论文精选 81arXiv

Manacá-1B-Instruct-2: Reproducible Safety Alignment for a Native Brazilian-Portuguese LLM. Negative DPO Results, a Controlled Composition Scan, and a...· Manacá-1B-Instruct-2: 巴西葡萄牙语大模型的安全对齐

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-mo

领域：cs.CL作者：Bruno Leonardo Santos Menezes、Carlos Leonardo Souza Cardoso、Fabio André Machado Porto

#cs.CL

📎 发布于 10-15 01:53🔗 阅读原文（arxiv.org）📄 PDF

Low-Bit N-Gram Embeddings for Language Models· 大模型低精度 n-gram 嵌入

Many recent open-weight language models use n-gram embedding tables to improve performance…

领域：cs.CL作者：Stephen Chung、Wenyu Du

📎 arXiv🕒 10-23 01:56🔗 arxiv.org

论文精选 82

Can an Agent Debug a Web Corpus? Agentic Auditing of Slovak FineWeb2· LLM 能否调试网络语料库？

Multilingual pipelines such as FineWeb2 apply one recipe across a thousand languages, leav…

领域：cs.CL作者：Marek Šuppa、Ivan Vykopal、Jaroslav Kopčan

📎 arXiv🕒 10-23 01:52🔗 arxiv.org

论文精选 83

Bilingual Language Models Capture Cross-Linguistic Influence in Second-Language Reading· 双语模型捕捉二语阅读的跨语言影响

While monolingual language model (LM) surprisal is widely established as a predictor of re…

领域：cs.CL作者：Abderrahmane Issam、Irene E. Winther、Stefan L. Frank

📎 arXiv🕒 10-23 01:42🔗 arxiv.org

论文精选 82

How Ground Truth Shapes Retrieval Scores: Evidence-Aware Evaluation of a Retrieval-Augmented Generation System for Breast Lesions· 乳腺病变管理文献中的检索系统评估

Purpose: To describe an evidence-aware retrieval evaluation protocol for a clinical retrie…

领域：cs.CL作者：Miguel Mochizuki Silva、Samuel Ismael、Mateus C. Dias

📎 arXiv🕒 10-23 01:36🔗 arxiv.org

论文精选 80

Would an LLM Match into a Residency? Uncertainty, Reasoning and Image Use of Open Language Models on the Spanish MIR Examination· 大语言模型能通过医学考试吗？

Accuracy on medical examinations is common evidence offered for the clinical use of langua…

领域：cs.CL作者：Miguel Fernandez-de-Retana、Unai Zulaika、Ruben Sanchez-Corcuera

📎 arXiv🕒 10-23 01:23🔗 arxiv.org

本站内容由 LLM 精选聚合，原文版权归 arXiv 所有 · 摘录仅供参考