easyminnn/utt-s0-checkpoints

UTT (Unified Tactile Tokenizer) — S0 pretraining checkpoints

Work-in-progress checkpoints of the UTT sensor-agnostic tactile encoder/decoder, published during training for evaluation. Code: jiminlx/cosmos-framework, branch tactile-tower-tva, cosmos_framework/model/generator/tokenizers/utt/.

• <stage>/index.json: uploaded steps with val/loss, latest_step, best_step (lowest val/loss).
• <stage>/step_XXXXXXX/model.pt: model weights only (fp32, no optimizer) + UTTConfig + val metrics.
• <stage>/step_XXXXXXX/metrics.json: validation metrics at that step.

import torch
from cosmos_framework.model.generator.tokenizers.utt import UTT, UTTConfig
ck = torch.load("model.pt", map_location="cpu", weights_only=True)
model = UTT(UTTConfig(**ck["config"])); model.load_state_dict(ck["model"]); model.eval()

Stages: s0a = Sparsh-DINO image stem frozen; s0b = stem fine-tuned (0.1x LR), initialized from s0a. Training data: easyminnn/utt-tactile-prepared v0 pretrain split. Several source datasets carry non-commercial / research-only licenses (see each SOURCE.md there); use these weights accordingly.