subset_id stringclasses 1
value | runtime_protocol stringclasses 1
value | episodes int64 1k 1k | clean_seen int64 500 500 | randomized_unseen int64 500 500 | tasks int64 50 50 | episodes_per_task_setting int64 10 10 | repo_id stringclasses 1
value | hf_revision stringclasses 1
value | episodes_sha256 stringclasses 1
value | protocol_sha256 stringclasses 1
value | sha256sums_sha256 stringclasses 1
value | github_data_commit stringclasses 1
value | repo_type stringclasses 1
value | path stringclasses 1
value | required_files listlengths 4 4 | relationship_to_training stringclasses 1
value | initial_migration_release_contains_evaluation bool 1
class |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
robotwin-eval-subset-v2-seen-unseen | hybrid_seen_unseen_v2_rt | 1,000 | 500 | 500 | 50 | 10 | ar-mine/robotwin-eval-subset-v2-seen-unseen | 73277c2022be224d3afb195b503a3cff56ded46d | 5807f2a42ef9513f6efa685431f4a5de460ba06171d42bf4fc0d2deffcbc96e2 | 30a5a691f001c3845b5ca41ca3f4ea111610b52efb82509311c24e4066f96efb | a8a41f7c5162732cce287e77628840bf754f631825d3a6a343d51c648f932eb9 | b5abe512dafd400736c6dca5905f2fd262dd9029 | dataset | benchmarks/robotwin_eval_subset_v2_seen_unseen | [
"benchmarks/robotwin_eval_subset_v2_seen_unseen/episodes.jsonl",
"benchmarks/robotwin_eval_subset_v2_seen_unseen/protocol.json",
"benchmarks/robotwin_eval_subset_v2_seen_unseen/SHA256SUMS",
"benchmarks/robotwin_eval_subset_v2_seen_unseen/verify.py"
] | Independent frozen simulator episode specifications; not a held-out split of the 2500 demonstrations | false |
robotwin-eval-subset-v2-seen-unseen | hybrid_seen_unseen_v2_rt | 1,000 | 500 | 500 | 50 | 10 | ar-mine/robotwin-eval-subset-v2-seen-unseen | 73277c2022be224d3afb195b503a3cff56ded46d | 5807f2a42ef9513f6efa685431f4a5de460ba06171d42bf4fc0d2deffcbc96e2 | 30a5a691f001c3845b5ca41ca3f4ea111610b52efb82509311c24e4066f96efb | a8a41f7c5162732cce287e77628840bf754f631825d3a6a343d51c648f932eb9 | b5abe512dafd400736c6dca5905f2fd262dd9029 | dataset | benchmarks/robotwin_eval_subset_v2_seen_unseen | [
"benchmarks/robotwin_eval_subset_v2_seen_unseen/episodes.jsonl",
"benchmarks/robotwin_eval_subset_v2_seen_unseen/protocol.json",
"benchmarks/robotwin_eval_subset_v2_seen_unseen/SHA256SUMS",
"benchmarks/robotwin_eval_subset_v2_seen_unseen/verify.py"
] | Independent frozen simulator episode specifications; not a held-out split of the 2500 demonstrations | false |
robotwin-eval-subset-v2-seen-unseen | hybrid_seen_unseen_v2_rt | 1,000 | 500 | 500 | 50 | 10 | ar-mine/robotwin-eval-subset-v2-seen-unseen | 73277c2022be224d3afb195b503a3cff56ded46d | 5807f2a42ef9513f6efa685431f4a5de460ba06171d42bf4fc0d2deffcbc96e2 | 30a5a691f001c3845b5ca41ca3f4ea111610b52efb82509311c24e4066f96efb | a8a41f7c5162732cce287e77628840bf754f631825d3a6a343d51c648f932eb9 | b5abe512dafd400736c6dca5905f2fd262dd9029 | dataset | benchmarks/robotwin_eval_subset_v2_seen_unseen | [
