Datasets:
task_id stringclasses 18
values | sample_id stringlengths 42 63 | source_step int64 1 274 | action_type stringclasses 2
values | messages listlengths 2 2 |
|---|---|---|---|---|
aerial-cactus-identification__00 | aerial-cactus-identification__00__sample_00001 | 1 | code | [
{
"role": "user",
"content": "# Introduction\n\nYou are a Kaggle grandmaster attending a 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. We will now provide a description of the task.\n\n# Task ... |
aerial-cactus-identification__00 | aerial-cactus-identification__00__sample_00002 | 3 | code | [
{
"role": "user",
"content": "# Introduction\n\nYou are a Kaggle grandmaster attending a 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. We will now provide a description of the task.\n\n# Task ... |
aerial-cactus-identification__00 | aerial-cactus-identification__00__sample_00003 | 4 | code | [
{
"role": "user",
"content": "# Introduction\n\nYou are a Kaggle grandmaster attending a 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. We will now provide a description of the task.\n\n# Task ... |
aerial-cactus-identification__00 | aerial-cactus-identification__00__sample_00004 | 6 | code | [{"role":"user","content":"# Introduction\n\nYou are a Kaggle grandmaster attending a competition. Y(...TRUNCATED) |
aerial-cactus-identification__00 | aerial-cactus-identification__00__sample_00005 | 7 | code | [{"role":"user","content":"# Introduction\n\nYou are a Kaggle grandmaster attending a competition. Y(...TRUNCATED) |
aerial-cactus-identification__00 | aerial-cactus-identification__00__sample_00006 | 8 | code | [{"role":"user","content":"# Introduction\n\nYou are a Kaggle grandmaster attending a competition. Y(...TRUNCATED) |
aerial-cactus-identification__00 | aerial-cactus-identification__00__sample_00007 | 9 | code | [{"role":"user","content":"# Introduction\n\nYou are a Kaggle grandmaster attending a competition. Y(...TRUNCATED) |
aerial-cactus-identification__00 | aerial-cactus-identification__00__sample_00008 | 10 | code | [{"role":"user","content":"# Introduction\n\nYou are a Kaggle grandmaster attending a competition. Y(...TRUNCATED) |
aerial-cactus-identification__00 | aerial-cactus-identification__00__sample_00009 | 12 | code | [{"role":"user","content":"# Introduction\n\nYou are a Kaggle grandmaster attending a competition. Y(...TRUNCATED) |
aerial-cactus-identification__00 | aerial-cactus-identification__00__sample_00010 | 13 | code | [{"role":"user","content":"# Introduction\n\nYou are a Kaggle grandmaster attending a competition. Y(...TRUNCATED) |
长程数据
这是已训练 Qwen long 对照模型实际使用的历史批次的发布版:807 条训练样本、149 条验证样本,来自 22 个独立构造任务。原训练从 Qwen3.8-27B 基座初始化新 LoRA,最终选择第 3 个完整 epoch 的 checkpoint-303。
教师为 DeepSeek v4.1 Flash,开启思考模式,每任务最长 24 小时。数据保留真实教师思考、代码动作、原生 review 动作及当时可见的历史执行反馈;仅当前 assistant 响应参与监督。
加载与字段
from datasets import load_dataset
dataset = load_dataset("jankin123/mle-data-long-1106")
assert len(dataset["train"]) == 807
assert len(dataset["validation"]) == 149
每行包含以下字段:
| 字段 | 含义 |
|---|---|
task_id |
构造任务标识,对应任务目录及环境压缩包 |
sample_id |
发布版本中的唯一、匿名样本编号 |
source_step |
原真实执行轨迹中的动作步号 |
action_type |
code 或 review |
messages |
两条消息:历史 user 上下文、当前 assistant 响应 |
messages[].loss |
user 为 false,assistant 为 true |
训练集包含 374 个 code 动作、433 个 review 动作;验证集包含 70 个 code 动作、79 个 review 动作。同一步可以同时有代码生成与执行后 review,不能按 task_id + source_step 合并或去重。sample_id 提供唯一标识。
assistant 内容保留 <think>...</think>,随后是完整 Python 程序,或原生 submit_review 工具调用文本。review 并非新增解题答案,也不是人工编造的评语。历史任务、代码、日志和反馈随原上下文保留,均不作为当前训练目标。
