d-matrix/Llama-3.1-8B-Instruct

Overview of Llama 3.1 8B Instruct

This model is the 8 billion parameter instruction-tuned variant from Meta's Llama 3.1 family, released on July 23, 2024. It leverages an optimized transformer architecture, fine-tuned using supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF) for enhanced helpfulness and safety. Pretrained on over 15 trillion tokens with a knowledge cutoff of December 2023, it features a substantial 128k token context length and Grouped-Query Attention (GQA) for improved inference scalability.

Key Capabilities

• Multilingual Dialogue: Optimized for assistant-like chat in English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai.
• Enhanced Performance: Shows improvements over Llama 3 8B Instruct across various benchmarks, including MMLU (69.4%), HumanEval (72.6% pass@1), and MATH (51.9% final_em).
• Tool Use: Significantly improved tool-use capabilities, scoring 82.6% on API-Bank and 76.1% on BFCL, with support for advanced function calling.
• Code Generation: Strong performance in coding tasks, achieving 72.6% on HumanEval pass@1 and 72.8% on MBPP++.
• Reasoning & Math: Demonstrates robust reasoning (83.4% on ARC-C) and mathematical problem-solving (84.5% on GSM-8K).

Good for

• Commercial and Research Applications: Suitable for a wide range of natural language generation tasks and assistant-like chat.
• Multilingual AI Systems: Ideal for applications requiring robust performance across its 8 explicitly supported languages.
• Agentic Systems: Designed to be integrated into AI systems with additional safety guardrails, supporting tool use and external integrations.
• Synthetic Data Generation: Can be used to generate synthetic data and for model distillation to improve other models.