Generation Engine

Headline Generator

Length-conditioned headline generation for Sinhala newsrooms. Generates engaging, contextually accurate headlines with strict adherence to designated word-count bands for different editorial layouts.

In-Band Accuracy

79.7%

Overall across 3 bands

Artifact Rate

1.1%

Scraper tags effectively eliminated

Length Control

3 Bands

Short, Medium, and Long formats

Adapter Version

v19

Current production deployment

Architecture & Strategy

Length-Conditioning Mechanics

The model generates headlines specifically tailored to three non-overlapping word-count bands to support dynamic editorial layouts. The band defines the constraints for the prompt and controls minimum/maximum token bounds dynamically during inference.

  • Short Band

    3 to 5 words. Ideal for tight columns, mobile grids, or breaking news tickers.

  • Medium Band

    6 to 7 words. The standard newsroom format for primary article thumbnails.

  • Long Band

    8 to 10 words. Perfect for feature articles and detailed editorial deep-dives.

Artifact Suppression

A common challenge with language models trained on scraped data is the retention of metadata artifacts (e.g., tags like (වීඩියෝ) or [Photos]). Through rigorous regex-based dataset cleaning procedures, we've successfully addressed this.

The v19 Breakthrough

By explicitly stripping trailing tags from every reference headline in the training set and recomputing word bands based on the cleaned text, the model learned to naturally suppress these artifacts.

Result: Artifact rate plummeted ~10x (from 11.2% down to 1.1%) with no regression in length-conditioning accuracy.