wrong-reasoning-traces
收藏资源简介:
Wrong-Reasoning Trace Bank (Completeness Cliff) 是一个用于研究语言模型推理过程的数据集,专注于分析完整性悬崖现象,即语言模型在错误推理完成前仍能纠正并得出正确答案的能力。它包含模型生成的推理轨迹,作为研究语言模型在错误推理完成前逃脱错误推理的注入材料。数据以JSONL格式存储,每个样本代表一次基线推演,主要包括两种类型:correct == false的行作为自我错误注入源(模型自身的错误推理),correct == true的行作为损坏源(正确框架但数字被扰动),并用于其他问题的无关/跨领域供体。数据字段涵盖运行标识、问题ID、数据集来源、任务名称、正确答案、样本索引、模型答案、正确性标志、截断标志、生成令牌数和推理轨迹文本。数据集使用了多种模型(如Qwen3-4B-Thinking-2507、Qwen3-4B-Instruct-2507和Gemma-4-E2B-it)在多个问题集(包括AIME 24+25、推理迷宫和迷你数独)上生成,每个问题采样16个样本。数据集不包含问题原文,仅通过问题ID引用AIME问题,但包含正确答案。推理轨迹受原始模型许可条款约束,而数据集卡片和结构以CC-BY-4.0许可发布。
The Wrong-Reasoning Trace Bank (Completeness Cliff) is a dataset designed for studying the reasoning processes of language models, with a specific focus on the completeness cliff phenomenon—the ability of language models to correct themselves and arrive at correct answers before completing erroneous reasoning. It contains model-generated reasoning traces that serve as injection materials for researching how language models escape erroneous reasoning before its completion. The data is stored in JSONL format, with each sample representing a baseline inference run. It primarily includes two types of rows: those with correct == false are self-error injection sources (the models own erroneous reasoning), while those with correct == true are corrupted sources (correct frameworks with numbers perturbed near cutoff points) and act as irrelevant/cross-domain donors for other problems. Data fields encompass run identifiers, question IDs, dataset sources, task names, correct answers, sample indices, model answers, correctness flags, truncation flags, generated token counts, and reasoning trace texts. The dataset was generated using multiple models (such as Qwen3-4B-Thinking-2507, Qwen3-4B-Instruct-2507, and Gemma-4-E2B-it) on various problem sets (including AIME 24+25, Reasoning Maze, and Mini Sudoku), with 16 samples per problem. It does not include the original problem texts but references AIME problems via question IDs and includes correct answers. The reasoning traces are subject to the original model licensing terms, while the dataset card and structure are released under the CC-BY-4.0 license.
数据集概述:Wrong-Reasoning Trace Bank (Completeness Cliff)
基本信息
- 数据集名称: Wrong-Reasoning Trace Bank (Completeness Cliff)
- 许可证: CC-BY-4.0
- 语言: 英语
- 任务类型: 文本生成
- 数据集规模: 1K < n < 10K 条样本
- 数据集配置: 默认配置,训练集数据存储在
data/*.jsonl文件中
数据集用途
该数据集包含模型生成的推理轨迹,用于研究 “完整性悬崖”(Completeness Cliff) 现象——即语言模型在推理完成之前无法摆脱错误推理。每条数据是一次采样的基线 rollout,其推理内容可以被拼接回新的提示中,测试模型是否仍能恢复正确答案(pass@k)。
- 代码与论文: GitHub 仓库
- 博客文章: 即将发布在 https://alvinzh04.github.io/blog/completeness-cliff.html
数据结构
每条 JSONL 行包含以下字段:
| 字段 | 类型 | 含义 |
|---|---|---|
run |
string | 基线运行的名称 |
qid |
string | 问题ID(如 aime25-3) |
dataset |
string | 数据集名称:aime24_25、rg_maze、rg_mini_sudoku |
rg_task |
string/null | reasoning-gym 任务名称(如适用) |
gold |
string | 正确答案 |
sample_index |
int | N=16 次采样中的索引 |
answer |
string | 模型对该样本提取的答案 |
correct |
bool | 答案是否匹配 gold |
truncated |
bool | 是否达到生成长度上限 |
n_tokens |
int | 生成的 token 数量 |
trace_text |
string | 推理通道内容(推理模型的 <think> 内容或指令模型的完整响应) |
数据用途说明
correct == false的行:作为 self_wrong 注入源(模型自身的错误推理)correct == true的行:作为 corrupted 源(带有扰动的正确推理框架),用作其他问题的 irrelevant / cross-domain 捐赠者
覆盖的模型与问题
| 模型 | 类型 | AIME 24+25 | reasoning-gym |
|---|---|---|---|
| Qwen3-4B-Thinking-2507 | 推理模型 | 是 | 迷宫、迷你数独 |
| Qwen3-4B-Instruct-2507 | 指令模型(匹配版本) | 是(含32k预算重运行) | 迷宫 |
| Gemma-4-E2B-it | 混合模型(跨家族) | 是 | 迷宫 |
- 每个问题采样 N=16 次,采用官方推荐的采样方案,使用 vLLM 框架
- 每个运行的基线指标(pass@k、adoption、truncation、token 数量)存储在
baseline_summaries.json中
数据加载示例
python from datasets import load_dataset ds = load_dataset("AZH04/wrong-reasoning-traces", split="train") wrong = ds.filter(lambda r: not r["correct"]) # 获取 self_wrong 注入轨迹
或直接加载单个运行文件: python import json rows = [json.loads(l) for l in open("data/base_qwen3-4b-thinking_aime24_25.jsonl")]
注意事项
- 仅保留提炼后的字段,完整的原始 rollout(含后推理答案文本和 token 级元数据,约2.4GB)未发布
- 问题文本未重新分发,仅通过
qid引用 AIME 问题,并包含标准答案 - AIME 问题归美国数学协会所有
- 推理轨迹来自 Qwen3(Apache-2.0)和 Gemma(Gemma 使用条款)模型
- 数据集卡片和结构采用 CC-BY-4.0 发布;模型输出受源模型条款约束




