2026-08-20-difficult-advice-gemini-716-smoke
收藏资源简介:
该数据集是“Difficult-advice SFT corpus”的变体,基于“Teaching Claude Why”配方构建,将整个生成器堆栈从Anthropic的基线(difficult_advice.yaml)替换为Gemini模型。数据集为合成生成,主要用于监督微调(SFT)中的对齐训练,特别是针对困难建议场景。数据生成过程分为多个阶段:首先对宪法(constitution)进行分块(stage_1),然后编写场景(stage_2),对场景进行去重(stage_3),草拟提示(stage_4),修订提示(stage_5),草拟响应(stage_6),修订响应(stage_7),最后导出SFT数据(stage_8)。每个阶段都有对应的JSONL数据文件。具体而言,使用google/gemini-3.6-flash生成场景、提示和草拟响应(阶段2/3/5),使用google/gemini-3.1-pro-preview根据完整宪法修订提示和响应(阶段4/6,对齐决策步骤)并对语料库进行评判。数据集的宪法文件为constitutions/claude_distilled_12_principles_mid/constitution.md。数据生成日期为20260820_163330。源代码仓库为https://github.com/Matthew-Bozoukov/teaching_claude_why_replication.git。所有模型调用通过OpenRouter路由,并设置safety_settings BLOCK_NONE。由于gemini-3.7-flash的安全层会阻止最难场景(CBRN/网络)的调用,因此生成运行在gemini-3.6-flash上。该数据集与基线语料库的行为差异可归因于生成器模型的替换。数据规模未明确给出,但包含多个阶段的JSONL文件,每个文件包含相应阶段的中间结果。适用任务包括监督微调、对齐训练、生成困难建议场景的提示-响应对。
This dataset is a variant of the Difficult-advice SFT corpus, built based on the Teaching Claude Why recipe, replacing the entire generator stack from Anthropics baseline (difficult_advice.yaml) with Gemini models. The dataset is synthetically generated and primarily used for alignment training in supervised fine-tuning (SFT), especially for difficult advice scenarios. The data generation process consists of multiple stages: first chunking the constitution (stage_1), then writing scenarios (stage_2), deduplicating scenarios (stage_3), drafting prompts (stage_4), revising prompts (stage_5), drafting responses (stage_6), revising responses (stage_7), and finally exporting SFT data (stage_8). Each stage has a corresponding JSONL data file. Specifically, google/gemini-3.6-flash is used to generate scenarios, prompts, and draft responses (stages 2/3/5), and google/gemini-3.1-pro-preview is used to revise prompts and responses according to the full constitution (stages 4/6, alignment decision steps) and to judge the corpus. The constitution file of the dataset is constitutions/claude_distilled_12_principles_mid/constitution.md. The data generation date is 20260820_163330. The source code repository is https://github.com/Matthew-Bozoukov/teaching_claude_why_replication.git. All model calls are routed through OpenRouter with safety_settings BLOCK_NONE. Since gemini-3.7-flashs safety layers block calls for the hardest scenarios (CBRN/cyber), generation runs on gemini-3.6-flash. Behavioral differences between this dataset and the baseline corpus can be attributed to the replacement of the generator model. The data scale is not explicitly given, but includes multiple JSONL files for each stage, each containing intermediate results of the corresponding stage. Applicable tasks include supervised fine-tuning, alignment training, and generating prompt-response pairs for difficult advice scenarios.
数据集概述
该数据集是一个困难建议(difficult-advice)监督微调(SFT)语料库的“全Gemini分支”版本,其核心特征是将Teaching Claude Why配方中的整个生成器堆栈从Anthropic模型(difficult_advice.yaml基线)替换为Gemini模型。
生成流程与模型分工
| 阶段 | 模型 | 任务 |
|---|---|---|
| 阶段 2/3/5 | google/gemini-3.6-flash | 生成场景、提示词和草稿回答 |
| 阶段 4/6 | google/gemini-3.1-pro-preview | 根据完整宪法重写提示词和回答(对齐决策步骤) |
| 阶段 7(评判) | google/gemini-3.1-pro-preview | 对语料库进行评判(autorate/scan) |
所有调用通过OpenRouter路由,并设置safety_settings BLOCK_NONE;采样设置和每阶段token使用情况记录在manifest.json中。
数据构成
数据集包含多个配置(config),对应流水线的不同阶段快照:
dataset:主数据集(dataset.jsonl)stage_1_chunk_constitution至stage_8_export_sft:各中间阶段的JSONL快照,包括:- 宪法分块(stage_1)
- 场景编写(stage_2)
- 场景去重(stage_3)
- 提示词草稿(stage_4)
- 提示词修订(stage_5)
- 回答草稿(stage_6)
- 回答修订(stage_7)
- SFT导出(stage_8)
- 多个
.partial阶段性中间文件
关键元数据
- 生成日期:20260820_163330
- 宪法:
constitutions/claude_distilled_12_principles_mid/constitution.md - 来源仓库:https://github.com/Matthew-Bozoukov/teaching_claude_why_replication.git(提交哈希:6eabb024100ccddef0aa1deb514e6787ed10e407)
- 配置:
configs/data/synth/difficult_advice_gemini_716.yaml - 标签:training-data、kind:synth、pipeline:difficult-advice、constitution:claude_distilled_12_principles_mid、smoke、backfilled-2026-08-25
特殊说明
生成运行使用gemini-3.6-flash而非更新的3.7-flash,原因是3.7的安全层即使在BLOCK_NONE设置下也会持续阻止约3%的draft_prompts调用,且这些被阻止的内容完全集中在最困难的t4(CBRN/网络)场景——即语料库最有价值的内容。已核实3.6/3.5/2.5-flash能通过3.7被阻止的相同提示词,确认该收紧是3.7特有的,而非共享的不可配置核心。
研究意义
与基线语料库的行为差异可归因于生成器模型的不同,该数据集可用于研究不同生成模型对最终SFT语料库质量与构成的影响。




