quant-experiments
收藏资源简介:
该数据集是用于在Prism能力基板上进行模型量化实验的配套资源,专注于通过压缩权重比特数(如NF4、int8、int4量化)来减少模型大小,同时保持下游任务性能不下降。数据集与prism-capability-extraction和circuit-discovery项目相关联,采用模块化结构,每个实验以独立文件夹形式组织,包含完整的实验记录:README.md文件描述实验目的、方法和运行命令;MANIFEST.json提供机器可读的元数据(包括基板、堆栈、评估和结果);code/目录包含量化框架和工具脚本;substrate_meta/存储适配器、分支和掩码的元数据收据;configs/存放实验配置文档;reports/包含每次运行的标准化评估摘要;logs/记录实验运行日志。数据集示例包括针对Qwen3-8B模型在特定基板上的注意力优先量化实验。需要注意的是,大尺寸的权重和掩码数据存储在独立的circuit-discovery数据集中,而本数据集主要包含实验代码、配置、元数据和结果报告。
This dataset is a companion resource for model quantization experiments on the Prism capability substrate, focusing on reducing model size by compressing weight bit numbers (e.g., NF4, int8, int4 quantization) while maintaining downstream task performance without degradation. It is associated with the prism-capability-extraction and circuit-discovery projects, and adopts a modular structure where each experiment is organized in an independent folder containing complete experimental records: a README.md file describing the experiments purpose, methods, and execution commands; MANIFEST.json providing machine-readable metadata (including substrate, stack, evaluation, and results); a code/ directory containing quantization frameworks and tool scripts; substrate_meta/ storing metadata receipts for adapters, branches, and masks; configs/ holding experiment configuration documents; reports/ including standardized evaluation summaries for each run; and logs/ recording experimental run logs. Dataset examples include attention-prioritized quantization experiments for the Qwen3-8B model on specific substrates. It is important to note that large-scale weight and mask data are stored in the separate circuit-discovery dataset, while this dataset primarily contains experimental code, configurations, metadata, and result reports.
数据集概述
数据集名称:quant-experiments
数据集描述: 该数据集是量化实验的集合,旨在对Prism能力子结构进行量化研究——在提取的任务子结构上压缩每个权重的位数,同时不降低下游得分。该数据集与以下仓库相关联:
数据布局:
每个实验是一个自包含的文件夹,命名格式为 issue<N>_<slug>_v<k>/。每个实验文件夹包含以下内容:
README.md:实验的动机、方法以及运行命令。MANIFEST.json:机器可读格式的子结构、堆栈、评估和结果信息。code/:量化脚本及附带的测试工具代码。substrate_meta/:适配器、分支、掩码的接收信息(小型JSON文件)。configs/:实验配置或文档。reports/:每次运行的评估摘要(标准化准确率)。logs/:运行日志。
实验列表:
| 文件夹名称 | 关联问题 | 子结构 | 关注点 |
|---|---|---|---|
issue5_attention_first_quant_v1 |
#5 | Qwen3-8B + b007 + issue12 v13 MACE-90 | 注意力优先量化(NF4 / int8 / int4),MLP在后 |
重要说明:
- 大型权重和掩码文件存储在
TokenBender/circuit-discovery数据集中,此数据集仅包含代码、接收信息、评估结果和日志,不含基础权重文件。




