anssio/Llama-Poro-2-8B-Instruct

Poro 2 8B Instruct: Multilingual Conversational AI

Poro 2 8B Instruct is an 8 billion parameter instruction-following model, part of the Poro 2 family, developed through a collaboration between AMD Silo AI, the TurkuNLP group, and HPLT. Built upon the Llama 3.1 8B architecture, this model is specifically designed for conversational AI and instruction following in both Finnish and English.

Key Capabilities & Training:

• Bilingual Proficiency: Optimized for strong performance in both Finnish and English, with a focus on improving Finnish instruction-following.
• Advanced Fine-Tuning: Created via supervised fine-tuning (SFT) on 1.4 million instruction examples (English and Finnish) and further refined with Direct Preference Optimization (DPO) using the HelpSteer3 dataset for enhanced response quality.
• Robust Base: The base Poro 2 8B model underwent continued pretraining on 165 billion tokens of Finnish, English, code, and math data.
• Context Length: Supports a maximum sequence length of 8192 tokens.

Performance Highlights:

• Finnish Excellence: Significantly outperforms Llama 3.1 8B Instruct, Gemma-2-9B-it, and EuroLLM-9B-Instruct on Finnish instruction-following benchmarks (IFEval, MTBench, AlpacaEval 2).
• English Maintenance: Maintains strong English instruction-following performance, comparable to or slightly below Llama 3.1 8B Instruct and Gemma-2-9B-it.
• Win Rate: Achieves an 85% win rate against Llama 3.1 8B Instruct in Finnish MTBench pairwise comparisons.

Good for:

• Conversational AI applications requiring bilingual support (Finnish and English).
• Question answering, information retrieval, and content generation in both languages.
• Educational tools and customer service applications targeting Finnish and English speakers.