Key Specifications

Vendorother
Versionfalcon-40b
Release Date2023-05-25
Context Window2048 tokens
Input Modalitiestext
Output Modalitiestext
LicenseTII Falcon LLM License
Documentationhttps://huggingface.co/models

Benchmark Performance

BenchmarkScoreUnitEvaluated AtNotesSource
MMLU56.7%2023-05-255-shotview
HUMANEVAL39.7pass@12023-05-25view
GSM8K61.3%2023-05-250-shot CoTview
MATH21.3%2023-05-250-shot CoTview
BBH50.6%2023-05-253-shot CoTview
GPQA21.1%2023-05-250-shotview
IFEVAL57.2%2023-05-25prompt_strictview
ARC82.9%2023-05-25challengeview
MUSR38.1%2023-05-250-shotview
WINOGRANDE65.6%2023-05-250-shotview

Pricing

TierPriceCurrency
Input$0.4 / MtokUSD
Output$0.4 / MtokUSD
Cache Read$0 / MtokUSD
Cache Write$0 / MtokUSD

Source: https://huggingface.co/models · as of 2023-05-25

Compliance

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

Falcon 40B

모델 개요

TII Falcon 40B 中型开源模型, 2K 上下文, 400 亿参数, 在 40B 规模上曾领先 Llama 2 70B。

핵심 사양

공급업체버전출시일컨텍스트 창입력 모달리티출력 모달리티라이선스
Otherfalcon-40b2023-05-252KtexttextTII Falcon LLM License

벤치마크 성능

벤치마크점수단위비고
MMLU (Massive Multitask Language Understanding)56.7%5-shot
HumanEval39.7pass@1
GSM8K (Grade School Math 8K)61.3%0-shot CoT
MATH21.3%0-shot CoT
BBH (BIG-Bench Hard)50.6%3-shot CoT
GPQA21.1%0-shot
IFEval57.2%prompt_strict
ARC82.9%challenge
MUSR38.1%0-shot
WinoGrande65.6%0-shot

가격

입력출력캐시 읽기캐시 쓰기

백만 토큰당

강점

  • 可靠的通用模型。

약점

  • MMLU 仅 56.7,知识推理偏弱。
  • HumanEval 39.7,代码能力较弱。
  • 闭源专有模型,不支持自托管。
  • 上下文窗口 2K 偏小。

사용 사례

  • 通用对话与问答

참고문헌