Nomodit is a command-line tool (cli+tui) for linguistic refinement using a fine-tuned version of Google's Gemma 3n model, bringing language processing capabilities to run locally on most consumer hardware (>8GB RAM/VRAM) while fostering a learning environment through transparent reasoning on what changed and other forms of feedback. The model specializes in various language-related tasks including:
- Grammatical Error Correction (GEC)
- Text clarity improvement
- Coherence fixing
- Text neutralization
- Paraphrasing
- Text simplification
nomodit -i "instruction" "text to modify"
Example:
nomodit -i "Fix grammatical errors" "I has went to the store yesterday."
Running nomodit without arguments launches an interactive text user interface with separate input areas for instructions and text.
The interactive TUI provides features to help users understand and learn from text modifications.
- Reasoning Traces: View the model's step-by-step reasoning process for each modification, helping you understand why specific changes were made
- Word-Level Diffs: See precise word-by-word differences between original and modified text, making it easy to identify exactly what was changed
- These features work together to help users identify patterns in their writing mistakes and learn from the corrections.
nomodit is also the name of a series of fine-tuned version of Google's Gemma 3n model using the CoEdit-cot-reasoning dataset which augments the CoEdit dataset with reasoning traces to allow for nomodit models to have reasoning capabilities.
The training runs for fine-tuning the models with unsloth's FastModel framework utlize PEFT techniques like QLoRA and add reasoning capabilites by using GRPO and rubrics as rewards for language tasks which are inherently non-verifiable.
You can find the subsequent notebooks for training the model, creating the reasoning traces using a larger llm and the benchmarking against other models on GEC here.
Note: This work is currently in progress.