Llama Guard 2 8B: Complete Benchmark Performance Guide

meta guard-2-8b 2024-04-18 Llama 3 Community License

In-depth analysis of Llama Guard 2 8B performance on MMLU, HumanEval, GSM8K, MATH, BBH benchmarks, including pricing and use cases.

Key Specifications

Vendormeta
Versionguard-2-8b
Release Date2024-04-18
Context Window8192 tokens
Input Modalitiestext
Output Modalitiestext
LicenseLlama 3 Community License
Documentationhttps://llama.meta.com/docs/

Benchmark Performance

BenchmarkScoreUnitEvaluated AtNotesSource
MMLU51%2024-04-185-shotview
HUMANEVAL26.6pass@12024-04-18—view
GSM8K46.3%2024-04-180-shot CoTview
MATH26.3%2024-04-180-shot CoTview
BBH47.6%2024-04-183-shot CoTview
GPQA18%2024-04-180-shotview
IFEVAL51.6%2024-04-18prompt_strictview
ARC76.3%2024-04-18challengeview
MUSR34.9%2024-04-180-shotview
WINOGRANDE66.5%2024-04-180-shotview

Pricing

TierPriceCurrency
Input$0.18 / MtokUSD
Output$0.18 / MtokUSD
Cache Read$0 / MtokUSD
Cache Write$0 / MtokUSD

Source: https://ai.meta.com/blog/ · as of 2024-04-18

Compliance

  • Data Residency: self-host
  • SOC2: ✗
  • HIPAA: ✗
  • GDPR: ✗
  • ISO 27001: ✗

Llama Guard 2 8B

Model Overview

Meta Llama Guard 2 8B 内容安全分类模型, 8K 上下文, 用于检测输入输出中的有害内容。

Core Specifications

VendorVersionRelease DateContext WindowInput ModalitiesOutput ModalitiesLicense
Metaguard-2-8b2024-04-188KtexttextLlama 3 Community License

Benchmark Performance

BenchmarkScoreUnitNotes
MMLU (Massive Multitask Language Understanding)51.0%5-shot
HumanEval26.6pass@1—
GSM8K (Grade School Math 8K)46.3%0-shot CoT
MATH26.3%0-shot CoT
BBH (BIG-Bench Hard)47.6%3-shot CoT
GPQA18.0%0-shot
IFEval51.6%prompt_strict
ARC76.3%challenge
MUSR34.9%0-shot
WinoGrande66.5%0-shot

Pricing

InputOutputCache ReadCache Write
————

per million tokens

Strengths

  • Reliable general-purpose model.

Weaknesses

  • MMLU 51.0, weak knowledge reasoning.
  • HumanEval 26.6, coding weak.
  • Proprietary, not self-hostable.
  • Context window 8K is limited.

Use Cases

  • General chat and Q&A

References