GR00T N1.6 + HAMLET + Persistent TTT β RoboMME NMT T=4
Private archival checkpoint for Nuri's persistent test-time-training experiment on RoboMME. It extends the HAMLET K=4 attention-memory policy with recurrent fast weights trained using a next-moment-token inner objective.
Architecture
- Starting checkpoint:
morealcholplz/ttt-vla-robomme-stage2-hamlet-k4 - Four HAMLET moment tokens per observation
- Memory implementation: persistent TTT fast weights plus the HAMLET attention branch
- Sequence length T=4, attention window k=4
- Memory stride: 16 environment steps
- Three inner updates per four-block sequence
- Inner objective: next moment token (
ttt_inner_loss=nmt) - Sequential outer loss enabled
- Repeated/padded-block masking enabled
- TTT mini-batch: 1 block
- Inner learning rate: 0.1, learnable
- Inner gradient clipping: 1.0
- One fast SwiGLU layer
- Fast branch gate initialization: 0.001
Training
- Base model:
nvidia/GR00T-N1.6-3B/nvidia/Eagle-Block2A-2B-v2 - Dataset: RoboMME (
Yinpei/robomme_data_lerobot) - Learning rate: 1e-4
- Global batch size: 32 (16 per GPU)
- GPUs: 2
- Training steps: 60,000 / 60,000
- Total runtime: approximately 35 h 33 min
- Final mean training loss: approximately 0.00907
- Final step loss: approximately 0.0088
- Original experiment:
nuri_nmt_seq_t4_2gpu_batch16_masked_r4 - W&B: https://wandb.ai/khnuri-rwh/ttt-vla-robomme/runs/4xl6seab
Source
- Project: https://github.com/Star-ry/ttt-vla
- Persistent TTT implementation:
test/nuri/ttt-vla/ - The preserved Nuri workspace was introduced in commit
f493f3e; the exact runtime commit was not embedded in the checkpoint artifact.
Checkpoint choice
This repository contains the final 60k model. A separate 40k intermediate checkpoint existed in the original archive but was omitted from the default Hub upload to avoid an additional approximately 6.4 GB. No RoboMME rollout evaluation was found alongside the archive, so 60k is the final training step rather than a verified best-performing checkpoint.
Notes
DeepSpeed restart state, scheduler/RNG state, duplicate checkpoint copies, and the local training log were excluded. Loading requires the custom GR00T/HAMLET/TTT implementation from the source repository.
Use and redistribution remain subject to the licenses and terms of the NVIDIA base model, HAMLET implementation, and RoboMME dataset.
- Downloads last month
- 9
Model tree for morealcholplz/ttt-vla-robomme-ttt-nmt-t4
Base model
nvidia/GR00T-N1.6-3B