AI TrackerFrontier desk · AEST
Back to news
normalaitech

Manacá-1B-Instruct-2: Reproducible Safety Alignment for a Native Brazilian-Portuguese LLM. Negative DPO Results, a Controlled Composition Scan, and a Silent Failure Mode of the LLM Judge

arXiv cs.CL · 15 Oct 2026, 03:53 AEST

The Manacá-1B-Instruct-2 model aims to improve safety alignment for Brazilian Portuguese language models. It documents negative results from DPO attempts and highlights a silent failure mode in the evaluation process.

Key facts

Read aloud

The Manacá-1B-Instruct-two model aims to improve safety alignment for Brazilian Portuguese language models. It documents negative results from DPO attempts and highlights a silent failure mode in the evaluation process. 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. We present Manacá-1B-Instruct-two, the second-generation post-training alignment of Manacá-1B one.72B parameters, trained from scratch for PT-BR, with an SFT corpus of two hundred ninety one thousand two hundred twenty three examples. We document three DPO attempts, all negative, including a case where the training reward margin reaches one point zero while free generation does not change at all. A controlled scan of the fraction of general data in the safety-SFT mix ten percent to sixty percent reveals a non-monotonic over-refusal curve with a knee point at forty-fifty percent. We also report a methodological finding for L L M-judge evaluation: on two of six safety items in MT-Bench-PT, the judge categorically refuses to assign a score regardless of the evaluated response.

Sources in this cluster