Update README.md
Browse files
README.md
CHANGED
|
@@ -21,39 +21,46 @@ tags:
|
|
| 21 |
- gguf
|
| 22 |
---
|
| 23 |
|
| 24 |
-
|
| 25 |
|
| 26 |
-
|
|
|
|
|
|
|
| 27 |
- **Name**: LLaMA 3.2 1B Instruct
|
| 28 |
- **Parameter Size**: 1B (1.23B)
|
| 29 |
-
|
| 30 |
-
## 2. **Quantization Information**
|
| 31 |
-
- **Available Formats**:
|
| 32 |
-
- **ggml-model-q8_0.gguf**: 8-bit quantization for resource efficiency and good performance.
|
| 33 |
-
- **ggml-model-f16.gguf**: Half-precision (16-bit) floating-point format for enhanced precision.
|
| 34 |
-
- **Quantization Library**: llama.cpp
|
| 35 |
-
- **Use Cases**: Recommended for tasks such as multilingual dialogue, text generation, and summarization.
|
| 36 |
-
|
| 37 |
-
## 3. **Model Brief**
|
| 38 |
-
LLaMA 3.2 1B Instruct is a multilingual instruction-tuned language model, optimized for various dialogue tasks. It has been trained on a diverse set of publicly available data and performs well on common NLP benchmarks. The model architecture leverages improved transformer optimizations, making it effective for both text-only and code tasks.
|
| 39 |
-
|
| 40 |
-
- **Purpose**: Multilingual dialogue generation and summarization.
|
| 41 |
- **Model Family**: LLaMA 3.2
|
| 42 |
- **Architecture**: Auto-regressive Transformer with Grouped-Query Attention (GQA)
|
|
|
|
| 43 |
- **Training Data**: A mix of publicly available multilingual data, covering up to 9T tokens.
|
| 44 |
- **Supported Languages**: English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai.
|
| 45 |
- **Release Date**: September 25, 2024
|
| 46 |
- **Context Length**: 128k tokens
|
| 47 |
- **Knowledge Cutoff**: December 2023
|
| 48 |
|
| 49 |
-
|
| 50 |
-
- **
|
| 51 |
-
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 52 |
- **Model Base**: [meta-llama/Llama-3.2-1B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct)
|
|
|
|
|
|
|
|
|
|
| 53 |
|
| 54 |
-
|
| 55 |
-
LLaMA 3.2 1B
|
| 56 |
|
| 57 |
-
- **Testing and Risk Assessment**: Initial testing has focused on English
|
| 58 |
-
- **Limitations**:
|
| 59 |
-
- **Responsible Use Guidelines**:
|
|
|
|
| 21 |
- gguf
|
| 22 |
---
|
| 23 |
|
| 24 |
+
## LLaMA 3.2 1B Instruct
|
| 25 |
|
| 26 |
+
LLaMA 3.2 1B Instruct is a multilingual instruction-tuned language model with 1.23 billion parameters. Designed for diverse multilingual dialogue and summarization tasks, it offers effective performance on a range of NLP benchmarks.
|
| 27 |
+
|
| 28 |
+
### Model Information
|
| 29 |
- **Name**: LLaMA 3.2 1B Instruct
|
| 30 |
- **Parameter Size**: 1B (1.23B)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 31 |
- **Model Family**: LLaMA 3.2
|
| 32 |
- **Architecture**: Auto-regressive Transformer with Grouped-Query Attention (GQA)
|
| 33 |
+
- **Purpose**: Multilingual dialogue generation, text generation, and summarization.
|
| 34 |
- **Training Data**: A mix of publicly available multilingual data, covering up to 9T tokens.
|
| 35 |
- **Supported Languages**: English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai.
|
| 36 |
- **Release Date**: September 25, 2024
|
| 37 |
- **Context Length**: 128k tokens
|
| 38 |
- **Knowledge Cutoff**: December 2023
|
| 39 |
|
| 40 |
+
### Quantized Model Files
|
| 41 |
+
- **Available Formats**:
|
| 42 |
+
- **ggml-model-q8_0.gguf**: 8-bit quantization for resource efficiency and good performance.
|
| 43 |
+
- **ggml-model-f16.gguf**: Half-precision (16-bit) floating-point format for enhanced precision.
|
| 44 |
+
- **Quantization Library**: llama.cpp
|
| 45 |
+
- **Use Cases**: Multilingual dialogue, summarization, and text generation.
|
| 46 |
+
|
| 47 |
+
### Core Library
|
| 48 |
+
LLaMA 3.2 1B Instruct can be deployed using `llama.cpp` or `transformers`, with a focus on streamlined integration into the Hugging Face ecosystem.
|
| 49 |
+
|
| 50 |
+
- **Primary Framework**: `llama.cpp`
|
| 51 |
+
- **Alternate Frameworks**:
|
| 52 |
+
- `transformers` for Hugging Face model support.
|
| 53 |
+
- `vLLM` for optimized inference and low-latency deployments.
|
| 54 |
+
|
| 55 |
+
**Library and Model Links**:
|
| 56 |
- **Model Base**: [meta-llama/Llama-3.2-1B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct)
|
| 57 |
+
- **Models**: [meta-llama/llama-stack](https://github.com/meta-llama/llama-stack)
|
| 58 |
+
- **Inference Support**: [meta-llama/llama](https://github.com/meta-llama/llama)
|
| 59 |
+
- **Quantization**: [ggerganov/llama.cpp](https://github.com/ggerganov/llama.cpp)
|
| 60 |
|
| 61 |
+
### Safety and Responsible Use
|
| 62 |
+
LLaMA 3.2 1B has been designed with safety in mind but may produce biased, harmful, or unpredictable outputs, especially for less-covered languages or specific prompts.
|
| 63 |
|
| 64 |
+
- **Testing and Risk Assessment**: Initial testing has primarily focused on English; coverage for other languages is ongoing.
|
| 65 |
+
- **Limitations**: LLaMA 3.2 may not fully adhere to user instructions or safety guidelines, and may exhibit unexpected behaviors.
|
| 66 |
+
- **Responsible Use Guidelines**: Refer to the [Responsible Use Guide](https://ai.meta.com/llama/responsible-use-guide/) for more details.
|