tags:
- robotics
- lingbot-vla
E5+C3 original training code and evaluation documentation
Public release, 2026-10-25 (Asia/Singapore). This Hugging Face dataset hosts a Git source
repository, cloneable with git clone https://huggingface.co/datasets/ar-mine/e5-c3-code.
It is not a training dataset or a generic AutoModel implementation.
Epoch2 weights · Epoch6 weights · Training method and reproduction · Evaluation method · Hardware · Full results.
| Model | Clean / seen | Randomized / unseen | Overall |
|---|---|---|---|
| ep2 | 381/500 (76.2%) | 204/500 (40.8%) | 585/1000 (58.5%) |
| ep6 | 416/500 (83.2%) | 178/500 (35.6%) | 594/1000 (59.4%) |
These are completed historical evaluations; the release task ran no new rollouts or training. Epoch2 and epoch6 differ in global batch (64/192), LR (1e-4/1.5e-4), GPU count (4/8) and epoch count. They are not a pure epoch ablation. Clean/seen versus randomized/unseen also changes the instruction pool, so it is not a pure visual randomization effect. All 50 tasks appear in training; seen/unseen describes instruction vocabulary/template pools, not unseen tasks. These are the original fixed-table runs, not later low-table adaptation or interpolation experiments.
The actual original training archives supply the shared runtime, training entry, E5 history
logic, C3 augmentation, normalization and configs. Selected shared files are byte-identical
between endpoints before the documented portable path changes. SOURCE_MANIFEST.json
records original/export SHA256 values, original archive SHA256 values and portable edits.
Original archive contents, machine-specific mappings and records remain local: 原始记录保存在本地。
Upstream licenses are preserved in third_party/lingbot-vla-v2/LICENSE and nested dependency
LICENSE files. Original upstream illustrations/loss plots are omitted; they are not these runs'
curves. Deployment allocation scripts and private machine layouts are not required by the
portable training entry. See the explicit remaining reproduction limits in the method documents.