لخّصلي

3 Meta AI's Large Language Model (LLaMA) The LLaMA is a powerful family of autoregressive language models designed to provide efficient, high-quality language understanding for both general and specialized applications.In 2024, Meta introduced LLaMA 3, representing a new generation of foundation models designed to support multilingual capabilities, advanced reasoning, tool use, and multimodal functionality across text, image, and speech.The dialogue-optimized variant, LLaMA 2-Chat, incorporates techniques such as red-teaming, safety tuning, and rejection sampling to improve alignment with human expectations for helpfulness and safety [251].LLaMA models are based on a transformer architecture and are pre-trained on a mixture of publicly available data sources, including Common Crawl, C4, GitHub, Wikipedia, books, and scientific articles.For example, LLaMA 3.2, which reaches a maximum of 405 billion parameters, shows strong capabilities in handling long-context applications and demonstrates reduced rates of hallucination.The dataset comprises around 1.4 trillion tokens, carefully curated to prioritize high-quality content and filtered to avoid duplicates and irrelevant information.

3 Meta AI’s Large Language Model (LLaMA)
The LLaMA is a powerful family of autoregressive language models designed to provide efficient, high-quality language
understanding for both general and specialized applications. Since the release of its first version in February 2023,
Meta has developed several versions, including models with up to 65.2 billion parameters. LLaMA models are based
on a transformer architecture and are pre-trained on a mixture of publicly available data sources, including Common
Crawl, C4, GitHub, Wikipedia, books, and scientific articles. The dataset comprises around 1.4 trillion tokens, carefully
curated to prioritize high-quality content and filtered to avoid duplicates and irrelevant information. Parameter sizes in
the first version included 7 billion, 13 billion, 33 billion, and 65 billion [250].
The improved LLaMA 2 model, released later, introduced enhanced context sensitivity, dialogue capabilities, and
alignment with user preferences. LLaMA 2 scales up to 69 billion parameters and is trained on 2 trillion tokens of
publicly available data, allowing it to handle longer, more complex inputs. Key advancements include grouped-query
attention (GQA) for larger models, which enhances inference scalability. LLaMA 2 also includes specialized finetuning for dialogue through supervised training and Reinforcement Learning with Human Feedback (RLHF). The
dialogue-optimized variant, LLaMA 2-Chat, incorporates techniques such as red-teaming, safety tuning, and rejection
sampling to improve alignment with human expectations for helpfulness and safety [251].
In 2024, Meta introduced LLaMA 3, representing a new generation of foundation models designed to support multilingual capabilities, advanced reasoning, tool use, and multimodal functionality across text, image, and speech. With up to
70.6 billion parameters, LLaMA 3 was trained on a 15.6 trillion token dataset. Compared to earlier versions, LLaMA 3
features an improved data curation pipeline for pre-training and post-training, enhancing its language understanding
and complex reasoning skills. Subsequent updates, including LLaMA 3.1 and LLaMA 3.2, have added further features
and fine-tuned performance. For example, LLaMA 3.2, which reaches a maximum of 405 billion parameters, shows
strong capabilities in handling long-context applications and demonstrates reduced rates of hallucination.
Similar to other large language models, LLaMA models offer a range of opportunities when integrated with smart
environments, particularly for handling complex tasks. For example, the study in [252] introduces “Harmony”, an
intelligent home assistant system powered by the LLaMA 3-8B model, designed to maintain user privacy and operate
locally without requiring an Internet connection. Harmony’s architecture consists of three components: a Message
Handler, an Agent, and a Controller. The Message Handler processes sensor data and user commands, inferring user
needs through both short-term and long-term memory functions. The Agent then formulates action plans based on
these inferences, consulting memory for contextual relevance. Finally, the Controller translates the Agent’s plans
into JSON-formatted commands to control devices, ensuring actions align with the smart home’s setup. Harmony
demonstrates high accuracy (about 90%) in executing tasks, comparable to cloud-based solutions such as GPT-4, while
significantly reducing hallucination rates. Harmony exemplifies how small-scale LLaMA models can enable effective,
privacy-preserving applications for smart spaces without even the need for cloud resources.