Downloads · 30 days
0
hmkang/frontier_demo_gr00t_n17_allex_v1v6_enc
frontier_demo_gr00t_n17_allex_v1v6_enc is a machine learning model from hmkang. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Files here are AES-256-CBC encrypted (openssl enc -aes-256-cbc -pbkdf2 -iter 200000 -salt), one .enc per original file, no compression. Frontier-trained weights are never published in the clear; the repo is public so…
Downloads · 30 days
0
Access
Public
Updated Sep 23, 2026
Repo size
111 GB
Likes
0
Public
Click a slice to open those files.
.enc111 GB · 100%
From the Hugging Face model README
Files here are AES-256-CBC encrypted (openssl enc -aes-256-cbc -pbkdf2 -iter 200000 -salt),
one .enc per original file, no compression. Frontier-trained weights are never published in
the clear; the repo is public so the team can find it, the contents need the key.
| folder | steps | plate | normalization | aug |
|---|---|---|---|---|
A_eef_relative_q0199_h24_30k_2gpu/ | 30k | 2 GPU, global batch 32 | 1 / 99 (1/99 overlay) | WAM mim-only port |
A_eef_relative_q001q999_h24_cjdavid_bs64_50k_2gpu/ | 50k | 2 GPU, global batch 64 | 0.1 / 99.9 (--use_eef_relative_percentile_999) | mim-only port + DAVID colour jitter |
Both: GR00T N1.7-3B on allex merged v1..v6 (480 episodes), 2-view ego stereo at 256, state dropout 0.3, action horizon 24, arms EEF-RELATIVE rot6d, hands + neck ABSOLUTE, no waist, 50-D.
…_cjdavid_bs64_50k_2gpu) — what the seniors' 2026-09-20 directive changed--save_steps 10000 → rungs 10k / 20k / 30k / 40k / 50k, all five published here.--gradient_accumulation_steps 1. per_device = global_batch_size // num_gpus (experiment.py:215 — GA is not folded in), so this is
per-device 32, double the 30k run's regime.--warmup_ratio 0.1 (5k of 50k), --lr_scheduler_type cosine_with_min_lr --min_lr_rate 0.1 → the floor is 1e-5.albumentations.ColorJitter after crop/rotate via a replay, so
the geometry draw is unchanged and only the photometry differs. (GR00T_MIMONLY_AUG=1.)--use_eef_relative_percentile_999, reading david's original
meta/relative_stats.json (which records _percentiles=[0.001, 0.999]).
★This is the opposite choice from the 30k A folder above, which deliberately read a 1/99
overlay. The field NAMES are q01/q99 in both cases — gr00t/data/stats.py writes the
0.1/99.9 values into fields named q01/q99 when use_q999 is set. _percentiles is the
ground truth, not the field name.
★And the split is asymmetric in both runs: the ARMS read relative_stats.json (0.1/99.9
here), while hands + neck read stats.json, which carries no marker and is therefore 1/99.Both runs here were trained with a defect that is now understood and fixed, but neither checkpoint here carries the fix — they are what they are, and this note says exactly what that means.
The action is zero-padded from 50 real dims to max_action_dim = 132, so 82 columns
(62%) are filler. action_mask removes those columns from the loss only — that is its
sole use in the model (gr00t_n1d7.py:405-407). Meanwhile the flow-matching noise is drawn
over all 132 columns (:318), and the inference ODE re-reads and updates all 132 through
action_encoder, a dense Linear, at every Euler step (:610, :670).
Those columns are never supervised. In training their distribution is (1-t)·noise — it
shrinks to zero — and the velocity that produces that shrinkage (= -noise) is exactly the
term the loss mask deletes. So at inference they do not shrink: at the last ODE step the
encoder sees std ≈ 1 where training had std ≈ 0, and that step is the one that sets the
emitted action (num_inference_timesteps = 4).
The model did not learn to ignore them. Measured on checkpoint-50000:
| encoder input-weight column norm | value | vs. its untrained initialisation |
|---|---|---|
| the 50 real dims | 2.341 | 3.0× (trained) |
| the 82 padded dims | 0.721 | 0.92× (essentially untouched) |
GR00T_ACTION_PAD_NOISE_MASK (default OFF, purely additive, +102/−0) masks the noise on the
padded columns in training and holds them at zero through the ODE. Code and reasoning:
hyeonmo/frontier-gr00t on ismty0805/Isaac-GR00T-AlinVLA (see its README).
A run with the flag ON is training as of 2026-09-20 and will be published as a third folder here. The two folders above differ from it only in that env var.
GR00T_ACTION_PAD_NOISE_MASK=1 when loading them. Their config.json carries
no action_valid_dim, and masked inference on such a checkpoint raises rather than
guessing the real width. That refusal is deliberate.None, and the flag resolves OFF — verified against checkpoint-50000.Optimizer state (global_step*), RNG (*.pth), scheduler.pt, latest, training_args.bin
and zero_to_fp32.py are excluded. Each checkpoint publishes the two safetensors shards,
statistics.json, and experiment_cfg/final_processor_config.json.
Every .enc here was verified by decrypting it back and comparing sha256 against the source
before upload.
for f in checkpoint-XXXXX/*.enc checkpoint-XXXXX/experiment_cfg/*.enc; do
openssl enc -d -aes-256-cbc -pbkdf2 -iter 200000 -pass file:/path/to/key.txt -in "$f" -out "${f%.enc}"
done