文件组织
train.jsonl
validation.jsonl
tasks.jsonl
tasks/<task_id>/task.md
tasks/<task_id>/task.json
task_data/<task_id>.tar.gz
configs/training.json
configs/deepspeed.json
configs/rollout.json
configs/environment.json
configs/encoding.json
training/exact_template.py
training/inference_template.jinja
training/train.py
environment/prepare_task.py
environment/grade_submission.py
environment/metrics.py
requirements.txt
tasks.jsonl 记录每个任务的源竞赛、train/validation 划分、动作数、描述、配置和数据包位置。18 个完整环境用于训练,4 个完整环境用于验证。验证环境固定为:
jigsaw-toxic-comment-classification-challenge__07plant-pathology-2020-fgvc7__01spooky-author-identification__06tabular-playground-series-may-2022__06
划分按完整环境预先固定,不按样本随机混拆。不同变体可能共享源竞赛及原始公开记录;验证结果衡量共享源数据池上的保留任务拟合情况。
| 源竞赛家族 | 训练任务 | 验证任务 | 训练样本 | 验证样本 |
|---|---|---|---|---|
| aerial-cactus-identification | 1 | 0 | 42 | 0 |
| detecting-insults-in-social-commentary | 3 | 0 | 125 | 0 |
| dog-breed-identification | 1 | 0 | 35 | 0 |
| jigsaw-toxic-comment-classification-challenge | 3 | 1 | 126 | 19 |
| nomad2018-predict-transparent-conductors | 2 | 0 | 92 | 0 |
| plant-pathology-2020-fgvc7 | 1 | 1 | 52 | 66 |
| spooky-author-identification | 2 | 1 | 90 | 37 |
| tabular-playground-series-dec-2021 | 2 | 0 | 89 | 0 |
| tabular-playground-series-may-2022 | 1 | 1 | 30 | 27 |
| text-normalization-challenge-english-language | 2 | 0 | 126 | 0 |
| 合计 | 18 | 4 | 807 | 149 |
真实任务输入与评分
任务从源竞赛授权公开 TRAIN 数据构造,保留原生文本、表格、图像或晶体结构等输入,并按任务划分生成训练与独立构造留出集。没有原 MLE-Bench 私有测试数据或私有测试标签。
每个压缩包包含 input/ 与 grading/。grading/hidden_answers.csv 是本构造任务的独立评分标签,解题时应由评分端保管。公开数据与代码引用中的匿名标识保持对应。CSV、内嵌 ZIP、归档文件名和图片元数据均纳入发布检查。
python environment/prepare_task.py \
--task detecting-insults-in-social-commentary__00 \
--output task_work
从 task_work 运行解题代码,读取 input,将结果写到 submission/submission.csv。独立评分:
python environment/grade_submission.py \
--task detecting-insults-in-social-commentary__00 \
--task-dir task_work \
--split hidden \
--predictions task_work/submission/submission.csv
这里计算的是构造任务指标。各任务的标签、预测列与指标以 tasks/<task_id>/task.json 为准;这些指标不能跨竞赛直接平均。
原始生成配置
| 参数 | 设置 |
|---|---|
| 解题代理 | AIDE v6.3.3,包含上游 PR85 与有界超时收尾封装 |
| 教师 | DeepSeek v4.1 Flash |
| 思考模式 | 开启,保留实际思考文本 |
| 每任务最长时间 | 86400 秒(24 小时) |
| 最大探索执行步数 | 500 |
| 单次执行/API 超时 | 32400 秒(9 小时) |
| 输入输出共享上下文窗口 | 262144 token |
| 单次最大输出 | 131072 token |
| temperature / top_p | 1.0 / 0.95 |
| presence_penalty | 0.0 |
| 达到留出目标后提前结束 | 关闭 |
解题过程使用真实执行日志、训练折内指标和提交格式反馈。最终候选冻结后,构造留出评分只用于独立质量筛选,不作为解题代理的搜索反馈。发布 JSONL 是原训练实际选中的动作切片,包含各动作当时的完整上下文,并非额外发布所有原始网络交换或完整运行目录。
对应的 Qwen long 训练
| 参数 | 设置 |
|---|---|
| 初始化 | 原始 Qwen3.8-27B,新 LoRA |
| 完整训练/验证样本数 | 807 / 149 |
| 精度 | BF16 |
| 进程数 / 原生 sequence parallel | 4 / 4 |
| 实际数据并行组数 | 1 |
| 全局 batch | 8 |
| 每设备 micro batch / 梯度累积 | 1 / 8 |
| LoRA rank / alpha / dropout | 16 / 32 / 0 |
| LoRA 模块 | all-linear |
| 学习率 / warmup ratio | 5e-5 / 0.05 |
| 实际完整 epoch / 优化步数 | 3 / 303 |
| 最大上下文长度 | 262144 token,不截断样本 |
| 损失范围 | 当前响应,忽略思考起始前缀;历史上下文全部掩码 |
| 检查点选择 | 完整 epoch 的整环境验证 NLL |
原批次验证 NLL 从 0.5760932 降至 0.46868446;检查点选择没有使用 benchmark 分数。这是原训练的历史记录。脱敏公开数据的编码统计见 configs/encoding.json。
在支持该模型与原生 SP4 的 ms-swift 环境中,可参考:
NPROC_PER_NODE=4 python training/train.py
可通过 BASE_MODEL 和 OUTPUT_DIR 指定模型位置及输出位置。入口保留原训练超参数;软件兼容性与实际显存需求取决于所用 Swift、PyTorch 和注意力实现。
发布检查
公开版全部 956 条已通过实际 Swift 数据加载、编码及损失掩码检查。未删除、未截断样本;历史上下文掩码、当前思考与最终响应监督范围、推理前缀均逐条验证。
| 划分 | 样本数 | 完整序列最短 | P50 | P95 | 最长 |
|---|---|---|---|---|---|
| train | 807 | 3010 | 12382.0 | 38357.2 | 109641 |
| validation | 149 | 3321 | 10698.0 | 29805.2 | 38608 |
- Downloads last month
- 102