axis-training-sets
收藏资源简介:
Axis Training Sets 数据集是一个专为Google Gemma-3大语言模型设计的微调与对齐实验训练集。它包含两个核心文件:1. `targeted_sft_v2.jsonl`:这是一个Targeted SFT v2训练集,采用JSON Lines格式,由265道经过验证的教师解答和265道回放题目组合而成,总计530行数据,对应265个训练目标。该文件通过脚本从已评分的教师解题数据中,依据良率与能力墙标准筛选生成。2. `targeted_sft_v2_manifest.json`:这是训练集构建过程的详细元数据清单文件,记录了数据过滤条件、难度分布、时间戳以及验证信息等,在训练集生成时自动创建。该数据集适用于大型语言模型的监督微调(SFT)与对齐任务,旨在提升模型在特定领域或任务上的性能与安全性。
The Axis Training Sets dataset is a training collection designed for fine-tuning and alignment experiments with the Google Gemma-3 large language model. It includes two core files: 1. `targeted_sft_v2.jsonl`: This is a Targeted SFT v2 training set in JSON Lines format, consisting of 265 verified teacher correct solutions and 265 replay pool items, totaling 530 rows of data corresponding to 265 training objectives. The file is generated via a script that filters from scored teacher solution data based on yield and capability wall criteria. 2. `targeted_sft_v2_manifest.json`: This is a detailed metadata manifest file documenting the dataset construction process, including data filtering conditions, difficulty distribution, timestamps, and validation information, automatically created during training set generation. The dataset is tailored for supervised fine-tuning (SFT) and alignment tasks of large language models, aiming to enhance model performance and safety in specific domains or tasks.
数据集概述:Axis Training Sets
该数据集收录了用于 Google Gemma-3 模型微调与对齐实验中的训练集及相关 Manifest 文件。
文件列表与说明
-
targeted_sft_v2.jsonl- 内容:Targeted SFT v2 训练集,包含 265 题老师验证的解答和 265 题 replay 题目,共 530 行数据,对应 265 个标的。
- 生成方式:由脚本
程式與腳本/build_targeted_sft_v2.py基于teacher_wall_pro_scored.jsonl中的筛选条件(良率与 0/8 能力墙)组合而成。 - 对应字段:
AI_AGENT_WORKFLOW.md§1 账本第 47 列(标记为“targeted v2(265 验证解 + 265 replay)”)。
-
targeted_sft_v2_manifest.json- 内容:Targeted SFT v2 训练集的详细 metadata 与 Manifest,包含过滤条件、难度分布、时间戳及验证信息。
- 生成方式:由脚本
程式與腳本/build_targeted_sft_v2.py在构建训练集时自动同步生成并写入。 - 对应字段:
AI_AGENT_WORKFLOW.md§1 账本第 47 列。




