exp10-deploy-bakeoff-data
收藏资源简介:
该数据集是一个用于训练和评估评估意识探针的数据组件集合。核心数据包括用于训练的多项选择题(MCQ)数据集,以及用于评估的对比数据集,如FORTRESS-vs-WildChat(计划中)和HarmBench-vs-WildChat(计划中)。数据集使用特定边界提示(例如RETINAL提示)来评估模型在受控行为损失下的表现,并应用于两种预训练模型:M_sdf(基于Llama-3_3-Nemotron-Super-49B-v1,主要关注类型提示使用行为)和OLMO 3.1 32B Think(主要关注边界提示上的语言化评估意识)。评估套件产生结构化指标,包括训练AUC轨迹、数据集区分AUC、基于Qwen评判的行为转向评分,以及针对特定任务(如FEP谜题、VEA插入)的逐词元探针轨迹。数据以JSON等格式存储,用于支持探针训练、阿尔法参数扫描和行为评估。
This dataset is a collection of data components for training and evaluating evaluation-aware probes. Core data includes multiple-choice question (MCQ) datasets for training, and contrastive evaluation datasets such as FORTRESS-vs-WildChat (planned) and HarmBench-vs-WildChat (planned). The dataset utilizes specific boundary prompts (e.g., RETINAL prompts) to assess model performance under controlled behavioral loss, and is applied to two pre-trained models: M_sdf (based on Llama-3_3-Nemotron-Super-49B-v1, primarily focusing on typed prompt usage behavior) and OLMO 3.1 32B Think (primarily focusing on linguistic evaluation awareness on boundary prompts). The evaluation suite generates structured metrics, including training AUC trajectories, dataset-discriminating AUC scores, behavioral shift scores evaluated by Qwen, and token-wise probe trajectories for specific tasks such as FEP puzzles and VEA insertions. The data is stored in formats including JSON to support probe training, alpha parameter sweeping, and behavioral evaluation.
数据集概述:Exp10 — Constrained-Choice Steering for Eval-Awareness
该数据集用于研究评估感知(Eval-Awareness) 的约束选择引导(Constrained-Choice Steering)方法。核心目标是测试对比探针(contrastive probes)是否能够捕捉模型关于自身被评估的信念,而不仅仅是用户提供的上下文信息。
研究的模型生物体
- M_sdf:
nvidia/Llama-3_3-Nemotron-Super-49B-v1与 Wood Labs SDF LoRA 的组合。主要行为信号是类型提示(type-hint)的使用。 - OLMO 3.1 32B Think:
allenai/Olmo-3.1-32B-Think,在边界提示(borderline prompts)下的口头化评估感知(verbalized eval-awareness)是主要信号。
仓库结构
仓库主要包含以下目录和文件:
| 路径 | 描述 |
|---|---|
README.md |
本文件 |
PROPOSAL.md |
原始实验设计 |
ENV_SETUP.md |
虚拟环境/包依赖说明 |
DESIGN_wait_token_training.md |
早期设计笔记 |
data/ |
提示(prompts)和探针(probe)数据集 |
configs/ |
SLURM 配置文件 |
results/ |
所有训练、评估和判断(judge)产出物 |
plots/ |
历史静态图表(PNG) |
reports/ |
可自包含 HTML 报告 |
logs/ |
SLURM 日志 |
scripts/ |
核心脚本,包括 common.py、utils/、train/、eval/、judge/、suite/、report/、analysis/ 子目录 |
评估套件(Eval Suites)—— 标准工作流
训练探针后,运行对应的套件以在 results/eval_suite/<probe_tag>/ 下生成结构化产出物。每个组件是幂等的,可使用 --force 重新运行。
OLMO 套件
