PaLM 2: Complete Benchmark Performance Guide

google palm-2 2023-05-10 Proprietary

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

Key Specifications

Vendorgoogle
Versionpalm-2
Release Date2023-05-10
Context Window8192 tokens
Input Modalitiestext
Output Modalitiestext
LicenseProprietary
Documentationhttps://ai.google.dev/gemini-api/docs

Benchmark Performance

Benchmark Score Unit Evaluated At Notes Source
MMLU 68.4 % 2023-05-10 5-shot view
HUMANEVAL 44.8 pass@1 2023-05-10 — view
GSM8K 49 % 2023-05-10 0-shot CoT view
MATH 29.9 % 2023-05-10 0-shot CoT view
BBH 52.9 % 2023-05-10 3-shot CoT view
GPQA 20.3 % 2023-05-10 0-shot view
IFEVAL 44.7 % 2023-05-10 prompt_strict view
ARC 89.2 % 2023-05-10 challenge view
MUSR 29.3 % 2023-05-10 0-shot view
WINOGRANDE 68.6 % 2023-05-10 0-shot view

Pricing

Tier Price Currency
Input$1 / MtokUSD
Output$2 / MtokUSD
Cache Read$0 / MtokUSD
Cache Write$0 / MtokUSD

Source: https://ai.google.dev/pricing · as of 2023-05-10

Compliance

  • Data Residency: US
  • SOC2: ✓
  • HIPAA: ✗
  • GDPR: ✓
  • ISO 27001: ✓

PaLM 2

Model Overview

Google PaLM 2 大型语言模型, 8K 上下文, 改进多语言/推理/编码能力, Bard 初始版本所用模型。

Core Specifications

Vendor Version Release Date Context Window Input Modalities Output Modalities License
Google palm-2 2023-05-10 8K text text Proprietary

Benchmark Performance

Benchmark Score Unit Notes
MMLU (Massive Multitask Language Understanding) 68.4 % 5-shot
HumanEval 44.8 pass@1 —
GSM8K (Grade School Math 8K) 49.0 % 0-shot CoT
MATH 29.9 % 0-shot CoT
BBH (BIG-Bench Hard) 52.9 % 3-shot CoT
GPQA 20.3 % 0-shot
IFEval 44.7 % prompt_strict
ARC 89.2 % challenge
MUSR 29.3 % 0-shot
WinoGrande 68.6 % 0-shot

Pricing

Input Output Cache Read Cache Write
— — — —

per million tokens

Strengths

  • Reliable general-purpose model.

Weaknesses

  • HumanEval 44.8, coding weak.
  • Proprietary, not self-hostable.
  • Context window 8K is limited.

Use Cases

  • General chat and Q&A

References