Downloads · 30 days
0
Demomasterlqx/VTLA-RL-sft-lora-xarm-no-adverb-pirl-2026-9-16
VTLA-RL-sft-lora-xarm-no-adverb-pirl-2026-9-16 is a reinforcement learning model from Demomasterlqx. Use it for the reinforcement learning task on the model card, and read the license before you ship it in a product. It is set up for openpi.
Base model: Demomasterlqx/VTLA-RL-sft-lora-xarm-no-adverb at main. Actual initialization: globalstep30000/model, SFT 30k EMA merged; immutable base revision dfbbc4c2e72732bd770c4aa77d77a5ba2114f416. See basemodel.json…
Downloads · 30 days
0
Access
Public
Updated Sep 17, 2026
Repo size
61.2 GB
Likes
0
Public
Click a slice to open those files.
.distcp60.4 GB · 99%
From the Hugging Face model README
Base model: Demomasterlqx/VTLA-RL-sft-lora-xarm-no-adverb at main.
Actual initialization: global_step_30000/model, SFT 30k EMA merged; immutable base revision dfbbc4c2e72732bd770c4aa77d77a5ba2114f416. See base_model.json for hashes.
Training: 300 RLinf global iterations, GPU 0-7, 128 environments, micro/global batch 32/256, action-expert LoRA rank32 and value head, no-state Pi0.5 TacField. Task6 cookie, xy_1cm and physics randomization, gradient checkpointing disabled. A snapshot is saved with each checkpoint.
Every 10 iterations:
resume_checkpoints/step_N: complete RLinf actor DCP model, optimizer, scheduler and RNG, trainable weights, training configuration and integrity audit. To resume, download this directory as a local global_step_N directory and set runner.resume_dir to it; RLinf parses the step from that local name. Use the same audited SFT base and topology. Simulator trajectories are reinitialized on recovery.bundle/step_N: T2-VLA loadable LoRA bundle; use the SFT base above plus --lora-bundle /local/bundle/step_N, config pi05_lora_tacfield_realworld_replayed_task820_firm_mixed_no_state. The bundle excludes the RL value head.W&B: https://wandb.ai/183842220-hkust/tabero-rlinf/runs/0882a381
Uploads are asynchronous; individual steps are only marked verified after remote file-set, size and content-hash verification. Cookie training was stopped by the user; its latest saved checkpoint is step 20. The new Vitasoy experiment is documented in vitasoy/README.md and uses separate artifact directories.
Deployment runtime: use peft==0.20.0 and transformers==4.53.2. The T2-VLA loader strictly checks the PEFT version recorded in each bundle; do not synchronize back to PEFT 0.19.1. Step 10 has passed actual T2-VLA CPU bundle loading and reconstructed tensor-hash verification.