Robotics
Safetensors
world-action-models
robotwin
FreDERicia commited on
Commit
1193e91
·
verified ·
1 Parent(s): b33f9f2

Update concise Chinese model card

Browse files
Files changed (1) hide show
  1. README.md +45 -15
README.md CHANGED
@@ -1,33 +1,63 @@
1
  ---
2
  tags:
3
  - robotics
 
4
  - robotwin
5
  - safetensors
6
  datasets:
 
7
  - FreDERicia/SWAM-Posttrain
8
  ---
9
 
10
- # SWAM checkpoints
11
 
12
- This private repository contains the paired pretraining and posttraining checkpoints selected for SWAM.
13
 
14
- | Directory | Checkpoint | Source |
15
- | --- | --- | --- |
16
- | `pretrain/` | 25-4 pretraining, global step 80,000 | Embodied-model storage; initialization checkpoint for the posttraining run below |
17
- | `posttrain/` | 25-4 RoboTwin Clean posttraining, global step 12,000 | DNA evaluation checkpoint |
18
 
19
- Both checkpoints were trained with DMuon. Each contains seven original Safetensors shards and a weight index. `manifest.json` provides file sizes and SHA-256 checksums. Weight files are copied without conversion or modification. Optimizer state and training YAML files are not included.
20
 
21
- The 12k checkpoint's audited RoboTwin evaluation is Clean **89.04%** (4,452/5,000), Aug **63.56%** (3,178/5,000), and overall **76.30%** (7,630/10,000).
22
 
23
- ## Companion resources
 
 
 
24
 
25
- - Posttraining data: [FreDERicia/SWAM-Posttrain](https://huggingface.co/datasets/FreDERicia/SWAM-Posttrain)
26
- - Pretraining data: [FreDERicia/SWAM-Pretrain](https://huggingface.co/datasets/FreDERicia/SWAM-Pretrain)
27
- - Existing project code: [shaaaaron/SWAM](https://github.com/shaaaaron/SWAM)
28
 
29
- ## Loading scope
30
 
31
- The original `config.json` files are retained as checkpoint metadata. Their component and tokenizer paths refer to the original training/evaluation machines and require local overrides. External Wan/UMT5 assets are not bundled. Posttraining normalization metadata is retained in `posttrain/experiment_cfg/metadata.json`.
32
 
33
- This upload verifies the checkpoint index, shard structure, sizes, and SHA-256 integrity. It does not certify a standalone `from_pretrained` workflow or alignment with the current public training examples. No training scripts or GitHub changes are part of this upload.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  tags:
3
  - robotics
4
+ - world-action-models
5
  - robotwin
6
  - safetensors
7
  datasets:
8
+ - FreDERicia/SWAM-Pretrain
9
  - FreDERicia/SWAM-Posttrain
10
  ---
11
 
12
+ # SWAM
13
 
14
+ [![SWAM 项目封面](https://huggingface.co/datasets/FreDERicia/SWAM-Pretrain/resolve/main/visual-previews/swam-cover-cropped.png)](https://char-lotte-xia.github.io/swam-official/)
15
 
16
+ [项目主页](https://char-lotte-xia.github.io/swam-official/) · [论文](https://arxiv.org/abs/2611.12870) · [GitHub](https://github.com/shaaaaron/SWAM)
 
 
 
17
 
18
+ 本仓库提供 **SWAM — Scaling World Action Models with Cross-Modal Alignment** 的预训练权重与 RoboTwin 后训练权重。
19
 
20
+ ## 模型权重
21
 
22
+ | 权重 | 下载目录 |
23
+ | --- | --- |
24
+ | 预训练模型 | [pretrain](https://huggingface.co/FreDERicia/SWAM/tree/main/pretrain) |
25
+ | RoboTwin 后训练模型 | [posttrain](https://huggingface.co/FreDERicia/SWAM/tree/main/posttrain) |
26
 
27
+ 后训练模型以上述预训练权重初始化,使用 RoboTwin Clean 数据训练。每份权重约 **27.98 GB**,以 Safetensors 分片存储。
 
 
28
 
29
+ ## 下载与使用
30
 
31
+ 安装 Hugging Face CLI。仓库当前为 private,下载前需使用具备访问权限的账号登录。
32
 
33
+ ```bash
34
+ pip install -U huggingface_hub
35
+ hf auth login
36
+
37
+ # 下载预训练权重
38
+ hf download FreDERicia/SWAM --include "pretrain/*" --local-dir ./SWAM
39
+
40
+ # 下载后训练权重
41
+ hf download FreDERicia/SWAM --include "posttrain/*" --local-dir ./SWAM
42
+ ```
43
+
44
+ 训练与评测入口见 [SWAM 代码仓库](https://github.com/shaaaaron/SWAM)。使用时需将加载配置中的资源路径改为本地路径,并准备所需的 Wan/UMT5 组件;这些外部资源未包含在本仓库中。
45
+
46
+ ## 配套数据
47
+
48
+ [预训练数据](https://huggingface.co/datasets/FreDERicia/SWAM-Pretrain) · [后训练数据](https://huggingface.co/datasets/FreDERicia/SWAM-Posttrain)
49
+
50
+ ## 引用
51
+
52
+ 如果本项目对你的研究有帮助,请引用 SWAM 论文。
53
+
54
+ ```bibtex
55
+ @misc{zhou2026swam,
56
+ title = {SWAM: Scaling World Action Models with Cross-Modal Alignment},
57
+ author = {Yu Zhou and Xinyuan Xia and Xuankun Yang and Yinghao Yang and Wendi Chen and Bowen Liu and Jianhan Ma and Weicheng Duan and Jun Lv and Cewu Lu and Yanwei Li},
58
+ year = {2026},
59
+ eprint = {2611.12870},
60
+ archivePrefix = {arXiv},
61
+ url = {https://arxiv.org/abs/2611.12870}
62
+ }
63
+ ```