Dataset Viewer
Auto-converted to Parquet Duplicate
task_id
string
sample_id
string
source_step
int64
action_type
string
messages
list
teacher_model
string
benchmark_leaked
bool
finish_reason
string
response_complete
bool
has_reasoning
bool
source_exchange_id
string
aerial-cactus-identification
aerial-cactus-identification__sample_00001
1
code
[ { "role": "system", "content": "You are a redacted_person attending a high-stakes competition. In order to win this competition, you need to come up with an excellent and creative plan for a solution and then implement this solution in Python.\n", "loss": false }, { "role": "user", "content"...
deepseek-v4.1-flash
true
stop
true
true
event_00001
aerial-cactus-identification
aerial-cactus-identification__sample_00002
2
code
[ { "role": "system", "content": "You are a redacted_person attending a high-stakes competition. In order to win this competition, you need to come up with an excellent and creative plan for a solution and then implement this solution in Python.\n", "loss": false }, { "role": "user", "content"...
deepseek-v4.1-flash
true
stop
true
true
event_00002
aerial-cactus-identification
aerial-cactus-identification__sample_00003
3
code
[{"role":"system","content":"You are a redacted_person attending a high-stakes competition. In order(...TRUNCATED)
deepseek-v4.1-flash
true
stop
true
true
event_00003
aerial-cactus-identification
aerial-cactus-identification__sample_00004
4
review
[{"role":"system","content":"You are an expert machine learning experiment analyst.\nYour job is to (...TRUNCATED)
deepseek-v4.1-flash
true
stop
true
true
event_00004
aerial-cactus-identification
aerial-cactus-identification__sample_00005
5
review
[{"role":"system","content":"You are an expert machine learning experiment analyst.\nYour job is to (...TRUNCATED)
deepseek-v4.1-flash
true
stop
true
true
event_00005
aerial-cactus-identification
aerial-cactus-identification__sample_00006
6
review
[{"role":"system","content":"You are an expert machine learning experiment analyst.\nYour job is to (...TRUNCATED)
deepseek-v4.1-flash
true
stop
true
true
event_00006
aerial-cactus-identification
aerial-cactus-identification__sample_00007
7
review
[{"role":"system","content":"You are an expert machine learning experiment analyst.\nYour job is to (...TRUNCATED)
deepseek-v4.1-flash
true
stop
true
true
event_00007
aerial-cactus-identification
aerial-cactus-identification__sample_00008
8
code
[{"role":"system","content":"You are a redacted_person attending a high-stakes competition. In order(...TRUNCATED)
deepseek-v4.1-flash
true
stop
true
true
event_00008
aerial-cactus-identification
aerial-cactus-identification__sample_00009
9
review
[{"role":"system","content":"You are an expert machine learning experiment analyst.\nYour job is to (...TRUNCATED)
deepseek-v4.1-flash
true
stop
true
true
event_00009
aerial-cactus-identification
aerial-cactus-identification__sample_00010
10
code
[{"role":"system","content":"You are a redacted_person attending a high-stakes competition. In order(...TRUNCATED)
deepseek-v4.1-flash
true
stop
true
true
event_00010
End of preview. Expand in Data Studio

MLE-Lite benchmark 泄漏训练数据

教师模型:deepseek-v4.1-flash。数据来自既有的 22 个 MLE-Lite 原始 benchmark 任务,供 benchmark 污染对照训练使用。数据已脱敏,路径已改为相对引用。

内容 数量
完整已记录任务轨迹 22
实际模型请求 2,215
真实响应训练样本 2,214
正常结束响应 1,820
原 API 长度上限响应 394
未收到响应的请求 1
实际执行步骤 1,131

train.jsonl 包含全部真实响应;train_complete.jsonl 包含正常结束的回答,覆盖全部 22 题。394 条原本以 length 结束的响应保留标记;未响应请求只保存在轨迹中,不补造训练目标。记录中的思考、正文、工具调用和历史上下文均保留,敏感信息使用中性占位符替换。

“完整轨迹”指完整保留实际记录的调用和响应内容。轨迹包含分支搜索事件,不能将不同分支当作一条线性聊天。所有成功响应均保留,不按得分或提交成败筛选。本次转换不调用教师、不重新评分、不续写响应、不截断训练文本。

数据结构

  • train.jsonl:2,214 条训练样本。
  • train_complete.jsonl:1,820 条正常结束响应。
  • trajectories.jsonl.gz:22 行,每题一条轨迹;trajectory_json 是可解析的 JSON 字符串。
  • tasks.jsonl:任务索引、数量和相对文件引用。
  • tasks/<task_id>/:任务说明、索引、逐决策训练样本及完整轨迹。
  • source_exchanges/<task_id>.tar.gz:脱敏后的请求、响应、传输尝试与回执。流式响应按选择序号拼接文本字段,保存为 attempt_N.stream.json,保留结束状态及额外记录;不再保存分片形式。
  • execution_artifacts/<task_id>.tar.gz:脱敏后的代码、提示词、思考、执行日志、反馈和提交产物。
  • environment/encode.py:不截断的训练编码工具。
  • validation_report.json:实际验证结果。
  • tokenization_audit.jsonl.gz:逐样本 token 数。

轨迹中的 event_00001 等编号与归档目录一致。归档不包含原始机器、账户或文件系统位置的来源元数据。代码里的相对路径需要由使用者根据本地数据目录配置;执行档案是历史记录,并非可直接复现的运行环境。

归档内的匿名摘要文件使用顺序编号,保证解压时没有重名覆盖,所有文件均保留。

训练字段

每行包含 task_id、sample_id、source_step、action_type、messages、teacher_model、benchmark_leaked、finish_reason、response_complete、has_reasoning 和 source_exchange_id。

source_step 为任务内请求顺序,从 1 开始。messages 保留当时的 system/user 上下文以及当前 assistant 目标;上下文的 loss 为 false,仅最后一条 assistant 的 loss 为 true。真实思考使用 <think> 标记,工具调用使用 <tool_call> 标记。原响应的结构化字段保存在对应轨迹中。

读取和编码

下载数据集后,在数据集目录读取:

import json
from datasets import load_dataset

train = load_dataset('json', data_files='train.jsonl', split='train')
complete = load_dataset('json', data_files='train_complete.jsonl', split='train')
trajectories = load_dataset('json', data_files='trajectories.jsonl.gz', split='train')
first_trajectory = json.loads(trajectories[0]['trajectory_json'])

使用兼容 Qwen 的 tokenizer 编码:

python -m pip install -r environment/requirements.txt
python environment/encode.py \
  --input train_complete.jsonl \
  --tokenizer models/Qwen3.8-27B \
  --output encoded_complete \
  --max-length 262144

编码只监督当前 assistant,并保留其全部思考和回答。超出指定长度时直接报错,不静默截断;编码无需模型权重或 GPU。输出为含 input_ids、labels、lengths 的训练数据,上下文标签统一为 -100。使用全部已记录响应时将输入改为 train.jsonl。

这些任务本身就是训练来源,没有独立的验证或测试划分。在同一组任务上的成绩只能用于污染对照,不能作为独立泛化成绩。

Downloads last month
-