本仓库记录第一阶段 AMD 八卡全参数训练的 globalstep10000。配置、脚本、逐任务评测结果公开供团队复现与分析。 完整 DCP 原件为 76.589 GB / 71.329 GiB（文件字节合计）；加配套配置与 modelassets 的续训共享目录为 76.604 GB。模型含 6,375,907,511 个张量元素；FP32 权重数据为 25.504 GB / 23.752 GiB，本次推理目录含配置实际 25.520 GB。另有 15.887 MB Qwen架构/tokenizer依赖，整体共享文件约 102.140 GB。GB 按十进制，GiB 按二进制。目录清单与实际 SHA256 见 HF 的 manifest.json。 - 数据：RoboTwin 2.0，Aloha-AgileX，50 个 clean 任务，每任务 50 条示范；共 2,500 条。原云端元数据审计合计 549,787 帧。 - 训练：10,000 个优化步，8 卡 FSDP2/full shard，全参数 AdamW；每卡 micro batch=16，梯度累积=2，global batch=256；动作 chunk=50。 - 学习率：1e-4 → 5e-5，cosine；保存间隔 5,000 步；原训练未启用视觉增强。 - 原配置 enablefp32=false、enablemixedprecision=true。DCP 保存的参数张量全部为 FP32；保存精度与训练计算精度是不同概念。…
Independent publisher
Fantasyyy
Fantasyyy
Models in Library1
Datasets in Library0
Models on Hugging Face2
Followers—
Compare the published evidence
This profile connects the 1 model and dataset record shown below to their upstream cards and sources. The library counts describe SAVRN's coverage, rather than this publisher's entire output.
Published tasks: Robotics.
Compare model specifications, source dates, and hardware estimates on the individual records. Publisher identity verification is separate from a model's quality, deployment readiness, or suitability for your work.