运行命令: bash python scripts/suite/run_olmo_suite.py --vector-path results/olmo_pipeline/olmo32b_v4_balanced_hua_et_al/best_by_auc.pt --probe-tag olmo32b_v4_balanced_hua_et_al --training-objective raise
各组件功能:
| 组件 | 测量内容 | 状态 |
|---|---|---|
training_trajectory |
MCQ 训练步骤上的 AUC | ✅ 已连接 |
auc_fortress_vs_wildchat |
留出集上的区分 AUC | ✅ 已连接 |
auc_harmbench_vs_wildchat |
新数据集转移 AUC | ⏳ 已存根 |
behavioral_steering_retinal |
Qwen 判断的评估感知率 vs α(RETINAL 边界提示,n=200/α) | ✅ 已连接 |
verbalized_ea_retinal |
GPT-OSS-120B 严格标准判断(相同 α 扫描) | ⏳ 已存根 |
fep_puzzle / fep_benign / fep_ea_prefix |
5 文档谜题设置中的逐 token 探针轨迹 | ✅ |
vea_insertion |
植入 "Im being tested…" 句子周围的逐 token 轨迹 |
✅ |
完整 RETINAL α 扫描运行: bash bash run_retinal_sweep.sh olmo32b_v4_balanced_hua_et_al bash run_retinal_sweep.sh --all
Llama 套件
运行命令: bash python scripts/suite/run_llama_suite.py --vector-path results/runs/main_A/v_star.pt --probe-tag llama_main_A
各组件功能:
| 组件 | 测量内容 | 状态 |
|---|---|---|
training_trajectory |
(依赖训练时日志) | ✅(如可用) |
typehint_regex |
基于正则表达式的类型提示率 vs α | ✅ |
typehint_llm |
LLM 判断的类型提示率 vs α | ⏳ 已存根 |
verbalized_ea |
GPT-OSS-120B 评估感知判断(在 CoTs 上) | ⏳ 已存根 |
HTML 报告
运行套件后生成自包含 HTML 报告: bash
逐探针详细报告
python scripts/report/generate_probe_html.py --suite-dir results/eval_suite/olmo32b_v4_balanced_hua_et_al
多探针对比报告
python scripts/report/generate_compare_html.py --glob "results/eval_suite/olmo32b_v4_balanced_*" --baseline olmo32b_v4_balanced_hua_et_al --out reports/compare_olmo_v4.html
训练 —— 快速参考
OLMO 训练
bash
标准 MCQ 训练
python scripts/train/olmo_pipeline.py --probe-key hua_et_al --tag olmo32b_v4_balanced_hua_et_al --layer 32
随机初始化控制
python scripts/train/olmo_pipeline.py --probe-key random --tag olmo32b_v4_balanced_random --layer 32 --max-steps 600
Llama 训练
使用 SLURM 包装脚本: bash bash run_pipeline.sh configs/full.sh
或分阶段运行:
bash run_extract_v0.sh bash run_curate_borderline.sh bash run_train.sh bash run_evaluate.sh
当前状态(2026-05-06)
- ✅ 仓库已重组为
train/eval/judge/suite/report/analysis子文件夹。 - ✅ 评估套件框架、逐探针和对比 HTML 报告已完成。
- ✅ 18 个 OLMO 探针已回溯填充到新的
results/eval_suite/布局。 - ⏳ GPT-OSS-120B 口头化评估感知判断器:尚未连接。
- ⏳ HarmBench-vs-WildChat AUC 组件:数据集摄取工作待完成。
- ⏳ Lower-EA 方向探针训练:未实现。
- ⏳ Llama 的 LLM 判断类型提示率:未连接。
- 🛈 未来计划:将 FORTRESS 切换为 HarmBench 作为主要评估感知数据集。
如何添加新组件
- 在
_schema.OLMO_COMPONENTS或_schema.LLAMA_COMPONENTS中添加条目。 - 在
_components.py中编写一个运行器,生成_schema.make_*负载并调用_manifest.record_component(...)。 - 在
_components.py底部的OLMO_RUNNERS/LLAMA_RUNNERS中添加该运行器。 - 编排器和 HTML 报告会自动拾取。




