Brazilian Portuguese remains underserved by open-source language models, and the few that are safety-aligned rarely document what did not work or validate comparisons with statistical significance. A new arXiv paper on Manacá-1B-Instruct-2 addresses this gap by reporting negative results and methodological pitfalls alongside the model itself.

The paper explicitly presents negative DPO (Direct Preference Optimization) results, indicating that this alignment technique did not yield the expected gains. It also describes a controlled composition scan, likely a systematic test of how the model handles composed or multi-part instructions, and identifies a silent failure mode of the LLM judge—a situation where the judge used for evaluation fails without obvious signs, potentially skewing results.

Because the abstract is brief, the exact details of the scan and the judge failure are not fully specified, but the authors' emphasis on reproducibility and statistical significance suggests these findings are meant to guide future work on Portuguese-language alignment. The paper stands out for openly sharing what did not work, a practice still rare in LLM development.Strict JSON with keys: headline, dek, keyPoints, body. Ensure body is markdown, 2-4 short paragraphs. The source is only the abstract. I'll write accordingly.```json {