Lispector: fine-tuning language models for grammar and spelling correction in Brazilian Portuguese
Original (Portuguese): Lispector: Fine-tuning de Modelos de Linguagem para Revisão Gramatical e Ortográfica em Português Brasileiro
Paper · 2026 · Andresa Medeiros, Felipe Iszlaji, Claudia Sarmiento-Moreno et al.
Summary
Production proofreading tools for Brazilian Portuguese need corrections that are predictable and minimal, without unnecessary rewriting. The authors, from the company Clarice.ai, ask whether smaller fine-tuned LLMs can match or beat larger general-purpose models used only with few-shot prompting. They built Lispector, two supervised fine-tuned models: Lispector large, based on GPT-4.1, and Lispector small, based on GPT-4.1 nano. Training used 4,500 pairs of real user texts from Clarice.ai, each with a reference correction written by linguists. The baselines were GPT-5, GPT-4.1, GPT-5 nano and GPT-4.1 nano, all given the same few-shot prompts. Both Lispector variants beat every baseline on BLEU, GLEU, METEOR and ROUGE. The small variant scored about as well as the large one. Lispector small also had the lowest median latency of all models tested, while GPT-5 models took tens of seconds per request. The study suggests that task-specific fine-tuning of compact models is the better trade-off when cost and latency matter. However, the models, data and prompts are proprietary and have not been released.
Key points
- Dataset: 4,500 original/corrected pairs from real Clarice.ai users, split into 2,500 for training, 1,000 for validation and 1,000 for testing, with no overlap. The average record is about 130 tokens.
- Training setup: both models used 3 epochs, batch size 6 and a learning-rate multiplier of 0.1. Lispector small was trained on about 792,000 tokens.
- Quality: both Lispector models reached 83% BLEU and 89% GLEU, against 70–77% and 77–81% for the prompted baselines. This is a gain of up to +13 BLEU and +12 GLEU. They also scored 97% on METEOR and ROUGE-1 and 89–90% on ROUGE-2.
- Latency: Lispector small had a median of 1,378 ms and a P95 of 2,601 ms, and Lispector large had a median of 2,040 ms. GPT-5 had a median of about 41 s and a P95 of about 163 s.
- Limitations: the paper includes no comparison with LanguageTool and did not optimize the baseline prompts. Evaluation relied only on automatic metrics, with no human assessment. Lispector is already in production on Clarice.ai.
Fields, topics and concepts
📚 Shelves 2The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 works
🧭 Topics 2Research subjects and languages the work is about.Language models & pretraining · 3,894 worksGrammatical error correction & learner language · 162 works
Cite
Cite as Andresa Medeiros, Felipe Iszlaji, Claudia Sarmiento-Moreno, Camila Muniz, Larissa Ponciano, Larissa Dejigov, Ronald Monteiro, Pedro Kretikouski and Guilherme Chaves. 2026. Lispector: Fine-tuning de Modelos de Linguagem para Revisão Gramatical e Ortográfica em Português Brasileiro. Proceedings of the 17th International Conference on Computational Processing of Portuguese (PROPOR 2026) - Vol. 2.
@inproceedings{medeiros2026lispector,
title = {{Lispector: Fine-tuning de Modelos de Linguagem para Revisão Gramatical e Ortográfica em Português Brasileiro}},
author = {Andresa Medeiros, Felipe Iszlaji, Claudia Sarmiento-Moreno, Camila Muniz, Larissa Ponciano, Larissa Dejigov, Ronald Monteiro, Pedro Kretikouski and Guilherme Chaves},
booktitle = {Proceedings of the 17th International Conference on Computational Processing of Portuguese (PROPOR 2026) - Vol. 2},
year = {2026},
url = {https://aclanthology.org/2026.propor-2.8/}
}References
- Satanjeev Banerjee and Alon Lavie. 2005. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Transla-
Updated
This page in Markdown → https://md.research.elamur.ai/free/en/papers/medeiros-2026-lispector-fine-tuning-de-modelos/