Abstractive Compression

News Summarizer

Length-conditioned abstractive summarization. Designed to distill lengthy Sinhala journalism into concise executive briefs or structured narratives without hallucinating facts outside the source article.

Paradigm

Abstractive

Generates new sentences

Formatting

Llama-3 Chat

Utilizes native chat templates

Length Control

3 Bands

Short, Medium, and Long targets

Adapter Version

v06

Current production deployment

Implementation Details

Length-Conditioned Prompts

Introduced in version v06, the summarizer now natively understands distinct length constraints natively. When an editor asks for a "short" summary, the model leverages its conditioned training to deliver an ultra-concise brief rather than truncating a longer output.

Supported Targets:

  • Short: 2-3 sentence core briefs.
  • Medium: Standard paragraph summary.
  • Long: Multi-paragraph detailed synopsis retaining secondary facts.

Chat Template Utilization

While the Grammar, Headline, and Style tools rely on traditional Alpaca-style instruction prompts, the Summarizer is trained on the native Llama-3 Chat format (<|begin_of_text|><|start_header_id|>...).

This architectural divergence allows the model to better parse complex, multi-turn reasoning and handle document-scale texts with superior contextual retention.