Phi-2: Complete Benchmark Performance Guide

other phi-2 2023-12-11 MIT

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

Key Specifications

Vendorother
Versionphi-2
Release Date2023-12-11
Context Window2048 tokens
Input Modalitiestext
Output Modalitiestext
LicenseMIT
Documentationhttps://huggingface.co/models

Benchmark Performance

Benchmark Score Unit Evaluated At Notes Source
MMLU 57.5 % 2023-12-11 5-shot view
HUMANEVAL 39.7 pass@1 2023-12-11 — view
GSM8K 36.8 % 2023-12-11 0-shot CoT view
MATH 20.5 % 2023-12-11 0-shot CoT view
BBH 47.8 % 2023-12-11 3-shot CoT view
GPQA 21.8 % 2023-12-11 0-shot view
IFEVAL 48.8 % 2023-12-11 prompt_strict view
ARC 75.1 % 2023-12-11 challenge view
MUSR 23.7 % 2023-12-11 0-shot view
WINOGRANDE 72.8 % 2023-12-11 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 2023-12-11

Compliance

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

Phi-2

Model Overview

Microsoft Phi-2 2.7B 模型, 2K 上下文, 在 2.7B 规模上推理能力突出, 适合研究/教育用途。

Core Specifications

Vendor Version Release Date Context Window Input Modalities Output Modalities License
Other phi-2 2023-12-11 2K text text MIT

Benchmark Performance

Benchmark Score Unit Notes
MMLU (Massive Multitask Language Understanding) 57.5 % 5-shot
HumanEval 39.7 pass@1 —
GSM8K (Grade School Math 8K) 36.8 % 0-shot CoT
MATH 20.5 % 0-shot CoT
BBH (BIG-Bench Hard) 47.8 % 3-shot CoT
GPQA 21.8 % 0-shot
IFEval 48.8 % prompt_strict
ARC 75.1 % challenge
MUSR 23.7 % 0-shot
WinoGrande 72.8 % 0-shot

Pricing

Input Output Cache Read Cache Write
— — — —

per million tokens

Strengths

  • Reliable general-purpose model.

Weaknesses

  • MMLU 57.5, weak knowledge reasoning.
  • HumanEval 39.7, coding weak.
  • Proprietary, not self-hostable.
  • Context window 2K is limited.

Use Cases

  • General chat and Q&A

References