How to transform raw data into engaging narratives with LLMs
If you have ever wondered how to make the numbers in a report understandable to everyone, Large Language Models (LLM) offer a concrete answer. With a few well-structured prompts, you can have the models generate summaries, visualizations, and even presentation scripts, speeding up the data storytelling process.
Why LLMs are revolutionary for data storytelling in 2026
In today's landscape, LLMs are no longer just textual assistants. Platforms like IFM K2 Horizon offer open-source models ranging from 0.9-B to 375-B parameters, providing flexible options for every computational need. Additionally, tools like UC Berkeley's CUA-Lite and Meta FAIR's AI Research Preference Models (RPMs) simplify the creation of test environments and automatic prompt evaluation, making the end-to-end workflow more reliable.
Choosing the right model for your project
Choose based on three key factors:
- Size and resources:Smaller models (0.9-7-B) are ideal for rapid prototyping, while larger models (150-375-B) excel at analyzing complex datasets.
- License:Apache-2.0 ensures freedom to modify, perfect for custom pipelines.
- Integrations:Verify support for REST APIs and SDKs, especially if you use CUA-Lite for interactive simulations.
How to build a data storytelling workflow in 5 steps
- Data preparation:Normalize CSV, JSON or relational databases into a compact tabular format. Remove duplicates and standardize data types.
- Define the narrative goal:Clarify who the audience is, what decision they need to make, and what tone to use (formal, divulgative, technical).
- Select the LLM tool:Load the chosen model (e.g., a 7-B K2 Horizon) and configure API access.
- Design the prompt:Write a prompt that includes context, stylistic constraints, and the desired output format.
- Iterate and validate:Use RPMs to automatically evaluate the quality of generated text and iterate quickly.
Writing effective prompts: a template
A structured prompt works best. Here is a schema you can copy and adapt:
Generate a concise data story in markdown style for the target audience {audience}, highlighting the main trends {trends} and anomalies {anomalies}. Use a {tone} tone and include a summary table with key metrics. Limit the length to {max_length} words.Replace placeholders with concrete values and add constraints like"avoid technical jargon". Using explicit variables makes the prompt repeatable and easy to test with CUA-Lite.
Practical prompt examples and use cases
1. Monthly sales report
Generate an executive summary of April 2026 sales for the North American market. Highlight growing products, double-digit declines, and regions with above-average margins. Use a formal tone and include a table with total sales, % change, and contribution.2. SaaS product metrics overview
Create a blog-post-style narrative for the Q2 2026 analytics dashboard. The audience