COGERPHEREAILABS/meridian-ac-nano

Meridian-AC-Nano: Specialized Academic Proofreading Model

Meridian-AC-Nano, developed by COGERPHERE AI Labs, is a compact (~596M parameter) research proofreading and academic-writing model. It is the flagship model of the MERIDIAN 0.1 research program, fine-tuned from Qwen3-0.6B, and is specifically designed for precise text refinement with a focus on academic integrity.

Key Capabilities

• Research Proofreading: Corrects grammar, spelling, punctuation, and wording while preserving original meaning.
• Grammar Correction: Addresses grammatical and spelling errors in academic texts.
• LaTeX-aware Proofreading: Fixes syntax and formatting errors within LaTeX source code.
• Academic / Research Writing Refinement: Improves academic tone and overall text quality.
• Restrained Text Refinement: Performs minimal, meaning-preserving edits to avoid over-editing.

Performance Highlights

Evaluated on a held-out test set, the model achieved a ROUGE-L score of 0.929 and a BLEU-4 score of 0.8797. Notably, it demonstrated high performance on specific tasks, including 95.24% exact match for LaTeX correction and 91.30% for general proofreading. Training was conducted on a chat-first, multi-task corpus using an Unsloth LoRA pipeline.

Limitations and Intended Use

As a small model, Meridian-AC-Nano is not intended for general-purpose coding, factual knowledge recall, or production medical/legal/financial advice. Its conversational quality is best-effort, and its primary domain is academic and LaTeX-focused proofreading. Users should note the low exact match for open-ended chat tasks (6.06%), which is expected for such responses.