AuraTracer智迹闻
中文

EVENT DOSSIER

EuroAlpaca: Task-Preserving Localisation of Instruction Data for European Languages

2026-09-07 12:00 Models across 2 days 🔥 47.2 heat score
2sources
2days unfolding
47.2heat score
2mentions
SummaryAI generated

The researchers introduced a task named EuroAlpaca to maintain a localized pipeline and a multi-language benchmark for European -IFEval, aiming to address the issue of performance degradation in large language models caused by direct translation of instruction data. This pipeline covers 50 European languages and preserves key content and verifies consistency through domain-specific machine translation or reconstructing equivalent instances in the target language. In LoRA experiments with four large models, direct translation of data improved some evaluation scores, but it reduced the accuracy of European -IFEval by 29.8% compared to the unadapted baseline; after adaptation using EuroAlpaca, the accuracy increased by 12.9% compared to the same baseline, successfully reversing the performance degradation and achieving the highest ROUGE-L and F-BERT scores in the Aya evaluation suite.

Related eventsRELATED EVENTS
Key entitiesKEY ENTITIES
EuroAlpacaEuropean-IFEval

Coverage · reports per dayLANGUAGE SPLIT

Entity relations
EuroAlpaca × European-I…2

Integrated timelineUNIFIED TIMELINE

  1. 2026-09-04

    EuroAlpaca: Task-Preserving Localisatio…

    EuroAlpaca 项目发布了一种覆盖 50 种欧洲语言及方言的任务保持本地化管道和近乎平行的资源,并配套了多语言基准 European-IFEval。该管道根据示例应用字段级机器翻译以保留任务关键内容,或重构目标语言等效实例,随后进行…

  2. 2026-09-07

    EuroAlpaca: Task-Preserving Localisatio…

    研究人员推出 EuroAlpaca 任务保持本地化管道及欧洲 -IFEval 多语言基准,覆盖 50 种欧洲语言。该管道根据示例应用分领域机器翻译以保留关键内容,或重构目标语言等价实例,并验证跨领域一致性与语言一致性。在四个大语言模型的 …

SignalsSIGNALS

Keyword heat
  • EuroAlpaca2
  • European-IFEval2

All reports (2)SOURCES

A arXiv cs.CL en 2026-09-04 20:05

EuroAlpaca: Task-Preserving Localisation of Instruction Data for European Languages

EuroAlpaca 项目发布了一种覆盖 50 种欧洲语言及方言的任务保持本地化管道和近乎平行的资源,并配套了多语言基准 European-IFEval。该管道根据示例应用字段级机器翻译以保留任务关键内容,或重构目标语言等效实例,随后进行跨领域一致性和目标语言一致性验证。在针对四种大语言模型的 LoRA 实验中,直接翻译数据虽提升了 Aya 评估套件中的 ROUGE-L 和 F-BERT 分数,但使 European-IFEval 上的准确率较未适配基线下降 29.8%;而采用 EuroAlpaca 适配则使准确率较同一基线提升 12.9%,逆转了直接翻译造成的退化,同时取得了 Aya 上最高的 ROUGE-L 和 F-BERT 分数。

A arXiv cs.CL en 2026-09-07 12:00

EuroAlpaca: Task-Preserving Localisation of Instruction Data for European Languages

研究人员推出 EuroAlpaca 任务保持本地化管道及欧洲 -IFEval 多语言基准,覆盖 50 种欧洲语言。该管道根据示例应用分领域机器翻译以保留关键内容,或重构目标语言等价实例,并验证跨领域一致性与语言一致性。在四个大语言模型的 LoRA 实验中,直接翻译数据虽提升 Aya 评估套件中的 ROUGE-L 和 F-BERT 分数,但使欧洲 -IFEval 准确率较未适配基线下降 29.8%;而采用 EuroAlpaca 适配则使准确率较同一基线提高 12.9%,逆转直接翻译造成的退化并达到 Aya 最高 ROUGE-L 和 F-BERT 分数。