alam_pretrain

alam_pretrain_latent_action_tokenizer

A shared ALAM v3 latent-action tokenizer for MetaWorld MT50 and LIBERO, pretrained on Mix-11 (10 OXE video sources + CALVIN). It uses 7 latent-action slots, a 256-entry codebook, and 128-dimensional latent vectors.

The alam_pretrain_latent_action_tokenizer/ directory contains config.yaml and pytorch_model.bin. The project owner confirms that the shared tokenizer configuration has passed LIBERO validation.

From the GitHub code checkout:

.venvs/publish/bin/python workflows/publishing/download_huggingface.py \
  --artifact alam_pretrain
bash workflows/alam_pretraining/evaluate_checkpoint.sh cpu

Both downstream tasks use evaluation/checkpoints/alam/alam_pretrain_latent_action_tokenizer. For existing checkouts, set ALAM_METAWORLD_TOKENIZER and ALAM_LIBERO_TOKENIZER in .env to this path.

Use evaluate_checkpoint.sh cuda to check encoder execution. Full pretraining uses 128 NVIDIA H20 GPUs, per-GPU batch 32, and gradient accumulation 2.