T5-FLAN
T5 (Text-To-Text Transfer Transformer) model with the T5-FLAN checkpoint.T5 (Text-To-Text Transfer Transformer) is a pre-trained language model that casts all NLP problems in a unified text-to-text format.
Flan-T5 checkpoints were released as part of the paper "Scaling Instruction-Finetuned Language Models". They were initialized from the T5 1.1 LM-Adapted and instruction-finetuned. The Flan-T5 variants significantly outperform the LM-adapted counterparts. Unlike the vanilla T5 checkpoints, Flan-T5 can also be directly used for few-shot prompting and standard finetuning.
You can fine-tune and deploy this model using the provided Vertex AI Pipeline template. The following model variants are supported: flant5_small
, flant5_base
, flant5_large
, flant5_xl
.
Because T5 casts NLP problems in a unified text-to-text format, it can be used in a wide range of language-related tasks:
This model can be used in Vertex AI Pipelines. Click Open Fine-Tuning Pipeline to access the pipeline template.
The same pipeline template can be reused for all T5 models, but some parameter values differ by variant. The default and recommended values are populated when possible.
Parameters whose values depend on OSS model:
model_template
: Name of the Flan-T5X template model to use. List of supported T5 models are: flant5_small
, flant5_base
(default), flant5_large
, flant5_xl
.
machine_type
: The type of the machine to run the Vertex LLM Trainer Component. The default value is cloud-tpu
.
accelerator_type
and accelerator_count
: For the large model sizes, users should specify more accelerators than smaller model sizes. Subject to quota. If the accelerator is not a TPU, use the supported machined type.
batch_size
: Number of examples per batch. It should a multiple of accelerator_count/num_partitions
.
model_template | num_partitions |
flant5_small | 1 for V3, 1 for A100 and V100 |
flant5_base | 1 for V3, 1 for A100 and V100 |
flant5_large | 2 for V3, 1 for A100 and V100 |
flant5_xl | 4 for V3, 2 for A100 and V100 |
View the GPU compatibility table to see valid values for machineSpec.acceleratorCount
depending on your choices for machineSpec.machineType
and machineSpec.acceleratorType
.
Parameters that depend on input data:
feature_keys
and label_key
: Name of the input features and label in the data.train_data_path
and validation_data_path
: Path to the training and validation data in gs://path/to/file
format.Parameters that depend on user input:
project
, location
: GCP project ID and location.finetuning_steps
: Additional steps on top of the pre-trained steps determined by checkpoint. If not set, use the default value in GIN.task
: generation
.inputs_length
and targets_length
: Length of the input and target features.model_upload_location
: location for uploading model. Default: us-central1
model_display_name
: display name for uploading model. Default: flant5-finetuned-model
encryption_spec_key_name
: Customer-managed encryption key. Has the form: "projects/my-project/locations/my-location/keyRings/my-kr/cryptoKeys/my-key". If this is set, then all resources created by the CustomJob
will be encrypted with the provided encryption key. Currently not supported for TPU.model_encryption_spec_key_name
: Customer-managed encryption key spec for a model. If set, the model and its sub-resources will be secured by the key. Has the form: "projects/my-project/locations/keyRings/cryptoKeys/my-key". The key needs to be in the same region as where the compute resource is created.Resource ID | Release date | Release stage | Description |
---|---|---|---|
t5-flan-001 | 2023-05-10 | Experimental | Initial release |
Die Google Cloud Console konnte keine JavaScript-Quellen von www.gstatic.com laden.
Mögliche Gründe: