RWKV

by RWKV Foundation / BlinkDL community

System Card

OrganizationRWKV Foundation / BlinkDL community

Released2023-05

Architectureexternal-memory-network / Linear-attention RNN with receptance-weighted key-value

DetailsCombines Transformer-style parallelizable training with RNN-style linear-time inference through a receptance-weighted key-value (RWKV) attention. Constant memory, no KV cache, unbounded context length.

Parameters—

Domainlong-context

Open SourceYes

PaperView Paper

WebsiteVisit

CodeRepository

rnnlinear-attentionconstant-memoryefficient

Capability Profile

Benchmark Scores

6 of 14 benchmarks

Long-Context Retrieval

5/5

75.887p

71.215p

73.927p

603p

72.616p

Multi-Turn Recall

0/2

LoCoMo

no data

MemoryBank

no data

Cross-Session Memory

0/1

LongMemEval

no data

Multi-Hop QA

1/3

BABILong

77.272p

MultiHop-RAG

no data

HotpotQA

no data

Agent Task Memory

0/1

AgentBench-Mem

no data

Personalization

0/1

PerLTQA

no data

Factuality / Grounding

0/1

RAGAS

no data

Sources:RWKV paper (arXiv:2305.13048); evaluated on BABILong: Testing the Limits of LLMs with Long-Context Reasoning-in-a-Haystack (AIRI, 2406)RWKV paper (arXiv:2305.13048); evaluated on InfiniteBench: Extending Long Context Evaluation Beyond 100K Tokens (Tsinghua / OpenBMB, 2402)RWKV paper (arXiv:2305.13048); evaluated on LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding (Tsinghua KEG, 2308)RWKV paper (arXiv:2305.13048); evaluated on LooGLE: Can Long-Context Language Models Understand Long Contexts? (Peking University, 2311)RWKV paper (arXiv:2305.13048); evaluated on Needle in a Haystack (Greg Kamradt, 2024)RWKV paper (arXiv:2305.13048); evaluated on RULER: What's the Real Context Size of Your Long-Context Language Models (NVIDIA, 2404)