趋近智
参数
3B
上下文长度
32.768K
模态
Text
架构
Dense
许可证
TII Falcon-LLM License 2.0
发布日期
17 Dec 2024
知识截止
-
注意力结构
Multi-Query Attention
隐藏维度大小
1536
层数
32
注意力头
48
键值头
1
激活函数
SwigLU
归一化
RMS Normalization
位置嵌入
ROPE
不同量化方法和上下文大小的显存要求
Falcon-3B is a member of the Falcon 3 family of decoder-only large language models, developed by the Technology Innovation Institute (TII). This model variant, with 3 billion parameters, is engineered for efficient deployment on various hardware, including systems with limited resources such as laptops and single GPUs. Its primary purpose is to deliver robust performance across a spectrum of natural language processing tasks, focusing on reasoning, language understanding, instruction following, code generation, and mathematics. The Falcon-3B model also supports multilingual capabilities, specifically English, French, Spanish, and Portuguese.
The architectural foundation of Falcon-3B is a transformer-based causal decoder-only design. It incorporates several innovations to enhance efficiency and performance. Notably, it utilizes Grouped Query Attention (GQA), a mechanism that optimizes inference speed and reduces Key-Value (KV) cache memory consumption by sharing parameters among attention heads. The model employs SwiGLU as its activation function and RMSNorm for normalization, contributing to stable and effective learning. Positional embeddings are handled using Rotary Positional Embeddings (RoPE) to support extended context comprehension. Furthermore, the model leverages FlashAttention 2 for accelerated attention computations and features a high vocabulary size of 131,000 tokens, enabling improved compression and downstream performance.
Falcon-3B, along with its instruction-tuned counterpart, has been developed using techniques such as pruning and knowledge distillation from the larger Falcon3-7B-Base model, resulting in an efficient and performant compact model. The base variant supports a context length of 8,000 tokens, while the instruction-tuned variant extends this capability to 32,000 tokens, allowing it to process and generate responses for longer and more complex inputs. This design paradigm makes Falcon-3B a suitable choice for applications requiring advanced AI functionalities in environments where computational resources are a consideration.
The TII Falcon model family comprises causal decoder-only language models (7B, 40B). Their architecture, adapted from GPT-3, integrates rotary positional embeddings, Multi-Query Attention for inference efficiency, and FlashAttention for accelerated operations. Models are trained on the RefinedWeb dataset.
排名适用于本地LLM。
没有可用的 Falcon-3B 评估基准。