Downloads · 30 days
0
Riasok/gr00t-n15-libero-goal-ppo-fp32head-rankhook
gr00t-n15-libero-goal-ppo-fp32head-rankhook is a machine learning model from Riasok. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
Same recipe as Riasok/gr00t-n15-libero-spatial-ppo-fp32head-rankhook: RLinf FSDP1 on 4xH100, 64 envs x 8 rollout epochs, updateepoch 4, lr 5e-6 / value 1e-4, AdamW wd 0.01, clip 1.0, flow-SDE noise 0.5. Start: RLinf/R…
Downloads · 30 days
0
Access
Public
Updated Sep 24, 2026
Repo size
72.8 GB
Likes
0
Public
Click a slice to open those files.
.distcp48.6 GB · 67%
From the Hugging Face model README
Same recipe as Riasok/gr00t-n15-libero-spatial-ppo-fp32head-rankhook: RLinf FSDP1 on 4xH100, 64 envs x 8 rollout epochs,
update_epoch 4, lr 5e-6 / value 1e-4, AdamW wd 0.01, clip 1.0, flow-SDE noise 0.5. Start: RLinf/RLinf-Gr00t-SFT-Goal (N1.5).
Action head (322 tensors) in fp32 masters + fp32 optimizer state; backbone (585 tensors) bf16; bf16 compute.
global_step_{1,3,5}/actor/model_state_dict/full_weights.pt : full state dict (action_head.* fp32, rest bf16)global_step_{1,3,5}/actor/dcp_checkpoint/ : FSDP distributed checkpoint (model + optimizer) for resumerank_hook/captures/optimizer_step_{001,005,020}.pt, rank_hook/analysis/, rank_hook/code/| global step | 1 | 2 | 3 | 4 | 5 |
|---|---|---|---|---|---|
| success | 0.465 | 0.580 | 0.576 | 0.658 | 0.732 |
| stage | W2 timestep (emb 31) | AdaNorm shift (L0/5/10/15) |
|---|---|---|
| PPO gradient | 2 / 3 / 2 | 1-2 |
| Adam direction | 13 / 3 / 3 | 20-27 / 3 / 2-3 |
| fp32 cumulative update | 13 / 6 / 5 | 20-27 / 6 / 4-5 |
| bf16 endpoint difference | 900 / 686 / 530 | 920 / 700-777 / 574-634 |
True update is rank <= 27 (< 2% of full rank); the bf16 checkpoint diff is 35-60% of full rank = bf16 quantisation noise. AdamW-formula prediction vs actual fp32 delta: cosine >= 0.9999997.