vorn-mat-cross-family-results
收藏资源简介:
Vorn-Mat跨家族结果数据集是论文《Vorn-Mat: 家族条件KV缓存驱逐与粒度拯救谱》的配套发布,旨在支持论文中配对McNemar测试的可复现性。该数据集包含实验活动中产生的所有49个已发布结果工件、作为参考基线的原始逐夹具观察数据分片以及论文附录A中引用的图表。具体内容包括:JSON格式的结果工件(每个都配有Markdown摘要)、包含原始观察数据的10个gzipped JSON分片以及4个参考图表。数据规模包括49个结果工件、299个计数结果行、270个携带逐夹具观察结果的行、18,000个独立的逐夹具观察记录,实验覆盖9个模型家族(如Mistral 7B v0.3、Llama 3.1 8B、Qwen 3 30B-A3B等),总计消耗32.72 GPU小时。数据集主要基于RULER基准测试,使用`niah_multikey_1_4k`切片进行实验,模型以bf16精度运行,采用贪婪解码。该数据集适用于KV缓存管理、注意力机制优化、长上下文语言模型基准测试结果分析等研究任务,为相关领域的实验复现和结果验证提供了关键数据支持。
The Vorn-Mat Cross-Family Results Dataset is an accompanying release of the paper *Vorn-Mat: Family-Conditional KV Cache Eviction and Granularity Rescue Spectrum*, intended to support the reproducibility of the paired McNemar test described in the paper. This dataset encompasses all 49 published result artifacts generated throughout the experimental campaign, the original per-fixture observational data shards serving as reference baselines, and the figures cited in Appendix A of the paper. Specifically, it includes: result artifacts in JSON format (each paired with a Markdown summary), 10 gzipped JSON shards containing raw observational data, and 4 reference figures. In terms of scale, the dataset contains 49 result artifacts, 299 count result rows, 270 rows with per-fixture observational results, and 18,000 independent per-fixture observation records. The experiments cover 9 model families (e.g., Mistral 7B v0.3, Llama 3.1 8B, Qwen 3 30B-A3B, etc.), with a total GPU hour consumption of 32.72. The dataset is primarily based on the RULER benchmark, with experiments conducted using the `niah_multikey_1_4k` slice. All models are executed with bfloat16 precision and adopt greedy decoding. This dataset is applicable to research tasks including KV cache management, attention mechanism optimization, and benchmark result analysis for long-context language models, providing critical data support for experimental reproduction and result validation in relevant research fields.
Vorn-Mat: Cross-Family Results 数据集详情
数据集概述
该数据集是论文《Vorn-Mat: Family-Conditional KV-Cache Eviction and the Granularity Rescue Spectrum》的配套发布物,包含了实验活动的所有49个结果工件以及支持的原始观测数据碎片和图表。
数据规模
- 样本量:10K < n < 100K
- 语言:英语
- 许可证:MIT
数据集内容
| 路径 | 内容 |
|---|---|
*.json |
49个实验活动结果工件,每个配有Markdown摘要文件(*.md) |
vanilla-observation-2026-05-13-shards/ |
10个gzip压缩的JSON碎片,包含原始逐夹具vanilla(无驱逐)观测数据,作为参考基线 |
figures/ |
4个论文中引用的图表(对齐差距、排名稳定性、残差差距、按步骤的答案top-k命中率) |
数据结构
大多数结果工件是JSON封装,包含rows[]或专门的数组如sentence_rows[]和token_rows[]。遗留和探针工件使用较旧的结构,用于溯源。大多数承载主要结论的行包含逐夹具的observations[]数组,记录每个评估夹具的二元正确性结果和原始预测文本。
主要统计数字(可从本数据集复现)
- 49个结果工件
- 299个计数的结果行
- 270行携带逐夹具观测数据
- 18,000条独立逐夹具观测记录
- 32.72 GPU小时(跨9个模型家族)
- 总计估计计算成本:$77.78
方法论
- 基准测试:使用RULER基准(通过HuggingFace镜像
rbiswasfc/ruler),主要基于niah_multikey_1_4k切片(validation[:50]) - 测试的模型家族(9个):Mistral 7B v0.3、Llama 3.1 8B、Ministral 8B、Gemma 2 9B、Gemma 4 E4B-it、Qwen 3 8B、Gemma 3 12B-pt、Qwen 2.5 7B Instruct、Qwen 3 30B-A3B
- 模型使用:直接从HuggingFace获取,bf16精度,贪婪解码
- 实验硬件:主要使用NVIDIA A100 80GB GPU(通过Modal),Phase 2C Mistral后期预算重试使用H100 80GB
引用
bibtex @article{penney2026vornmat, title={Vorn-Mat: Family-Conditional KV-Cache Eviction and the Granularity Rescue Spectrum}, author={Penney, L.}, journal={arXiv preprint}, year={2026} }
相关资源
- 论文配套代码仓库:https://github.com/synapt-dev/vorn-mat
- LOCOMO基准:https://snap-research.github.io/locomo/
- RULER基准:https://huggingface.co/datasets/rbiswasfc/ruler





