FlashLearn makes it easy to build multi-step and multimodal workflows by chaining multiple tasks such as rewriting, sentiment analysis, or, as demonstrated below, dynamic label discovery followed by classification. In this example, we first use a skill to automatically discover labels from a combined dataset of text and images, and then we classify the same inputs using these discovered labels.
Note that this example leverages dynamic skill creation via its parameters. You can also choose to load or learn pre-built skills as needed.
import os
from openai import OpenAI
from flashlearn.skills.discover_labels import DiscoverLabelsSkill
from flashlearn.skills.classification import ClassificationSkill
def main():
os.environ["OPENAI_API_KEY"] = "YOUR_API_KEY"
# Example data (text reviews)
text_reviews = [
{"comment": "Battery life exceeded expectations, though camera was mediocre."},
{"comment": "Arrived late and cracked screen, but customer support was helpful."}
]
# Example data (images + brief text)
# Here, the "image_base64" field simulates an encoded image
image_reviews = [
{"image": "ENCODED_ISSUE_IMAGE", "comment": "WHZ BOTHER WITH IT?"},
{"image": "ENCODED_ISSUE_IMAGE", "comment": "This feature is amazing!! You should charge more!"}
]
# 1) Label Discovery (Aggregates the entire dataset at once)
discover_skill = DiscoverLabelsSkill(model_name="gpt-4o-mini", client=OpenAI())
column_modalities = {"image_base64": "image_base64", "comment": "text"}
tasks_discover = discover_skill.create_tasks(text_reviews + image_reviews)
discovered_labels = discover_skill.run_tasks_in_parallel(tasks_discover)['0']['labels']
print("Discovered labels:", discovered_labels)
# 2) Classification using discovered labels
classify_skill = ClassificationSkill(model_name="gpt-4o-mini", client=OpenAI(), categories=discovered_labels)
tasks_classify = classify_skill.create_tasks(text_reviews + image_reviews)
final_results = classify_skill.run_tasks_in_parallel(tasks_classify)
print("Classification results:", final_results)
if __name__ == "__main__":
main()In this advanced pipeline example:
- We first perform label discovery on a mixed dataset. The DiscoverLabelsSkill aggregates both text and image data into a unified set of labels.
- Then, we use the dynamically discovered labels as the categories for a classification task applied to the same dataset.
- The process orchestrates multi-step workflows seamlessly, passing tasks down the processing chain.
FlashLearn’s flexible framework lets you conveniently replace or enhance parts of your pipeline by loading or learning new skills as your workflow evolves.