datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
daily-paper-2026-07-28-kv-cache-tiering-pd-disagg-cost
KV Cache Tiering Meets Prefill-Decode Disaggregation: Mapping the Cost-Latency Frontier for MoE LLM Serving on H200
TL;DR — Combining prefill/decode disaggregation with tiered KV-cache offloading is analytically antagonistic: disaggregation raises cache hit value but consumes the TTFT slack the slowest tier needs. Empirical validation failed (vLLM init crash), yielding zero performance data.
ThakiCloud AI Research · 2026-07-28 · 📝 Tech blog (KO)
Problem
LLM… See the full description on the dataset page: https://huggingface.co/datasets/thaki-AI/daily-paper-2026-07-28-kv-cache-tiering-pd-disagg-cost.Neurips2026_evaluating_accuracy_KV-cache_reuse_techniques
BoxOffice Verified Seeds
This dataset contains the released BoxOffice seed datasets used in the
benchmark pipeline described in the accompanying paper. The release includes
ten verified seeds:
7
11
13
17
19
23
29
31
47
73
For each seed, we provide:
a full JSONL file containing warmup rows plus evaluation rows
an eval JSONL file containing only the evaluation rows
a manifest JSON file
a validation JSON file with directional warmup counts
Layout
viewer/
normalized… See the full description on the dataset page: https://huggingface.co/datasets/Boxoffice1280/Neurips2026_evaluating_accuracy_KV-cache_reuse_techniques.details_xformAI__opt-125m-gqa-ub-6-best-for-KV-cache
Dataset Card for Evaluation run of xformAI/opt-125m-gqa-ub-6-best-for-KV-cache
Dataset automatically created during the evaluation run of model xformAI/opt-125m-gqa-ub-6-best-for-KV-cache on the Open LLM Leaderboard.
The dataset is composed of 63 configuration, each one coresponding to one of the evaluated task.
The dataset has been created from 2 run(s). Each run can be found as a specific split in each configuration, the split being named using the timestamp of the run.The… See the full description on the dataset page: https://huggingface.co/datasets/open-llm-leaderboard-old/details_xformAI__opt-125m-gqa-ub-6-best-for-KV-cache.kvcache-quantization-logs-qwen7bdetails_xformAI__facebook-opt-125m-qcqa-ub-6-best-for-KV-cache
Dataset Card for Evaluation run of xformAI/facebook-opt-125m-qcqa-ub-6-best-for-KV-cache
Dataset automatically created during the evaluation run of model xformAI/facebook-opt-125m-qcqa-ub-6-best-for-KV-cache on the Open LLM Leaderboard.
The dataset is composed of 63 configuration, each one coresponding to one of the evaluated task.
The dataset has been created from 1 run(s). Each run can be found as a specific split in each configuration, the split being named using the timestamp… See the full description on the dataset page: https://huggingface.co/datasets/open-llm-leaderboard-old/details_xformAI__facebook-opt-125m-qcqa-ub-6-best-for-KV-cache.turboquant-tcq-kv-cache
Closing the Gap: Trellis-Coded Quantization for KV Cache at 2-3 Bits
Authors: buun, Claude (Anthropic)
This dataset accompanies the paper Closing the Gap: Trellis-Coded Quantization for KV Cache at 2-3 Bits. It contains the trained TCQ codebooks, codebook training scripts, and the full paper PDF.
Summary
We present the first application of trellis-coded quantization (TCQ) to KV cache compression in LLM inference. TCQ constrains quantization indices to follow a… See the full description on the dataset page: https://huggingface.co/datasets/spiritbuun/turboquant-tcq-kv-cache.details_saarvajanik__facebook-opt-6.7b-gqa-ub-16-best-for-KV-cache
Dataset Card for Evaluation run of saarvajanik/facebook-opt-6.7b-gqa-ub-16-best-for-KV-cache
Dataset automatically created during the evaluation run of model saarvajanik/facebook-opt-6.7b-gqa-ub-16-best-for-KV-cache on the Open LLM Leaderboard.
The dataset is composed of 63 configuration, each one coresponding to one of the evaluated task.
The dataset has been created from 1 run(s). Each run can be found as a specific split in each configuration, the split being named using the… See the full description on the dataset page: https://huggingface.co/datasets/open-llm-leaderboard-old/details_saarvajanik__facebook-opt-6.7b-gqa-ub-16-best-for-KV-cache.details_saarvajanik__facebook-opt-6.7b-qcqa-ub-16-best-for-KV-cache
Dataset Card for Evaluation run of saarvajanik/facebook-opt-6.7b-qcqa-ub-16-best-for-KV-cache
Dataset automatically created during the evaluation run of model saarvajanik/facebook-opt-6.7b-qcqa-ub-16-best-for-KV-cache on the Open LLM Leaderboard.
The dataset is composed of 63 configuration, each one coresponding to one of the evaluated task.
