Phi-3 Vision: Complete Benchmark Performance Guide

other phi-3-vision 2024-05-21 MIT

In-depth analysis of Phi-3 Vision performance on MMLU, HumanEval, GSM8K, MATH, BBH benchmarks, including pricing and use cases.

Key Specifications

Vendorother
Versionphi-3-vision
Release Date2024-05-21
Context Window4096 tokens
Input Modalitiestext, image
Output Modalitiestext
LicenseMIT
Documentationhttps://huggingface.co/models

Benchmark Performance

BenchmarkScoreUnitEvaluated AtNotesSource
MMLU56.4%2024-05-215-shotview
HUMANEVAL48.2pass@12024-05-21—view
GSM8K30%2024-05-210-shot CoTview
MATH24.3%2024-05-210-shot CoTview
BBH40.1%2024-05-213-shot CoTview
GPQA17.4%2024-05-210-shotview
IFEVAL56.4%2024-05-21prompt_strictview
ARC77.4%2024-05-21challengeview
MUSR32.1%2024-05-210-shotview
WINOGRANDE74.7%2024-05-210-shotview

Pricing

TierPriceCurrency
Input$0.2 / MtokUSD
Output$0.2 / MtokUSD
Cache Read$0 / MtokUSD
Cache Write$0 / MtokUSD

Source: https://huggingface.co/models · as of 2024-05-21

Compliance

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

Phi-3 Vision

Model Overview

Microsoft Phi-3 Vision 4.2B 多模态模型, 4K 上下文, 支持图像理解, 在 4B 规模上多模态能力领先。

Core Specifications

VendorVersionRelease DateContext WindowInput ModalitiesOutput ModalitiesLicense
Otherphi-3-vision2024-05-214Ktext, imagetextMIT

Benchmark Performance

BenchmarkScoreUnitNotes
MMLU (Massive Multitask Language Understanding)56.4%5-shot
HumanEval48.2pass@1—
GSM8K (Grade School Math 8K)30.0%0-shot CoT
MATH24.3%0-shot CoT
BBH (BIG-Bench Hard)40.1%3-shot CoT
GPQA17.4%0-shot
IFEval56.4%prompt_strict
ARC77.4%challenge
MUSR32.1%0-shot
WinoGrande74.7%0-shot

Pricing

InputOutputCache ReadCache Write
————

per million tokens

Strengths

  • Input Modalities: text, image, audio.

Weaknesses

  • MMLU 56.4, weak knowledge reasoning.
  • HumanEval 48.2, coding weak.
  • Proprietary, not self-hostable.
  • Context window 4K is limited.

Use Cases

  • Vision and image understanding

References