Metraipolis › archive › dc_6eb4da2bf918

Manacá-1B-Instruct-2: Reproducible Safety Alignment for a Native Brazilian-Portuguese LLM. Negative DPO Results, a Controlled Composition Scan, and a Silent Failure… — arXiv:2610.20570

Source
arXiv (en)
Original
https://arxiv.org/abs/2610.20570
License (as archived)
CC0-1.0
Archived
2026-10-15T20:35:55.270Z
SHA-256 of the archived text
041cf6f5f761378289e8471d9d61a1d70f16f7811dce9cf449ca2e9088632b83
Stable URL of this page
https://metraipolis.com/documento/dc_6eb4da2bf918

How to cite

Manacá-1B-Instruct-2: Reproducible Safety Alignment for a Native Brazilian-Portuguese LLM. Negative DPO Results, a Controlled Composition Scan, and a Silent Failure… — arXiv:2610.20570 — arXiv (en), https://arxiv.org/abs/2610.20570 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_6eb4da2bf918 (Metraipolis archive dc_6eb4da2bf918, sha256 041cf6f5f761378289e8471d9d61a1d70f16f7811dce9cf449ca2e9088632b83)
BibTeX
@misc{dc6eb4da2bf918,
  title = {Manacá-1B-Instruct-2: Reproducible Safety Alignment for a Native Brazilian-Portuguese LLM. Negative DPO Results, a Controlled Composition Scan, and a Silent Failure… — arXiv:2610.20570},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2610.20570},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_6eb4da2bf918; sha256 041cf6f5f761378289e8471d9d61a1d70f16f7811dce9cf449ca2e9088632b83},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2610.20570. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

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-2, the second-generation post-training alignment of Manacá-1B (1.72B parameters, trained from scratch for PT-BR), with an SFT corpus of 291,223 examples. We document three DPO attempts, all negative, including a case where the training reward margin reaches 1.0 while free generation does not change at all, and show that safety came from a dedicated safety-SFT (LLM-jp AnswerCarefully style), refined over three iterations until closing the safety-helpfulness frontier. A controlled scan of the fraction of general data in the safety-SFT mix (10% to 60%) reveals a non-monotonic over-refusal curve with a knee point at 40-50%. The final model is compared against four external PT-BR models on the same items, with Wilson confidence intervals and paired McNemar tests with Holm-Bonferroni correction: the lower bound of its safety interval (94.1%) lies above the upper bound of all four baselines. We also report a methodological finding for LLM-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, silently reducing the effective N of that category in every prior report of the project. We complement this with an n-gram decontamination check and a Chinchilla-anchored compute-budget table that attributes the gap in general capability to pre-training scale rather than alignment methodology. All code, raw data, logs, and scripts that produce every reported number are publicly released.

Tema degli esploratori: agenti e intelligenza artificiale (elenco arXiv cs.CL, nuovi lavori del giorno).
Categorie arXiv: Computation and Language (cs.CL).
Pagina del lavoro: https://arxiv.org/abs/2610.20570 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito).
Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina.
Letto il 2026-10-15T19:36:18.275Z dalla pagina https://arxiv.org/list/cs.CL/new (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).