only necessary for Adapter. The fully specified name of the to-be-used adapter from adapterhub.ml

Responses

Request samples

Content type
application/json
{
  • "input": [
    ],
  • "is_preprocessed": false,
  • "preprocessing_kwargs": { },
  • "model_kwargs": { },
  • "task_kwargs": { },
  • "explain_kwargs": { },
  • "attack_kwargs": { },
  • "adapter_name": ""
}

Response samples

Content type
application/json
{
  • "message": "string",
  • "task_id": "string"
}

Question Answering

Authorizations:
Auth
path Parameters
identifier
required
string (Identifier)
hf_username
required
string (Hf Username)
Request Body schema: application/json
required
Array of Input (strings) or Array of Input (strings) or Input (object) (Input)

Input for the model. Supports Huggingface Transformer inputs (i.e., list of sentences, or list of pairs of sentences), a dictionary with Transformer inputs, or a dictionary containing numpy arrays (as lists). For the numpy arrays, also set is_preprocessed=True.

Transformer/ Adapter:
Task 'question_answering' expects the input to be in the (question, context) format.

is_preprocessed
boolean (Is Preprocessed)
Default: false

Flag indicating that the input contains already pre-processed numpy arrays as list and that it needs no further pre-processing.

Transformer/ Adapter/ SentenceTransformer: 'is_preprocessed' is not supported.

preprocessing_kwargs
object (Preprocessing Kwargs)
Default: {}

Optional dictionary containing additional parameters for the pre-processing step.

SentenceTransformer: This is ignored.
Transformer/ Adapter: See the Huggingface tokenizer for possible parameters.

model_kwargs
object (Model Kwargs)
Default: {}

Optional dictionary containing parameters that are passed to the model for the forward pass to control what additional tensors are returned.

SentenceTransformer: This is ignored.
Transformer/ Adapter: See the forward method of the Huggingface models for possible parametersFor example, set ‘output_attentions=True’ to receive the attention results in the output.For adapter models, the following options are also available:1. set average_adapters to True to average the adapter weights

task_kwargs
object (Task Kwargs)
Default: {}

Optional dictionary containing additional parameters for handling of the task and task-related post-processing.

SentenceTransformer: This is ignored.
Transformer/ Adapter:
'sentence_classification':
- 'is_regression': Flag to treat output of models with num_labels>1 as regression, too, i.e., no softmax and no labels are returned
'token_classification':
- 'is_regression': Flag to treat output of models with num_labels>1 as regression, too, i.e., no softmax and no labels are returned
'embedding':
- 'embedding_mode: One of 'mean', 'max', 'cls', 'pooler', 'token'. The pooling mode used (or not used for 'token'). 'pooler' uses the pooler_output of a Transformer, i.e. the processed CLS token. Default value 'mean'.
'question_answering':
- 'topk': Return the top-k most likely spans. Default 1.
- 'max_answer_len': Maximal token length of answers. Default 128.
'generation':
- 'clean_up_tokenization_spaces': See parameter in Huggingface tokenizer.decode(). Default False
- See Huggingface model.generate() for all possible parameters that can be used. Note, 'model_kwargs' and 'task_kwargs' are merged for generation.
'normalize',boolen, 'True' for using normalized embedding, default 'False'

explain_kwargs
object (Explain Kwargs)
Default: {}

Optional dictionary containing additional parameters for explaining predictions
- 'method': explanation method such as 'simple_grads, integrated_grads,smooth_grads, attention or scaled_attention':
- 'top_k': number of word attributions to return:
- 'mode: One of 'question', 'context', 'all'. Returns respective attributions.

attack_kwargs
object (Attack Kwargs)
Default: {}

Optional dictionary containing additional parameters for attacking models
- 'method': explanation method such as 'hotflip', 'input_reduction'
'saliency_method': simple_grads, integrated_grads, smooth_grads, attention or scaled_attention :
- 'max_flips': number of words to flip in hotflip
- 'include_answer: Whether to remove answer from context while attacking model.

