flux3-chess-lora

LoRA adapters for FLUX 3 Action SO-101 (black-forest-labs/flux-3-action-so101 @ c9e13b2aca6a0a472b3ea03fd90cff32d3e85849), trained with LeRobot's examples/flux3/lora.json recipe on one SO-101 chess move (horse c3 → d5).

• Dataset: niru-5/chess_dataset_c3_6cm_to_d5_10cm_20261017_145741 @ 9b6b38719ab08509bdf4ce169a4de028573199de; last 0.2 of episodes held out (episodes 24–30); 2500 steps at batch 8 per GPU × accum 1 (1 × H100), seed 42.
• Repo root = step 2500 (lowest in-training eval loss so far), EMA weights. Every finished checkpoint is under checkpoints/<step>/ with pretrained_model/ (raw) and pretrained_model_ema/ (EMA). Optimizer state is not uploaded.
• Encoders: black-forest-labs/flux-3-action-base @ 62878e2925e59b7a89ec14463ce89932624c490d (video VAE, text encoder).
• Use is subject to the license of the base FLUX 3 Action models.

Checkpoints

Eval loss: LeRobot's in-training flow-matching loss on the held-out episodes, computed on the raw weights with random noise (differences of a few percent are noise). Offline MAE: teacher-forced action error of the EMA weights on the same held-out episodes (88 windows, stride 32, one sample each): predicted vs recorded commands over the executed horizon (first 32 of 42 steps); arm = mean of the 5 arm joints in degrees, gripper in %. Lower is better. Moving windows (same definition as the team's motion_windows.py): only windows where an arm command moves ≥ 10° from the last command; error on those, share of joint-windows (recorded end move ≥ 5°) where the model moves the same way by ≥ 10 % of it, and median coverage (1 = all of the motion, 0 = none, < 0 = wrong way). Hold the last command scores well on whole windows because most of each episode is slow or still; a model should beat it on moving windows and point the right way.

step	eval loss (raw)	offline MAE arm ° (EMA)	gripper %	moving windows: arm °	right direction	median coverage	folder
125	0.1180	9.97	4.79	–	–	–	checkpoints/000125/
250	0.1123	8.80	4.56	–	–	–	checkpoints/000250/
375	0.1109	7.66	4.43	–	–	–	checkpoints/000375/
500	0.1078	6.72	4.24	–	–	–	checkpoints/000500/
1000	0.0990	4.59	3.89	–	–	–	checkpoints/001000/
1500	0.0957	3.31	3.37	–	–	–	checkpoints/001500/
2000	0.0947	2.96	3.01	–	–	–	checkpoints/002000/
2500	0.0927	2.73	2.83	3.79	90%	0.86	checkpoints/002500/ root
base so101	–	12.49	5.82	–	–	–	black-forest-labs/flux-3-action-so101
hold the last command	–	4.05	2.08	7.31	0%	0.00	baseline

Per joint, root checkpoint (step 2500) vs base

joint	base so101	step 2500
shoulder_pan	10.70	1.30
shoulder_lift	14.92	4.15
elbow_flex	16.14	4.08
wrist_flex	12.91	2.96
wrist_roll	7.77	1.18
gripper	5.82	2.83

Load

from peft import PeftConfig, PeftModel
from lerobot.policies.flux3.modeling_flux3 import Flux3Policy

repo, sub = "Gawtam/flux3-chess-lora", None  # or sub="checkpoints/000250/pretrained_model_ema"
cfg = PeftConfig.from_pretrained(repo, subfolder=sub)
policy = PeftModel.from_pretrained(Flux3Policy.from_pretrained(cfg.base_model_name_or_path), repo, subfolder=sub)

Run: lora_c3d5_h100.