Datasets:

ar-mine
/

deepstack-all-gates-handoff-20261019

Languages:

Tags:

Dataset card Files Files and versions

xet

Community

Dataset Viewer

The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.

Couldn't infer the same data file format for all splits. Got {NamedSplit('validation'): ('json', {}), NamedSplit('test'): (None, {})}

Error code:   FileFormatMismatchBetweenSplitsError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

DeepStack 三路门控训练迁移包

本包接替已取消、尚未运行的NSCC作业25877127。任务是投影＋全部三路门控，沿用25844379的匹配训练配置，含真实训练smoke、原首步梯度/更新参考、完整冻结数据及agent任务说明。

接收机器的agent请先读 AGENT_TASK.md。 当前发布验证仅包含CPU/静态检查及原NSCC运行证据；接收机器必须先通过GPU smoke，才能开始正式训练。

下载与准备

需要Python3.12、NVIDIA CUDA12.8兼容驱动、4张至少32GiB GPU、建议256GiB主存和150GiB可用磁盘。源硬件是4张A100 40GB。无需NSCC账号。

# 使用发布者给出的完整COMMIT、MANIFEST_SHA256，以及BOOTSTRAP_SHA256。
export COMMIT=<full_release_commit>
export MANIFEST_SHA256=<publisher_manifest_sha256>
export BOOTSTRAP_SHA256=<publisher_bootstrap_sha256>
curl -fL "https://huggingface.co/datasets/ar-mine/deepstack-all-gates-handoff-20261019/resolve/$COMMIT/bootstrap.py" -o bootstrap.py
printf '%s  bootstrap.py\n' "$BOOTSTRAP_SHA256" | sha256sum -c -
python3.12 -m pip install huggingface-hub==0.34.3 hf-xet
python3.12 bootstrap.py --revision "$COMMIT" --manifest-sha256 "$MANIFEST_SHA256" \
  --package-dir /data/deepstack-package --workspace /data/deepstack-all

Bootstrap会匿名下载并验证本任务包与约4.2GiB原转换数据、约1.3GiB严格首步参考，再从固定commit下载父权重和依赖。父模型是 ar-mine/lingbot-vla-v2-6b-e5-c3-nscc25723372-step17180，不是旧迁移工具默认的官方base初始化。

环境和测试

新建隔离环境。已有兼容环境可复用，但仍须preflight。Ubuntu需FFmpeg及CUDA toolkit；FlashAttention可能需要编译。environment/source-freeze.txt是原环境审计清单，environment/install.sh与Dockerfile是安装起点，尚无干净目标机器安装验收。源码内第三方模块已包含，不需要GitHub账号。

# 示例：uv venv --python 3.12.13 --seed /data/deepstack-venv
export DEEPSTACK_PYTHON=/data/deepstack-venv/bin/python
cd /data/deepstack-package
bash environment/install.sh
export WORK=/data/deepstack-all
export CUDA_VISIBLE_DEVICES=0,1,2,3
export PYTHONPATH="$WORK/runtime/source:$WORK/runtime/source/scripts:$WORK/runtime/source/third_party/lingbot-vla-v2:$WORK/runtime/source/third_party/utils3d:$WORK/runtime/source/third_party/lingbot-vla-v2/lingbotvla/models/vla/vision_models/MoGe:$WORK/runtime/source/third_party/lingbot-vla-v2/lingbotvla/models/vla/vision_models/lingbot-depth"
"$DEEPSTACK_PYTHON" -m unittest discover -s "$WORK/runtime/training" -p 'test_*.py'
"$DEEPSTACK_PYTHON" "$WORK/tools/preflight.py" --workspace "$WORK" --gpu
"$DEEPSTACK_PYTHON" "$WORK/tools/run.py" smoke --workspace "$WORK"
"$DEEPSTACK_PYTHON" "$WORK/tools/run.py" train --workspace "$WORK"

smoke使用同一加载器、原图/C3双分支、micro1/GA16/global64，运行2步和小验证，保留原严格首步参考与FP32恒等/保存重载检查。正式训练重新从父权重开始，1000步、250帧开发验证、每250步保存。源A100完整投影训练约7小时；另一种GPU的时长需要先实测。所有日志写入WORK/runs/，输出路径不会覆盖旧尝试。

接收结果与边界

正式结果为runs/train_TIMESTAMP/checkpoint_001000/delta.pt及其complete.json、顶层与四rankacceptance.json、first_step_guard及验证记录。所有验收通过才报告完成。更换硬件若导致MoE数值不对齐，要停在smoke并保留诊断，不得放宽容差。此包不自动启动闭环评估。

冻结数据来自 TianxingChen/RoboTwin2.0。使用原2250/250划分、64k训练帧与逐样本noise/time；保留各上游源码license，权重条款以各公开仓库为准。任务只使用正常观测，不使用隐藏供体、分割真值或未见指令训练。

Downloads last month: 150

Total file size:

5.57 GB