NbAiLab/nb-notram-llama-3.3-70b-instruct
NbAiLab/nb-notram-llama-3.3-70b-instruct is a 70 billion parameter instruction-tuned language model developed by the National Library of Norway (NB-AiLab), built upon Meta’s Llama-3.3-70B-Instruct. It is specifically fine-tuned to enhance instruction-following in Norwegian Bokmål and Nynorsk, while maintaining strong English performance. This model is notable for its training exclusively on publicly available data and its tendency to produce shorter, more concise answers. It is primarily intended for dialogue systems, assistant-style applications, summarization, and Q&A in Norwegian and English.
Loading preview...
Model Overview
NbAiLab/nb-notram-llama-3.3-70b-instruct is a 70 billion parameter instruction-tuned model from the National Library of Norway (NB-AiLab), based on Meta's Llama-3.3-70B-Instruct. It is part of the "NB-Llama-3.x" and "NoTraM" series, focusing on adapting instruction-tuned models for Norwegian language, culture, and history using only publicly available data.
Key Capabilities
- Multilingual Performance: Strong in Norwegian Bokmål ("nb"), Norwegian Nynorsk ("nn"), and English ("en").
- Instruction Following: Fine-tuned primarily with supervised fine-tuning (SFT) and a light preference optimization (DPO) to improve instruction-following.
- Concise Responses: Tends to generate shorter, more direct answers, reflecting its current instruction-tuning recipe.
- Public Data Training: Uniquely trained exclusively on publicly available datasets, including CulturaX, HPLT monolingual v1.2, Norwegian Colossal Corpus, and Wikipedia, without legal deposit material.
- Data Curation: Utilizes a data selection methodology inspired by "FineWeb" and custom "Corpus Quality Classifiers" to prioritize educational value and linguistic quality.
Intended Use
- Dialogue systems and assistant-style applications in Norwegian and English.
- Summarization and Q&A tasks in Bokmål or Nynorsk.
Limitations
- May produce incorrect or fabricated information.
- Norwegian cultural/historical knowledge can be uneven or "pocketed" (prompt-sensitive).
- Safety alignment is limited; careful evaluation is recommended for specific use cases.