The dataset has been created from 1 run(s). Each run can be found as a specific split in each configuration, the split being named using the… See the full description on the dataset page: https://huggingface.co/datasets/open-llm-leaderboard-old/details_saarvajanik__facebook-opt-6.7b-qcqa-ub-16-best-for-KV-cache.stillwarm-kv-cache-artifact
A downloadable KV-cache save file — with the honest math
One llama-server slot save: the first 8,192 Llama-tokens of Frankenstein
(public domain), prefilled by Qwen2.5-7B-Instruct Q4_K_M (Apache-2.0 model —
chosen over Llama specifically for artifact licensing) and saved with a
stillwarm sidecar.
This file is USELESS unless your setup matches the sidecar exactly:
field
value
llama.cpp build
b9871 (ef2d770117db45b05aa7ecd1b0acca36370c5470) — advisory: ±5 weeks measured… See the full description on the dataset page: https://huggingface.co/datasets/vimalnakrani/stillwarm-kv-cache-artifact.dgx-spark-kv-cache-benchmark
KV Cache Quantization on NVIDIA DGX Spark GB10
Corrected benchmarks (v3, April 2026) — KV cache quantization behavior on the NVIDIA DGX Spark's GB10 Grace Blackwell unified memory architecture.
Author: Nathan Maine
Date: March 2026, corrected April 2026
Hardware: NVIDIA DGX Spark (GB10, compute 12.1, 128GB unified memory)
Correction Notice: The original v1 benchmarks (March 31) contained methodology errors. Memory was measured via RSS (wrong on unified memory) and some… See the full description on the dataset page: https://huggingface.co/datasets/Nathan-Maine/dgx-spark-kv-cache-benchmark.KVCacheskvcache-bench-results
Mingxin KV-Cache Tiered-Storage Benchmark Results
Measured results for LLM KV Cache tiered-storage acceleration on 8× AMD Instinct MI308X (ROCm 7.2, vLLM 0.20.1+rocm721, LMCache upstream mainline), model Qwen3-Coder-480B-FP8, published by Mingxin Technology.
Device under test: Mingxin FX100 all-flash NVMe-oF array (4-disk RAID0, RoCEv2, 100 GbE) vs local NVMe (PCIe Gen4) vs recompute-only baseline.
Files
kvcache_bench_results.json — all experiments in structured… See the full description on the dataset page: https://huggingface.co/datasets/wangqiyuan2026/kvcache-bench-results.kv_cache_logits_exp2camus-10-kv-cache-quantization
We attempted KV cache quantization to Q4 — and documented why it fails on the qwen2vl architecture.
KV Cache Quantization Attempt: type_k/type_v on qwen2vl Architecture
The Problem
The KV cache consumes significant memory bandwidth during autoregressive generation. On CPU-bound systems, memory bandwidth is the primary bottleneck. Quantizing the KV cache from FP16 to Q4 theoretically halves memory bandwidth requirements.
What We Built
We attempted… See the full description on the dataset page: https://huggingface.co/datasets/Anticloud/camus-10-kv-cache-quantization.kvcachedgx-spark-kv-cache-benchmark
KV Cache Quantization on NVIDIA DGX Spark GB10
Corrected benchmarks (v3, April 2026) — KV cache quantization behavior on the NVIDIA DGX Spark's GB10 Grace Blackwell unified memory architecture.
Author: Nathan Maine, Memoriant Inc.
Date: March 2026, corrected April 2026
Hardware: NVIDIA DGX Spark (GB10, compute 12.1, 128GB unified memory)
Correction Notice: The original v1 benchmarks (March 31) contained methodology errors. Memory was measured via RSS (wrong on unified memory) and… See the full description on the dataset page: https://huggingface.co/datasets/memoriant/dgx-spark-kv-cache-benchmark.camus-10-kv-cache-quantization
We attempted KV cache quantization to Q4 — and documented why it fails on the qwen2vl architecture.
KV Cache Quantization Attempt: type_k/type_v on qwen2vl Architecture
The Problem
The KV cache consumes significant memory bandwidth during autoregressive generation. On CPU-bound systems, memory bandwidth is the primary bottleneck. Quantizing the KV cache from FP16 to Q4 theoretically halves memory bandwidth requirements.
What We Built
We attempted… See the full description on the dataset page: https://huggingface.co/datasets/kleinnner/camus-10-kv-cache-quantization.kv-cache-compression-mbe
Matched-Budget Evaluation (MBE) — KV Cache Compression
A standardized reporting protocol for KV cache compression in LLM inference. MBE is
not a new task benchmark; it is a thin reporting layer that fixes which models, tasks,
and budgets results are reported at, so that numbers from different papers become
comparable.
Manifest (mbe_manifest.json): the frozen evaluation specification — model suite,
task suite (consuming existing benchmarks: LongBench, RULER, SCBench, GSM8K… See the full description on the dataset page: https://huggingface.co/datasets/Rohithreddybc/kv-cache-compression-mbe.fast-kv-compaction-cachekvcache_offload