Downloads · 30 days
7
25% of all-time downloads
easyminnn/flevel-k25-vlm-60k
flevel-k25-vlm-60k is a robotics model from easyminnn. 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.
Inference-only checkpoint (weights + config; optimizer/scheduler/RNG state stripped). Fine-tuned from nvidia/GR00T-N1.5-3B on RoboCasa MG 300 (7,200 episodes, 3 cameras 256x256, 20 fps) with VLM speed-ratio labels bak…
Downloads · 30 days
7
25% of all-time downloads
All-time downloads
28
Public
Parameters
3B
8.6 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors8.6 GB · 100%
How the weights are stored.
BF161.7B · 55%
From the Hugging Face model README
vlm, 60k stepsInference-only checkpoint (weights + config; optimizer/scheduler/RNG state stripped).
Fine-tuned from nvidia/GR00T-N1.5-3B on RoboCasa MG 300 (7,200 episodes, 3 cameras 256x256, 20 fps)
with VLM speed-ratio labels baked into the action tensor (ratio_label, dims 12:14).
| Decoders | level_ks=(1, 2.5): 1x stream (16 rows) + 2.5x compressed stream ([2,3,2,3,2,3] -> 6 rows / 15 steps) |
| conf target | conf (VLM speed-ratio confidence only) |
| conf head | scalar regression, MSE, ratio_loss_coef=0.1, readout detached from state encoder |
| Training | 60,000 steps, global batch 64 (32 x 2 GPU), AdamW lr 1e-4 cosine, warmup 0.05, seed 42, bf16, 2x H200 |
| Labels | prehj/robocasa-ratio-labels-contact |
| Code | TTKKWAN/F_level branch hj-two-decoder-conf @ 29a1090 (papers/reproducing/FLARE/scripts/train_flevel_ratio.py --mode ratio_current --decoder k25 --label-variant vlm) |
tau (conf_threshold) is an eval-time knob: pred >= tau selects the compressed stream. The stored value 0.5 is a default, not a trained constant.
Serve with serve_flevel_robocasa.py --model-path <this repo> --mode ratio_current; the client executes the returned rows as-is (no re-merging).