Phi-1: Complete Benchmark Performance Guide

other phi-1 2023-09-14 MIT

In-depth analysis of Phi-1 performance on MMLU, HumanEval, GSM8K, MATH, BBH benchmarks, including pricing and use cases.

Key Specifications

Vendorother
Versionphi-1
Release Date2023-09-14
Context Window2048 tokens
Input Modalitiestext
Output Modalitiestext
LicenseMIT
Documentationhttps://huggingface.co/models

Benchmark Performance

BenchmarkScoreUnitEvaluated AtNotesSource
MMLU68.1%2023-09-145-shotview
HUMANEVAL88.1pass@12023-09-14—view
GSM8K51.6%2023-09-140-shot CoTview
MATH25.2%2023-09-140-shot CoTview
BBH69.9%2023-09-143-shot CoTview
GPQA22.1%2023-09-140-shotview
IFEVAL53.8%2023-09-14prompt_strictview
ARC88.4%2023-09-14challengeview
MUSR49%2023-09-140-shotview
WINOGRANDE74.5%2023-09-140-shotview

Pricing

TierPriceCurrency
Input$0.1 / MtokUSD
Output$0.1 / MtokUSD
Cache Read$0 / MtokUSD
Cache Write$0 / MtokUSD

Source: https://huggingface.co/models · as of 2023-09-14

Compliance

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

Phi-1

Model Overview

Microsoft Phi-1 1.3B 代码模型, 2K 上下文, 专为代码生成训练, 在 1.3B 规模上 HumanEval 表现突出。

Core Specifications

VendorVersionRelease DateContext WindowInput ModalitiesOutput ModalitiesLicense
Otherphi-12023-09-142KtexttextMIT

Benchmark Performance

BenchmarkScoreUnitNotes
MMLU (Massive Multitask Language Understanding)68.1%5-shot
HumanEval88.1pass@1—
GSM8K (Grade School Math 8K)51.6%0-shot CoT
MATH25.2%0-shot CoT
BBH (BIG-Bench Hard)69.9%3-shot CoT
GPQA22.1%0-shot
IFEval53.8%prompt_strict
ARC88.4%challenge
MUSR49.0%0-shot
WinoGrande74.5%0-shot

Pricing

InputOutputCache ReadCache Write
————

per million tokens

Strengths

  • HumanEval 88.1, excellent code generation.

Weaknesses

  • Proprietary, not self-hostable.
  • Context window 2K is limited.

Use Cases

  • Code generation and debugging

References