Flan-UL2: Complete Benchmark Performance Guide

other flan-ul2 2023-03-03 Apache 2.0

In-depth analysis of Flan-UL2 performance on MMLU, HumanEval, GSM8K, MATH, BBH benchmarks, including pricing and use cases.

Key Specifications

Vendorother
Versionflan-ul2
Release Date2023-03-03
Context Window4096 tokens
Input Modalitiestext
Output Modalitiestext
LicenseApache 2.0
Documentationhttps://huggingface.co/models

Benchmark Performance

BenchmarkScoreUnitEvaluated AtNotesSource
MMLU58.7%2023-03-035-shotview
HUMANEVAL46.8pass@12023-03-03—view
GSM8K40.6%2023-03-030-shot CoTview
MATH22%2023-03-030-shot CoTview
BBH49.2%2023-03-033-shot CoTview
GPQA27.2%2023-03-030-shotview
IFEVAL55.6%2023-03-03prompt_strictview
ARC88.8%2023-03-03challengeview
MUSR37.2%2023-03-030-shotview
WINOGRANDE76.2%2023-03-030-shotview

Pricing

TierPriceCurrency
Input$0.7 / MtokUSD
Output$0.7 / MtokUSD
Cache Read$0 / MtokUSD
Cache Write$0 / MtokUSD

Source: https://huggingface.co/models · as of 2023-03-03

Compliance

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

Flan-UL2

Model Overview

Google Flan-UL2 20B 指令微调模型, 4K 上下文, 基于 UL2 框架, 适合多任务与零样本推理。

Core Specifications

VendorVersionRelease DateContext WindowInput ModalitiesOutput ModalitiesLicense
Otherflan-ul22023-03-034KtexttextApache 2.0

Benchmark Performance

BenchmarkScoreUnitNotes
MMLU (Massive Multitask Language Understanding)58.7%5-shot
HumanEval46.8pass@1—
GSM8K (Grade School Math 8K)40.6%0-shot CoT
MATH22.0%0-shot CoT
BBH (BIG-Bench Hard)49.2%3-shot CoT
GPQA27.2%0-shot
IFEval55.6%prompt_strict
ARC88.8%challenge
MUSR37.2%0-shot
WinoGrande76.2%0-shot

Pricing

InputOutputCache ReadCache Write
————

per million tokens

Strengths

  • Reliable general-purpose model.

Weaknesses

  • MMLU 58.7, weak knowledge reasoning.
  • HumanEval 46.8, coding weak.
  • Proprietary, not self-hostable.
  • Context window 4K is limited.

Use Cases

  • General chat and Q&A

References