Phi-1.5: Complete Benchmark Performance Guide

other phi-1-5 2023-09-14 MIT

In-depth analysis of Phi-1.5 performance on MMLU, HumanEval, GSM8K, MATH, BBH benchmarks, including pricing and use cases.

Key Specifications

Vendorother
Versionphi-1-5
Release Date2023-09-14
Context Window2048 tokens
Input Modalitiestext
Output Modalitiestext
LicenseMIT
Documentationhttps://huggingface.co/models

Benchmark Performance

Benchmark Score Unit Evaluated At Notes Source
MMLU 58.4 % 2023-09-14 5-shot view
HUMANEVAL 39.2 pass@1 2023-09-14 — view
GSM8K 34.2 % 2023-09-14 0-shot CoT view
MATH 27.2 % 2023-09-14 0-shot CoT view
BBH 49.5 % 2023-09-14 3-shot CoT view
GPQA 16.9 % 2023-09-14 0-shot view
IFEVAL 48 % 2023-09-14 prompt_strict view
ARC 75.9 % 2023-09-14 challenge view
MUSR 30.3 % 2023-09-14 0-shot view
WINOGRANDE 74.1 % 2023-09-14 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 2023-09-14

Compliance

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

Phi-1.5

Model Overview

Microsoft Phi-1.5 1.3B 模型, 2K 上下文, 改进推理能力, 在 1.3B 规模上接近 7B 模型水平。

Core Specifications

Vendor Version Release Date Context Window Input Modalities Output Modalities License
Other phi-1-5 2023-09-14 2K text text MIT

Benchmark Performance

Benchmark Score Unit Notes
MMLU (Massive Multitask Language Understanding) 58.4 % 5-shot
HumanEval 39.2 pass@1 —
GSM8K (Grade School Math 8K) 34.2 % 0-shot CoT
MATH 27.2 % 0-shot CoT
BBH (BIG-Bench Hard) 49.5 % 3-shot CoT
GPQA 16.9 % 0-shot
IFEval 48.0 % prompt_strict
ARC 75.9 % challenge
MUSR 30.3 % 0-shot
WinoGrande 74.1 % 0-shot

Pricing

Input Output Cache Read Cache Write
— — — —

per million tokens

Strengths

  • Reliable general-purpose model.

Weaknesses

  • MMLU 58.4, weak knowledge reasoning.
  • HumanEval 39.2, coding weak.
  • Proprietary, not self-hostable.
  • Context window 2K is limited.

Use Cases

  • General chat and Q&A

References