qwen3.5-9b-blogprovider-traces
收藏资源简介:
该数据集名为“Blog-Provider-ID — model inference traces (val + val_ood)”,包含由Hugging Face上CK0607 blog-provider集合中的多个模型在“3-way AI-provider classification task”任务上生成的推理轨迹。该分类任务的目标是判别给定博客内容是由CLAUDE、CHATGPT还是GEMINI这三种AI模型生成的。数据集包括验证集(val,n=414)和分布外验证集(val_ood,n=471)的推理记录。对于每个模型,数据以JSONL格式存储,每条记录对应一个博客,包含以下字段:提示与真实标签(prompt gold)、模型预测(prediction)、完整的推理与答案生成内容(full `<reason_why>/<answer>` completion)以及截断标志(truncation flag)。此外,每个模型文件夹还包含一个汇总文件(summary.json)。所有推理均在统一条件下生成,即使用“plain”版本的SYSTEM_PROMPT_3WAY(关闭模型的“思考”功能),以确保不同模型输出结果的可比性。数据集附带的analysis/目录提供了对AI提供者写作风格的研究,包括基于语义和文体测量学的UMAP可视化、线性探针可分离性分析以及各提供者的风格指纹。该数据集适用于AI生成文本检测、大型语言模型行为与可解释性分析、模型推理过程研究、以及AI写作风格对比分析等任务。README中提供的准确率表格展示了各模型在该任务上的性能,可作为基准参考。
The dataset is named Blog-Provider-ID — model inference traces (val + val_ood) and contains inference traces generated by multiple models from the CK0607 blog-provider collection on Hugging Face for the 3-way AI-provider classification task. This classification task aims to determine whether a given blog content was generated by one of three AI models: CLAUDE, CHATGPT, or GEMINI. The dataset includes inference records for the validation set (val, n=414) and the out-of-distribution validation set (val_ood, n=471). For each model, the data is stored in JSONL format, with each record corresponding to a blog and containing the following fields: prompt gold, prediction, full `<reason_why>/<answer>` completion, and a truncation flag. Additionally, each model folder includes a summary file (summary.json). All inferences are generated under uniform conditions, using the plain version of SYSTEM_PROMPT_3WAY (with the models thinking function disabled) to ensure comparability across different models. The dataset comes with an analysis/ directory that provides research on AI provider writing styles, including UMAP visualizations based on semantics and stylometry, linear probe separability analysis, and style fingerprints for each provider. This dataset is suitable for tasks such as AI-generated text detection, large language model behavior and interpretability analysis, model inference process research, and AI writing style comparison analysis. The README includes an accuracy table showing the performance of each model on this task, serving as a benchmark reference.
数据集概述
数据集名称:Blog-Provider-ID — model inference traces (val + val_ood)
许可证:Apache-2.0
任务类别:文本分类(text-classification)
标签:ai-text-detection, qwen3, prime-rl, reasoning-traces
数据集描述:该数据集包含针对三分类(CLAUDE / CHATGPT / GEMINI)AI生成文本提供商识别任务的模型推理轨迹。数据来自 CK0607 blog-provider collection 系列模型。
数据内容:每个模型文件夹中包含:
val.jsonl:验证集数据val_ood.jsonl:分布外验证集数据(每条记录包含:博客提示、真实标签、预测结果、完整的 <reason_why>/<answer> 补全内容、截断标志)summary.json:摘要文件
所有推理均使用 plain 版本的 SYSTEM_PROMPT_3WAY(关闭思考模式),因此结果具有直接可比性。
附加分析:analysis/ 文件夹包含提供商标记风格研究(语义 vs. 风格计量UMAP、线性探针可分性分析、每个提供商的风格指纹)。
模型性能(Plain-prompt准确率)
验证集(val)样本量 n=414,分布外验证集(val_ood)样本量 n=471。
| 模型 | val 准确率 | val_ood 准确率 |
|---|---|---|
coldstart-rl |
0.9106 | 0.9193 |
coldstart-sft |
0.6957 | 0.6815 |
qwen3.5-9b-blogprovider-rl-cheatsheet |
0.372 | 0.3737 |
qwen3.5-9b-blogprovider-rl-entropydecay |
0.2899 | 0.3036 |
qwen3.5-9b-blogprovider-rl-pureacc-peak |
0.3841 | 0.3822 |
qwen3.5-9b-blogprovider-opcd-e2 |
0.0314 | 0.0234 |
qwen3.5-9b-blogprovider-rlsd-e3 |
0.0024 | 0.0 |
qwen3.5-9b-blogprovider-selfgated-answeronly |
1.0 | 1.0 |
qwen3.5-9b-blogprovider-sft-goldcond |
1.0 | 1.0 |
qwen3.5-9b-blogprovider-star-selfdistill |
0.9517 | 0.9597 |
特别说明:
opcd-e2和rlsd-e3是OPCD / RLSD重新运行的 最终(step-40)检查点,它们最初在早期达到过渡推理通道峰值(约0.27 / 约0.35 val reward),随后出现截断崩溃,因此最终检查点的plain-prompt准确率接近0。完整的运行轨迹可参见博客/日志中的运行阶梯。




