Manacá-1B-Instruct-2
An open-source instruction-tuned large language model for Brazilian Portuguese, serving as the second-generation post-training alignment version of Manacá-1B. It achieves safety alignment through safety-oriented supervised fine-tuning and publicly releases negative DPO results.
Timeline (last 90 days)
Manacá-1B-Instruct-2 is the second-generation post-training aligned version of Manacá-1B, with an SFT corpus of 291,223 entries.
The paper discloses three failed DPO alignment attempts, one controlled data-mix sweep, and a silent failure mode in evaluation methodology.
The safety capabilities of Manacá-1B-Instruct-2 come from a dedicated safety-SFT, whose style is close to the construction method of the LLM-jp AnswerCarefully dataset, with three iterations.
In the controlled sweep of the general-data proportion in safety SFT, the over-refusal curve is non-monotonic, with the inflection point falling in the 40%–50% general-data proportion range.
The paper attributes the gap of Manacá-1B-Instruct-2 in general capabilities to pretraining scale (a comparison of compute budgets anchored on Chinchilla) rather than to the alignment method itself.
In a head-to-head comparison with four external PT-BR models, the lower bound of the Wilson confidence interval for the safety rate of Manacá-1B-Instruct-2 is 94.1%, higher than the upper bounds of all four baseline models.
All of the paper's code, raw data, logs, and the scripts that generate every reported number have been made public.
All Facts (7)
Manacá-1B-Instruct-2 is the second-generation post-training aligned version of Manacá-1B, with an SFT corpus of 291,223 entries.
50%UnverifiedThe paper discloses three failed DPO alignment attempts, one controlled data-mix sweep, and a silent failure mode in evaluation methodology.
50%UnverifiedThe safety capabilities of Manacá-1B-Instruct-2 come from a dedicated safety-SFT, whose style is close to the construction method of the LLM-jp AnswerCarefully dataset, with three iterations.
50%UnverifiedIn the controlled sweep of the general-data proportion in safety SFT, the over-refusal curve is non-monotonic, with the inflection point falling in the 40%–50% general-data proportion range.
50%UnverifiedThe paper attributes the gap of Manacá-1B-Instruct-2 in general capabilities to pretraining scale (a comparison of compute budgets anchored on Chinchilla) rather than to the alignment method itself.
50%UnverifiedIn a head-to-head comparison with four external PT-BR models, the lower bound of the Wilson confidence interval for the safety rate of Manacá-1B-Instruct-2 is 94.1%, higher than the upper bounds of all four baseline models.
50%UnverifiedAll of the paper's code, raw data, logs, and the scripts that generate every reported number have been made public.
50%