In this step, we create our first model. You can either use promptranker to help you find the best labels and hypotheses or skip this step with your preferred setting.
Set up your model
Select CREATE NEW from the side navigation
Select your dataset from the dropdown
Select the language of the text
Choose a model type
symanto_fast (for demonstration purposes, live demo, etc.)
symanto (for actual model training with a purpose)
OPTIONAL Select the multi_label option if the text classification task allows the prediction of multiple labels per single text instance. Otherwise, the task is considered as multi-class (or binary if only 2 labels are categorized).
Choose an embedding model
Choose symanto_brain_multilingual only if there is no specific language model for your selected language
OPTIONAL Add a model name or keep the automatically assigned one
Define your labels
As a brief reminder, when defining the text classification task, we need to associate a label text with the label name that we are intending to categorise. This label text, also referred sometimes as hypothesis or prompt, provides a semantic context to the model.
Expland the 'Check or modify labels '
2. Select the number of labels you want to have using the slider
3. Continue with/ without using Promptranker
3. Provide label names, e.g. POSITIVE, NEGATIVE and label texts, e.g. GREAT, BAD. You can include your hypothesis in the label text too, such as THE TEXT IS GREAT or THE TEXT IS BAD
Please note: Hypotheses could also be long and detailed explanations of your labels, e.g. if your labels come from a specific framework you can use the framework’s description of each label with/ or multiple descriptors (adjectives/ nouns for more clarity).
When formulating long hypotheses always think about what would help another person to understand my label better in order to end up with the most clear and understandable hypothesis.
2. OPTIONAL Enable ‘Inspect typical examples’ to see how your model performs
2.1. Use the slider to select up to 10 examples.
2.2. Change the label dropdown to see examples for each of your labels.
2.3. Change the assigned label if necessary
3. Click Create
This will run an unsupervised statistical Bayesian process that will end up outputting a reliability score of the best combination from the provided ones. It is recommended to use that combination for your model, both at zero and few-shot level.
Enable ‘Rank label texts’
Provide label names, e.g. POSITIVE, NEGATIVE
For each label name, provide a list of alternative descriptions. For example, for positive, you can provide GREAT WONDERFUL POSITIVE and for negative, BAD TERRIBLE NEGATIVE
Provide hypothesis variations, e.g. use an empty one and add several different variations such as THIS TEXT IS {}, THE AUTHOR EXPRESSES A {} SENTIMENT , …
The result is a table where each variation is ranked according to a calculated score. The higher the score, the more likely the given labels and hypothesis yield good results in a real scenario.
4. Click Create and the model will be automatically created with the prompts (labels and hypothesis) that scored the top.