Key Specifications
| Vendor | meta |
|---|
| Version | guard-3-8b |
|---|
| Release Date | 2024-07-23 |
|---|
| Context Window | 128000 tokens |
|---|
| Input Modalities | text |
|---|
| Output Modalities | text |
|---|
| License | Llama 3.2 Community License |
|---|
| Documentation | https://llama.meta.com/docs/ |
|---|
Benchmark Performance
| Benchmark | Score | Unit | Evaluated At | Notes | Source |
|---|
| MMLU | 40.1 | % | 2024-07-23 | 5-shot | view |
| HUMANEVAL | 28 | pass@1 | 2024-07-23 | — | view |
| GSM8K | 30.9 | % | 2024-07-23 | 0-shot CoT | view |
| MATH | 8.4 | % | 2024-07-23 | 0-shot CoT | view |
| BBH | 44.1 | % | 2024-07-23 | 3-shot CoT | view |
| GPQA | 27.9 | % | 2024-07-23 | 0-shot | view |
| IFEVAL | 50 | % | 2024-07-23 | prompt_strict | view |
| ARC | 76 | % | 2024-07-23 | challenge | view |
| MUSR | 22.8 | % | 2024-07-23 | 0-shot | view |
| WINOGRANDE | 68 | % | 2024-07-23 | 0-shot | view |
Pricing
| Tier | Price | Currency |
|---|
| Input | $0.18 / Mtok | USD |
| Output | $0.18 / Mtok | USD |
| Cache Read | $0 / Mtok | USD |
| Cache Write | $0 / Mtok | USD |
Source:
https://ai.meta.com/blog/
· as of 2024-07-23
Compliance
- Data Residency: self-host
- SOC2: ✗
- HIPAA: ✗
- GDPR: ✗
- ISO 27001: ✗
Llama Guard 3 8B
Descripción del modelo
Meta Llama Guard 3 8B 内容安全分类模型, 128K 上下文, 支持 MLCommons 安全分类标准。
Especificaciones principales
| Proveedor | Versión | Fecha de lanzamiento | Ventana de contexto | Modalidades de entrada | Modalidades de salida | Licencia |
|---|
| Meta | guard-3-8b | 2024-07-23 | 128K | text | text | Llama 3.2 Community License |
Rendimiento en benchmarks
| Benchmark | Puntuación | Unidad | Notas |
|---|
| MMLU (Massive Multitask Language Understanding) | 40.1 | % | 5-shot |
| HumanEval | 28.0 | pass@1 | — |
| GSM8K (Grade School Math 8K) | 30.9 | % | 0-shot CoT |
| MATH | 8.4 | % | 0-shot CoT |
| BBH (BIG-Bench Hard) | 44.1 | % | 3-shot CoT |
| GPQA | 27.9 | % | 0-shot |
| IFEval | 50.0 | % | prompt_strict |
| ARC | 76.0 | % | challenge |
| MUSR | 22.8 | % | 0-shot |
| WinoGrande | 68.0 | % | 0-shot |
Precios
| Entrada | Salida | Lectura caché | Escritura caché |
|---|
| — | — | — | — |
por millón de tokens
Fortalezas
Debilidades
- MMLU 仅 40.1，知识推理偏弱。
- HumanEval 28.0，代码能力较弱。
- 闭源专有模型，不支持自托管。
Casos de uso
Referencias