Select a task above to see how different instructions affect the embedding space.
About the Visualization
Each point represents a document from the 20 Newsgroups dataset. Colors indicate different categories. Hover over points to see the text preview. The UMAP projection shows how the task instruction reshapes the embedding space.
Task-Specific Embeddings: One model, different perspectives, no fine-tuning
This interactive visualization demonstrates how task instructions reshape
QWEN-3 embedding models. Watch the same 800 documents (from 10 newsgroup categories) reorganize themselves four different ways—all through task instructions.
What are embeddings?
Embeddings are vector representations of text that capture semantic meaning. Similar texts have similar embeddings, enabling machines to understand and compare documents.
What makes task instructions powerful?
By providing an instruction like "Classify the sentiment..." or "Identify the topic...", you guide the model to reorganize the embedding space around task-relevant features. Same 800 documents, four completely different organizations.
The Four Tasks
Default: General-purpose embeddings with no specific instruction
Topic: Optimized for identifying subject matter and themes
Sentiment: Optimized for detecting positive, negative, or neutral tone
Toxicity: Optimized for detecting harmful or toxic content
The Dataset
10 categories from the 20 Newsgroups dataset (politics, science, sports, religion). Each point is one document, colored by category. The 2D projection uses UMAP dimensionality reduction.
How to Use
Click the task buttons to switch between embedding spaces