Adapter Name (string) or Array of Adapter Name (strings) (Adapter Name)
Default: ""

Only necessary for Adapter. The fully specified name of the to-be-used adapter from adapterhub.ml

Responses

Request samples

Content type
application/json
{
  • "input": [
    ],
  • "is_preprocessed": false,
  • "preprocessing_kwargs": { },
  • "model_kwargs": { },
  • "task_kwargs": { },
  • "explain_kwargs": { },
  • "attack_kwargs": { },
  • "adapter_name": ""
}

Response samples

Content type
application/json
{
  • "message": "string",
  • "task_id": "string"
}

Question Answering

Authorizations:
Auth
path Parameters
identifier
required
string (Identifier)
query Parameters
hf_username
string (Hf Username)
Request Body schema: application/json
required
Array of Input (strings) or Array of Input (strings) or Input (object) (Input)

Input for the model. Supports Huggingface Transformer inputs (i.e., list of sentences, or list of pairs of sentences), a dictionary with Transformer inputs, or a dictionary containing numpy arrays (as lists). For the numpy arrays, also set is_preprocessed=True.

Transformer/ Adapter:
Task 'question_answering' expects the input to be in the (question, context) format.

is_preprocessed
boolean (Is Preprocessed)
Default: false

Flag indicating that the input contains already pre-processed numpy arrays as list and that it needs no further pre-processing.

Transformer/ Adapter/ SentenceTransformer: 'is_preprocessed' is not supported.

preprocessing_kwargs
object (Preprocessing Kwargs)
Default: {}

Optional dictionary containing additional parameters for the pre-processing step.

SentenceTransformer: This is ignored.
Transformer/ Adapter: See the Huggingface tokenizer for possible parameters.

model_kwargs
object (Model Kwargs)
Default: {}

Optional dictionary containing parameters that are passed to the model for the forward pass to control what additional tensors are returned.

SentenceTransformer: This is ignored.
Transformer/ Adapter: See the forward method of the Huggingface models for possible parametersFor example, set ‘output_attentions=True’ to receive the attention results in the output.For adapter models, the following options are also available:1. set average_adapters to True to average the adapter weights

task_kwargs
object (Task Kwargs)
Default: {}

Optional dictionary containing additional parameters for handling of the task and task-related post-processing.

SentenceTransformer: This is ignored.
Transformer/ Adapter:
'sentence_classification':
- 'is_regression': Flag to treat output of models with num_labels>1 as regression, too, i.e., no softmax and no labels are returned
'token_classification':
- 'is_regression': Flag to treat output of models with num_labels>1 as regression, too, i.e., no softmax and no labels are returned
'embedding':
- 'embedding_mode: One of 'mean', 'max', 'cls', 'pooler', 'token'. The pooling mode used (or not used for 'token'). 'pooler' uses the pooler_output of a Transformer, i.e. the processed CLS token. Default value 'mean'.
'question_answering':
- 'topk': Return the top-k most likely spans. Default 1.
- 'max_answer_len': Maximal token length of answers. Default 128.
'generation':
- 'clean_up_tokenization_spaces': See parameter in Huggingface tokenizer.decode(). Default False
- See Huggingface model.generate() for all possible parameters that can be used. Note, 'model_kwargs' and 'task_kwargs' are merged for generation.
'normalize',boolen, 'True' for using normalized embedding, default 'False'

explain_kwargs
object (Explain Kwargs)
Default: {}

Optional dictionary containing additional parameters for explaining predictions
- 'method': explanation method such as 'simple_grads, integrated_grads,smooth_grads, attention or scaled_attention':
- 'top_k': number of word attributions to return:
- 'mode: One of 'question', 'context', 'all'. Returns respective attributions.

attack_kwargs
object (Attack Kwargs)
Default: {}

Optional dictionary containing additional parameters for attacking models
- 'method': explanation method such as 'hotflip', 'input_reduction'
'saliency_method': simple_grads, integrated_grads, smooth_grads, attention or scaled_attention :
- 'max_flips': number of words to flip in hotflip
- 'include_answer: Whether to remove answer from context while attacking model.

