.zenodo.json

{
  "title": "ElasticKV: Exploiting Attention Haze for Structure-Preserving KV-Cache Compression",
  "description": "ElasticKV is a training-free KV-cache sparsification method for llama.cpp that removes low-magnitude, low-selectivity attention haze via per-head adaptive thresholds with sink token protection. This archival release bundles the CUDA hook, paper source, raw artifacts, and analysis scripts used in the accompanying manuscript.",
  "creators": [
    {
      "name": "Camargo, Guilherme de",
      "orcid": "0009-0004-8913-9419",
      "affiliation": "Independent Researcher",
      "type": "ProjectMember"
    }
  ],
  "keywords": [
    "kv-cache",
    "llama.cpp",
    "llm",
    "cuda",
    "long-context",
    "sparsity",
    "attention",
    "reproducibility"
  ],
  "license": "apache-2.0",
  "upload_type": "software",
  "access_right": "open",
  "language": "eng",
  "version": "0.1.1"
}