"benchmarks/robotwin_eval_subset_v2_seen_unseen/episodes.jsonl",
"benchmarks/robotwin_eval_subset_v2_seen_unseen/protocol.json",
"benchmarks/robotwin_eval_subset_v2_seen_unseen/SHA256SUMS",
"benchmarks/robotwin_eval_subset_v2_seen_unseen/verify.py"
] | Independent frozen simulator episode specifications; not a held-out split of the 2500 demonstrations | false |
robotwin-eval-subset-v2-seen-unseen | hybrid_seen_unseen_v2_rt | 1,000 | 500 | 500 | 50 | 10 | ar-mine/robotwin-eval-subset-v2-seen-unseen | 73277c2022be224d3afb195b503a3cff56ded46d | 5807f2a42ef9513f6efa685431f4a5de460ba06171d42bf4fc0d2deffcbc96e2 | 30a5a691f001c3845b5ca41ca3f4ea111610b52efb82509311c24e4066f96efb | a8a41f7c5162732cce287e77628840bf754f631825d3a6a343d51c648f932eb9 | b5abe512dafd400736c6dca5905f2fd262dd9029 | dataset | benchmarks/robotwin_eval_subset_v2_seen_unseen | [
"benchmarks/robotwin_eval_subset_v2_seen_unseen/episodes.jsonl",
"benchmarks/robotwin_eval_subset_v2_seen_unseen/protocol.json",
"benchmarks/robotwin_eval_subset_v2_seen_unseen/SHA256SUMS",
"benchmarks/robotwin_eval_subset_v2_seen_unseen/verify.py"
] | Independent frozen simulator episode specifications; not a held-out split of the 2500 demonstrations | false |
robotwin-eval-subset-v2-seen-unseen | hybrid_seen_unseen_v2_rt | 1,000 | 500 | 500 | 50 | 10 | ar-mine/robotwin-eval-subset-v2-seen-unseen | 73277c2022be224d3afb195b503a3cff56ded46d | 5807f2a42ef9513f6efa685431f4a5de460ba06171d42bf4fc0d2deffcbc96e2 | 30a5a691f001c3845b5ca41ca3f4ea111610b52efb82509311c24e4066f96efb | a8a41f7c5162732cce287e77628840bf754f631825d3a6a343d51c648f932eb9 | b5abe512dafd400736c6dca5905f2fd262dd9029 | dataset | benchmarks/robotwin_eval_subset_v2_seen_unseen | [
"benchmarks/robotwin_eval_subset_v2_seen_unseen/episodes.jsonl",
"benchmarks/robotwin_eval_subset_v2_seen_unseen/protocol.json",
"benchmarks/robotwin_eval_subset_v2_seen_unseen/SHA256SUMS",
"benchmarks/robotwin_eval_subset_v2_seen_unseen/verify.py"
] | Independent frozen simulator episode specifications; not a held-out split of the 2500 demonstrations | false |
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Lingbot VLA transfer — Hugging Face only
已完成实验：E5 allclean2epoch global64
作业25708998已完成17180步。此前的 e5-allclean-2epoch-gbs64-hf-only-b-20261009 是fresh复现包，
现已补发实际训练好的最终模型（FP32存储，6分片，约23.77GiB）。
模型固定revision：b3306e5b517d33815d7cafe392d5cc326af97bc2。
完成态实验与使用说明｜机器可读索引。
原source包和旧manifest保留；本页下面的八H200实验仍是另一项尚未训练的新配置。
迁入端不需要GitHub账号、git clone或GitHub网络。所有接收、配置和训练入口都在固定HF包内。
新实验：E5八H200，micro24，GA2，6epoch
- Release:
e5-8h200-micro24-ga2-gbs384-ep6-lr2e4-20261009 - Bundle commit:
9362d9b12da972fe48703192ba20b696c007dda0 - Manifest SHA256:
e7fb3eb016663192b255b5d7cd8a8b77e82aabdd885c78c18c325cb943134e5e - Bootstrap SHA256:
6987699047d681d850e27e71d558a0eed9dbe50591798db65d2b331e0acab7f9 - 完整接收与训练操作说明
- 固定目标配置
单节点8GPU，micro24×GA2×8=global384；6epoch，每epoch1431步，总8586步。 主LR2e-4、cosine最低LR1e-4、视觉编码器LR1e-6。与上一版四卡实验的global BS和epoch一致。 从同一官方基础模型fresh开始，BF16混合精度/FP32主权重，全量微调、无LoRA。 2500clean轨迹、549787帧全部用于训练；每1000步及最后8586步保存DCP和FP32 HF权重。
下载与运行
需要Python3.12（可用uv安装）、curl、sha256sum和HF网络；下载到新目录：
curl -fL --retry 3 -o bootstrap_receiver.py 'https://huggingface.co/datasets/ar-mine/lingbot-vla-transfer/resolve/9362d9b12da972fe48703192ba20b696c007dda0/releases/e5-8h200-micro24-ga2-gbs384-ep6-lr2e4-20261009/receiver/bootstrap_receiver.py'
printf '%s %s\n' '6987699047d681d850e27e71d558a0eed9dbe50591798db65d2b331e0acab7f9' bootstrap_receiver.py | sha256sum --check
python3.12 bootstrap_receiver.py --revision '9362d9b12da972fe48703192ba20b696c007dda0' --release 'e5-8h200-micro24-ga2-gbs384-ep6-lr2e4-20261009' --manifest-sha256 'e7fb3eb016663192b255b5d7cd8a8b77e82aabdd885c78c18c325cb943134e5e' --output /mnt/FAST/e5-8h200-transfer
也可使用根目录的 receive-e5-8h200-m24-ga2-ep6-20261009.sh 下载。48个载荷文件在READY发布前已全部回读校验。
按包内说明用uv准备兼容训练环境、下载固定模型并转换数据后：
export VLA_B_ROOT=/mnt/FAST/vla-B-project
export VLA_B_PAYLOAD=/mnt/FAST/e5-8h200-transfer/payload
export VLA_B_RUN=/mnt/FAST/e5-8h200-m24-ga2-ep6
"$VLA_B_ROOT/.venv/bin/python" "$VLA_B_PAYLOAD/receiver/e5_experiment.py" prepare --root "$VLA_B_ROOT" --bundle-dir "$VLA_B_PAYLOAD" --output "$VLA_B_RUN"
"$VLA_B_ROOT/.venv/bin/python" "$VLA_B_PAYLOAD/receiver/receiver_interfaces.py" "$VLA_B_RUN"
"$VLA_B_ROOT/.venv/bin/python" "$VLA_B_PAYLOAD/receiver/e5_experiment.py" train --output "$VLA_B_RUN"
train缺少smoke结果时先运行八卡10步smoke，全部八个rank通过后fresh启动正式训练。 单节点八卡启动器不支持直接跨两个四卡节点使用。不会自动改batch或复用旧smoke输出。 54项测试通过；未执行目标机器的H200训练，容量仍须由接收端smoke确认。 发布工具uv.lock不是训练环境锁；当前仍需按照文档建立兼容训练环境。
其他固定版本
上一版四卡六epoch：
e5-4h200-micro96-ga1-gbs384-ep6-lr2e4-20261009，commit46c450a2ebc2e9ff11ee4c626e9cdfff12083771。原四epoch实验：
e5-4h200-micro96-ga1-gbs384-ep4-lr2e4-20261009，commit80869ea26d27f26008e0db30566cfbe984d019f6。通用batch重规划接收器：
e5-hf-only-b-batch-plan-20261009，commitee21c0342170421288bc00d3eb9a098b2608473d。HF-only首版：
e5-allclean-2epoch-gbs64-hf-only-b-20261009，commit62aa778aeb50249e6437bcac0458e307c309a5ca。A端首版：
e5-allclean-2epoch-gbs64-sourceA-20261008，commitd5b7c57dee106e46e10a80ad5c5ab1c7a208eced。
旧release和下载入口保持不变；不存在跨机完整状态续训或多节点接收启动器。 源码保留各上游LICENSE，不对组合包赋予新的统一许可证。
- Downloads last month
- 477