Adapter Name (string) or Array of Adapter Name (strings) (Adapter Name)
Default: ""

Only necessary for Adapter. The fully specified name of the to-be-used adapter from adapterhub.ml

Responses

Request samples

Content type
application/json
{
  • "input": [
    ],
  • "is_preprocessed": false,
  • "preprocessing_kwargs": { },
  • "model_kwargs": { },
  • "task_kwargs": { },
  • "explain_kwargs": { },
  • "attack_kwargs": { },
  • "adapter_name": ""
}

Response samples

Content type
application/json
{
  • "message": "string",
  • "task_id": "string"
}

Generation

Authorizations:
Auth
path Parameters
identifier
required
string (Identifier)
hf_username
required
string (Hf Username)
Request Body schema: application/json
required
Array of Input (strings) or Array of Input (strings) or Input (object) (Input)

Input for the model. Supports Huggingface Transformer inputs (i.e., list of sentences, or list of pairs of sentences), a dictionary with Transformer inputs, or a dictionary containing numpy arrays (as lists). For the numpy arrays, also set is_preprocessed=True.

Transformer/ Adapter:
Task 'question_answering' expects the input to be in the (question, context) format.

is_preprocessed
boolean (Is Preprocessed)
Default: false

Flag indicating that the input contains already pre-processed numpy arrays as list and that it needs no further pre-processing.

Transformer/ Adapter/ SentenceTransformer: 'is_preprocessed' is not supported.

preprocessing_kwargs
object (Preprocessing Kwargs)
Default: {}

Optional dictionary containing additional parameters for the pre-processing step.

SentenceTransformer: This is ignored.
Transformer/ Adapter: See the Huggingface tokenizer for possible parameters.

model_kwargs
object (Model Kwargs)
Default: {}

Optional dictionary containing parameters that are passed to the model for the forward pass to control what additional tensors are returned.

SentenceTransformer: This is ignored.
Transformer/ Adapter: See the forward method of the Huggingface models for possible parametersFor example, set ‘output_attentions=True’ to receive the attention results in the output.For adapter models, the following options are also available:1. set average_adapters to True to average the adapter weights

task_kwargs
object (Task Kwargs)
Default: {}

Optional dictionary containing additional parameters for handling of the task and task-related post-processing.

SentenceTransformer: This is ignored.
Transformer/ Adapter:
'sentence_classification':
- 'is_regression': Flag to treat output of models with num_labels>1 as regression, too, i.e., no softmax and no labels are returned
'token_classification':
- 'is_regression': Flag to treat output of models with num_labels>1 as regression, too, i.e., no softmax and no labels are returned
'embedding':
- 'embedding_mode: One of 'mean', 'max', 'cls', 'pooler', 'token'. The pooling mode used (or not used for 'token'). 'pooler' uses the pooler_output of a Transformer, i.e. the processed CLS token. Default value 'mean'.
'question_answering':
- 'topk': Return the top-k most likely spans. Default 1.
- 'max_answer_len': Maximal token length of answers. Default 128.
'generation':
- 'clean_up_tokenization_spaces': See parameter in Huggingface tokenizer.decode(). Default False
- See Huggingface model.generate() for all possible parameters that can be used. Note, 'model_kwargs' and 'task_kwargs' are merged for generation.
'normalize',boolen, 'True' for using normalized embedding, default 'False'

explain_kwargs
object (Explain Kwargs)
Default: {}

Optional dictionary containing additional parameters for explaining predictions
- 'method': explanation method such as 'simple_grads, integrated_grads,smooth_grads, attention or scaled_attention':
- 'top_k': number of word attributions to return:
- 'mode: One of 'question', 'context', 'all'. Returns respective attributions.

attack_kwargs
object (Attack Kwargs)
Default: {}

Optional dictionary containing additional parameters for attacking models
- 'method': explanation method such as 'hotflip', 'input_reduction'
'saliency_method': simple_grads, integrated_grads, smooth_grads, attention or scaled_attention :
- 'max_flips': number of words to flip in hotflip
- 'include_answer: Whether to remove answer from context while attacking model.

