kmy17518/gr00t-n1.7-b1k-multitask

GR00T N1.7 — BEHAVIOR-1K 2026 multi-task fine-tune (eval snapshots)

Serving-only snapshots (bf16 model weights, no optimizer state) of a multi-task fine-tune of nvidia/GR00T-N1.7-3B on the BEHAVIOR-1K 2026 challenge dataset (100 tasks, 20,000 episodes, LeRobot v3.0), via the wensi-ai/Isaac-GR00T fork.

• R1Pro embodiment (examples/b1k/r1pro.py, NEW_EMBODIMENT tag); projector + diffusion action head tuned, backbone frozen.
• Effective batch 2048 (global 1024 x grad-accum 2), lr 1e-4, cosine over 2M steps with 10k warmup — not annealed: the schedule was still near peak LR at the final snapshot. Prefer evaluating several late snapshots over assuming the last is best.
• checkpoint-<N> = optimizer step N; snapshots every 5k steps to 80k, every 2k for 100k-200k, every 1k after; final = checkpoint-238000 (~487M samples).
• The policy is conditioned on the dataset's task-description strings (e.g. turning_on_radio) — serve with the matching string: python scripts/b1k/serve_b1k.py --model-path checkpoint-<N> --modality-config-path examples/b1k/r1pro.py

Base model use is subject to the NVIDIA license terms of GR00T N1.7 / Cosmos.