Downloads · 30 days
8
19% of all-time downloads
prehj/allex-atq-conf-gated-h24-bs256-30k
allex-atq-conf-gated-h24-bs256-30k is a robotics model from prehj. Use it for the robotics task on the model card, and read the license before you ship it in a product. The card lists the license as other.
GR00T N1.7 (Cosmos-Reason2-2B backbone) fine-tuned on ALLEX bimanual manipulation with the ATQ label-gated expert grid: four action decoders that differ in playback speed and horizon, plus a confidence head that decid…
Downloads · 30 days
8
19% of all-time downloads
All-time downloads
42
Public
Parameters
3.3B
7.1 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors7.1 GB · 100%
How the weights are stored.
BF163.6B · 100%
From the Hugging Face model README
GR00T N1.7 (Cosmos-Reason2-2B backbone) fine-tuned on ALLEX bimanual manipulation with the ATQ label-gated expert grid: four action decoders that differ in playback speed and horizon, plus a confidence head that decides which group of experts is allowed at inference.
Trained 2026-09-15/16, 30,000 steps, global batch 256, 2x H200.
action_horizon = 24)| idx | name | span | rows | speed |
|---|---|---|---|---|
| 0 | main | 24 | 24 | 1x, fine |
| 1 | mc | 24 | 12 | 2x, compressed |
| 2 | sc | 12 | 6 | 2x, compressed |
| 3 | sf | 12 | 12 | 1x, fine |
A compressed row is the block's LAST source step, not the sum (moe_merge: "last"). GR00T's action space is not a per-step increment: ABSOLUTE steps are
joint targets, and RELATIVE steps are measured from one fixed anchor (the last
observed state) for the whole chunk. Summing such steps inflates the target by
roughly the block size. Anything consuming a compressed row must execute it as
the pose to reach after K source steps, not as a delta to be split over K steps.
The gate is inference-only. In training all four experts are supervised every
step and the label is used for one thing: a regression head that predicts the
chunk's confidence (action.ratio_label = [conf, valid]). At inference the
predicted conf is thresholded (conf_threshold, default 0.5) to mask a group,
and the router picks a horizon inside that group.
frontier_demo/merged/allex/v1_v2 — 160 episodes, 447,162 frames, 30 fps.allex_rrc_eef_relative_no_waist_egostereo_conf — wrists RELATIVE (xyz + rot6d), hands/neck ABSOLUTE, ego stereo cameras, no waist.| base | nvidia/GR00T-N1.7-3B |
| steps | 30,000 |
| global batch | 256 (2 GPUs x 128) |
| precision | bf16, DeepSpeed ZeRO-2, gradient checkpointing |
| action horizon | 24 |
| experts | 4, moe_speed 2.0, moe_merge last, label-gated |
| conf loss coef | 0.1 |
| router warmup | 5,000 steps |
| state dropout | 0.3 |
| runtime | 11h05m (39,927 s), 0.75 steps/s |
Final: train_loss 0.0395; at the end of training loss/main 0.0104,
loss/mc 0.0128, loss/sc 0.0104, loss/sf 0.0103, loss/conf 0.0031,
conf MAE 0.042 against labels with sd 0.128. No NaN/Inf and no loss spikes over
the full run.
Router note. After the 5,000-step warmup the four experts settle at ~0.25 each (max−min ≈ 0.006), because their losses end up within 0.003 of each other. The router therefore carries little preference of its own; the conf gate is what separates fine from compressed at inference. Worth measuring before relying on the router's ranking.
Inference weights only. Optimizer/scheduler/RNG state and the DeepSpeed shards are not included, so this repo cannot resume training.
config.json
model-0000{1,2}-of-00002.safetensors
model.safetensors.index.json
processor/{processor_config.json,embodiment_id.json,statistics.json}
experiment_cfg/ # training config as run, incl. moe_merge and the init manifest
Code: the hj-atq-conf-gated branch of a GR00T N1.7 fork (ATQ ported from
GR00T-action-quantization, jimin-dev-label-gated).