Adapter Name (string) or Array of Adapter Name (strings) (Adapter Name)
Default: ""

Only necessary for Adapter. The fully specified name of the to-be-used adapter from adapterhub.ml

Responses

Request samples

Content type
application/json
{
  • "input": [
    ],
  • "is_preprocessed": false,
  • "preprocessing_kwargs": { },
  • "model_kwargs": { },
  • "task_kwargs": { },
  • "explain_kwargs": { },
  • "attack_kwargs": { },
  • "adapter_name": ""
}

Response samples

Content type
application/json
{
  • "message": "string",
  • "task_id": "string"
}

Generation

Authorizations:
Auth
path Parameters
identifier
required
string (Identifier)
query Parameters
hf_username
string (Hf Username)
Request Body schema: application/json
required
Array of Input (strings) or Array of Input (strings) or Input (object) (Input)

Input for the model. Supports Huggingface Transformer inputs (i.e., list of sentences, or list of pairs of sentences), a dictionary with Transformer inputs, or a dictionary containing numpy arrays (as lists). For the numpy arrays, also set is_preprocessed=True.

Transformer/ Adapter:
Task 'question_answering' expects the input to be in the (question, context) format.

is_preprocessed
boolean (Is Preprocessed)
Default: false

Flag indicating that the input contains already pre-processed numpy arrays as list and that it needs no further pre-processing.

Transformer/ Adapter/ SentenceTransformer: 'is_preprocessed' is not supported.

preprocessing_kwargs
object (Preprocessing Kwargs)
Default: {}

Optional dictionary containing additional parameters for the pre-processing step.

SentenceTransformer: This is ignored.
Transformer/ Adapter: See the Huggingface tokenizer for possible parameters.

model_kwargs
object (Model Kwargs)
Default: {}

Optional dictionary containing parameters that are passed to the model for the forward pass to control what additional tensors are returned.

SentenceTransformer: This is ignored.
Transformer/ Adapter: See the forward method of the Huggingface models for possible parametersFor example, set ‘output_attentions=True’ to receive the attention results in the output.For adapter models, the following options are also available:1. set average_adapters to True to average the adapter weights

task_kwargs
object (Task Kwargs)
Default: {}

Optional dictionary containing additional parameters for handling of the task and task-related post-processing.

SentenceTransformer: This is ignored.
Transformer/ Adapter:
'sentence_classification':
- 'is_regression': Flag to treat output of models with num_labels>1 as regression, too, i.e., no softmax and no labels are returned
'token_classification':
- 'is_regression': Flag to treat output of models with num_labels>1 as regression, too, i.e., no softmax and no labels are returned
'embedding':
- 'embedding_mode: One of 'mean', 'max', 'cls', 'pooler', 'token'. The pooling mode used (or not used for 'token'). 'pooler' uses the pooler_output of a Transformer, i.e. the processed CLS token. Default value 'mean'.
'question_answering':
- 'topk': Return the top-k most likely spans. Default 1.
- 'max_answer_len': Maximal token length of answers. Default 128.
'generation':
- 'clean_up_tokenization_spaces': See parameter in Huggingface tokenizer.decode(). Default False
- See Huggingface model.generate() for all possible parameters that can be used. Note, 'model_kwargs' and 'task_kwargs' are merged for generation.
'normalize',boolen, 'True' for using normalized embedding, default 'False'

explain_kwargs
object (Explain Kwargs)
Default: {}

Optional dictionary containing additional parameters for explaining predictions
- 'method': explanation method such as 'simple_grads, integrated_grads,smooth_grads, attention or scaled_attention':
- 'top_k': number of word attributions to return:
- 'mode: One of 'question', 'context', 'all'. Returns respective attributions.

attack_kwargs
object (Attack Kwargs)
Default: {}

Optional dictionary containing additional parameters for attacking models
- 'method': explanation method such as 'hotflip', 'input_reduction'
'saliency_method': simple_grads, integrated_grads, smooth_grads, attention or scaled_attention :
- 'max_flips': number of words to flip in hotflip
- 'include_answer: Whether to remove answer from context while attacking model.

