Linguistic Register Control

Style & Tone Rewriter

Dynamically shift the linguistic register of Sinhala news articles. Rewrite content instantly into one of five distinct newspaper styles, matching the target demographic without altering the underlying facts.

Registers

5

Distinct editorial tone styles

Latency

<800ms

Adapter switching overhead

Fact Fidelity

High

Preserves journalistic accuracy

Adapter Version

v07

Current production deployment

Supported Editorial Styles

Formal News

Standard broadcast and broadsheet tone. Objective, structured, and strictly adheres to formal journalistic syntax.

Sports Reporting

Dynamic and action-oriented vocabulary. Adapts verbs and pacing to match the excitement of live sports coverage.

Youth & Culture

Engaging, modern vernacular tailored for younger demographics and digital-first tabloid formats.

Editorial Opinion

Authoritative, persuasive, and reflective. Enhances sentences with richer adjectives to convey a strong viewpoint.

Feature Story

Narrative-driven and descriptive. Prioritizes storytelling flow and immersive language for long-form reading.

Zero-Degradation Hot Swapping

Unlike traditional monolithic deployments where every distinct task requires an entirely separate 8-billion-parameter model to be loaded into VRAM, SinAi utilizes a shared pre-merged foundational base (SinLLaMA-merged-base).

The Style Rewriter, alongside the grammar and summarization tools, is deployed as a highly optimized, lightweight Low-Rank Adaptation (LoRA) adapter. Our inference server (serve_sinai.py) dynamically swaps these adapters into the active context in milliseconds.

Efficiency Gain

This architecture allows a single GPU to serve all four major SinAi AI features concurrently without memory exhaustion, supporting rapid tone adjustments without user-facing loading delays.