ALAM pretrained tokenizer checkpoints
Repository ID: Mark-ZJTang/alam_pretrain.
This repository contains the two ALAM v3 tokenizer checkpoints used by the released downstream evaluations. Both use 7 latent-action slots, a 256-entry codebook, and 128-dimensional latent vectors.
| Directory | Downstream use | Training epoch/step |
|---|---|---|
metaworld_epoch19_step58216 |
MetaWorld MT50 | epoch 19, step 58,216 |
libero_epoch16_step49024 |
LIBERO Table 9 shared-checkpoint configuration | epoch 16, step 49,024 |
Each directory contains config.yaml and pytorch_model.bin. Verify the files
against evaluation/WEIGHTS_MANIFEST.sha256 in the GitHub code release. License
metadata must be completed by the copyright owner before public publication.
From the matching GitHub code checkout, restore both checkpoints and run strict CPU structure checks with:
.venvs/publish/bin/python workflows/publishing/download_huggingface.py \
--artifact alam_pretrain
bash workflows/alam_pretraining/evaluate_checkpoint.sh cpu
bash workflows/alam_pretraining/evaluate_checkpoint.sh cpu \
--checkpoint evaluation/checkpoints/alam/libero_epoch16_step49024
Use evaluate_checkpoint.sh cuda for real encoder execution; the preserved
encoder has a historical internal CUDA device assumption.
The project-owner resource specification for full ALAM pretraining is 128
NVIDIA H20 GPUs (one process per GPU), with per-GPU batch 32 and gradient
accumulation 2. Historical directory labels containing 64gpu are provenance
names, not the public resource contract.