Adapter Name (string) or Array of Adapter Name (strings) (Adapter Name)
Default: ""

Only necessary for Adapter. The fully specified name of the to-be-used adapter from adapterhub.ml

Responses

Request samples

Content type
application/json
{
  • "input": [
    ],
  • "is_preprocessed": false,
  • "preprocessing_kwargs": { },
  • "model_kwargs": { },
  • "task_kwargs": { },
  • "explain_kwargs": { },
  • "attack_kwargs": { },
  • "adapter_name": ""
}

Response samples

Content type
application/json
{
  • "message": "string",
  • "task_id": "string"
}

Get Task Results

Authorizations:
Auth
path Parameters
task_id
required
string (Task Id)

Responses

Response samples

Content type
application/json
null

Statistics

Returns the statistics of the model :return: the ModelStatistics for the model

Authorizations:
Auth
path Parameters
identifier
required
string (Identifier)
hf_username
required
string (Hf Username)

Responses

Response samples

Content type
application/json
{
  • "model_type": "string",
  • "model_name": "string",
  • "batch_size": 0,
  • "max_input": 0,
  • "model_class": "string",
  • "disable_gpu": true,
  • "return_plaintext_arrays": true,
  • "preloaded_adapters": true,
  • "transformers_cache": ".cache",
  • "model_path": "string",
  • "decoder_path": "string",
  • "onnx_use_quantized": true,
  • "is_encoder_decoder": true
}

Statistics

Returns the statistics of the model :return: the ModelStatistics for the model

Authorizations:
Auth
path Parameters
identifier
required
string (Identifier)
query Parameters
hf_username
string (Hf Username)

Responses

Response samples

Content type
application/json
{
  • "model_type": "string",
  • "model_name": "string",
  • "batch_size": 0,
  • "max_input": 0,
  • "model_class": "string",
  • "disable_gpu": true,
  • "return_plaintext_arrays": true,
  • "preloaded_adapters": true,
  • "transformers_cache": ".cache",
  • "model_path": "string",
  • "decoder_path": "string",
  • "onnx_use_quantized": true,
  • "is_encoder_decoder": true
}

Update

Update the model with the given parameters. (not all parameters can be updated through this method e.g. the model class is linked to the model, hence it can't be updated during runtime) :param updated_param: the new parameters :return: the information about the updated model

Authorizations:
Auth
path Parameters
identifier
required
string (Identifier)
hf_username
required
string (Hf Username)
Request Body schema: application/json
disable_gpu
boolean (Disable Gpu)
batch_size
integer (Batch Size)
max_input
integer (Max Input)
return_plaintext_arrays
boolean (Return Plaintext Arrays)

Responses

Request samples

Content type
application/json
{
  • "disable_gpu": true,
  • "batch_size": 0,
  • "max_input": 0,
  • "return_plaintext_arrays": true
}

Response samples

Content type
application/json
null

Update

Update the model with the given parameters. (not all parameters can be updated through this method e.g. the model class is linked to the model, hence it can't be updated during runtime) :param updated_param: the new parameters :return: the information about the updated model

Authorizations:
Auth
path Parameters
identifier
required
string (Identifier)
query Parameters
hf_username
string (Hf Username)
Request Body schema: application/json
disable_gpu
boolean (Disable Gpu)
batch_size
integer (Batch Size)
max_input
integer (Max Input)
return_plaintext_arrays
boolean (Return Plaintext Arrays)

Responses

Request samples

Content type
application/json
{
  • "disable_gpu": true,
  • "batch_size": 0,
  • "max_input": 0,
  • "return_plaintext_arrays": true
}

Response samples

Content type
application/json
null