StableLM Zephyr 3B: Complete Benchmark Performance Guide

other stablelm-zephyr-3b 2024-01-19 Stability AI Community License

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

Key Specifications

Vendorother
Versionstablelm-zephyr-3b
Release Date2024-01-19
Context Window4096 tokens
Input Modalitiestext
Output Modalitiestext
LicenseStability AI Community License
Documentationhttps://huggingface.co/models

Benchmark Performance

Benchmark Score Unit Evaluated At Notes Source
MMLU 44.3 % 2024-01-19 5-shot view
HUMANEVAL 47.1 pass@1 2024-01-19 — view
GSM8K 36.2 % 2024-01-19 0-shot CoT view
MATH 14.9 % 2024-01-19 0-shot CoT view
BBH 43.9 % 2024-01-19 3-shot CoT view
GPQA 21 % 2024-01-19 0-shot view
IFEVAL 55.3 % 2024-01-19 prompt_strict view
ARC 82.1 % 2024-01-19 challenge view
MUSR 23.5 % 2024-01-19 0-shot view
WINOGRANDE 71.6 % 2024-01-19 0-shot view

Pricing

Tier Price Currency
Input$0.12 / MtokUSD
Output$0.12 / MtokUSD
Cache Read$0 / MtokUSD
Cache Write$0 / MtokUSD

Source: https://huggingface.co/models · as of 2024-01-19

Compliance

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

StableLM Zephyr 3B

Model Overview

Stability AI StableLM Zephyr 3B 对话微调模型, 4K 上下文, 基于 StableLM 2 1.6B DPO 微调。

Core Specifications

Vendor Version Release Date Context Window Input Modalities Output Modalities License
Other stablelm-zephyr-3b 2024-01-19 4K text text Stability AI Community License

Benchmark Performance

Benchmark Score Unit Notes
MMLU (Massive Multitask Language Understanding) 44.3 % 5-shot
HumanEval 47.1 pass@1 —
GSM8K (Grade School Math 8K) 36.2 % 0-shot CoT
MATH 14.9 % 0-shot CoT
BBH (BIG-Bench Hard) 43.9 % 3-shot CoT
GPQA 21.0 % 0-shot
IFEval 55.3 % prompt_strict
ARC 82.1 % challenge
MUSR 23.5 % 0-shot
WinoGrande 71.6 % 0-shot

Pricing

Input Output Cache Read Cache Write
— — — —

per million tokens

Strengths

  • Reliable general-purpose model.

Weaknesses

  • MMLU 44.3, weak knowledge reasoning.
  • HumanEval 47.1, coding weak.
  • Proprietary, not self-hostable.
  • Context window 4K is limited.

Use Cases

  • General chat and Q&A

References