Llama-3.2-3B-abliterated: Multilingual Dialogue and Agentic LLM

mylesgoose/Llama-3.2-3B-abliterated is a 3.2 billion parameter model from Meta's Llama 3.2 collection, designed for multilingual text-in/text-out generative tasks. This instruction-tuned variant is optimized for dialogue, agentic retrieval, and summarization, leveraging an optimized transformer architecture with Grouped-Query Attention (GQA) for efficient inference. It was trained on up to 9 trillion tokens of publicly available data, with a knowledge cutoff of December 2023, and incorporates knowledge distillation from larger Llama 3.1 models.

Key Capabilities

• Multilingual Proficiency: Officially supports English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai, with broader training across other languages.
• Dialogue Optimization: Specifically fine-tuned for assistant-like chat applications.
• Agentic Tasks: Excels in knowledge retrieval, summarization, and query/prompt rewriting.
• Long Context: Supports a substantial context length of 32768 tokens.
• Performance: Outperforms many open-source and closed chat models on common industry benchmarks, particularly in multilingual contexts, achieving 58.2 EM on MGSM (CoT) and 63.4 MMLU (5-shot) in English.

Good for

• Developing multilingual chat assistants and conversational AI.
• Applications requiring efficient knowledge retrieval and summarization across multiple languages.
• Building mobile AI-powered writing assistants.
• Research and commercial use in diverse linguistic environments, provided compliance with the Llama 3.2 Community License.