Phi-3 Mini: Complete Benchmark Performance Guide

other phi-3-mini 2024-04-22 MIT

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

Key Specifications

Vendorother
Versionphi-3-mini
Release Date2024-04-22
Context Window4096 tokens
Input Modalitiestext
Output Modalitiestext
LicenseMIT
Documentationhttps://huggingface.co/models

Benchmark Performance

Benchmark Score Unit Evaluated At Notes Source
MMLU 43.6 % 2024-04-22 5-shot view
HUMANEVAL 26.9 pass@1 2024-04-22 — view
GSM8K 40.9 % 2024-04-22 0-shot CoT view
MATH 17.9 % 2024-04-22 0-shot CoT view
BBH 55.9 % 2024-04-22 3-shot CoT view
GPQA 20.7 % 2024-04-22 0-shot view
IFEVAL 54.9 % 2024-04-22 prompt_strict view
ARC 84.4 % 2024-04-22 challenge view
MUSR 33.5 % 2024-04-22 0-shot view
WINOGRANDE 67.1 % 2024-04-22 0-shot view

Pricing

Tier Price Currency
Input$0.1 / MtokUSD
Output$0.1 / MtokUSD
Cache Read$0 / MtokUSD
Cache Write$0 / MtokUSD

Source: https://huggingface.co/models · as of 2024-04-22

Compliance

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

Phi-3 Mini

Model Overview

Microsoft Phi-3 Mini 3.8B 轻量模型, 4K 上下文 (可扩展 128K), 训练数据高质量, 性能超越 Llama 2 7B。

Core Specifications

Vendor Version Release Date Context Window Input Modalities Output Modalities License
Other phi-3-mini 2024-04-22 4K text text MIT

Benchmark Performance

Benchmark Score Unit Notes
MMLU (Massive Multitask Language Understanding) 43.6 % 5-shot
HumanEval 26.9 pass@1 —
GSM8K (Grade School Math 8K) 40.9 % 0-shot CoT
MATH 17.9 % 0-shot CoT
BBH (BIG-Bench Hard) 55.9 % 3-shot CoT
GPQA 20.7 % 0-shot
IFEval 54.9 % prompt_strict
ARC 84.4 % challenge
MUSR 33.5 % 0-shot
WinoGrande 67.1 % 0-shot

Pricing

Input Output Cache Read Cache Write
— — — —

per million tokens

Strengths

  • Reliable general-purpose model.

Weaknesses

  • MMLU 43.6, weak knowledge reasoning.
  • HumanEval 26.9, coding weak.
  • Proprietary, not self-hostable.
  • Context window 4K is limited.

Use Cases

  • General chat and Q&A

References