speedy00/pi05-yPRTRonMbRGk

LingBot-VLA 2.0 — SN80 competition 2 candidate (d12, step 600)

Fine-tune of ApexUltron/pi05-KsYerY3Q8siH (revision 4ac5fea918107a26ba525e3d1b6bdbe615d4ee72), itself a fine-tune of openroboto-ai/lingbot-vla-v2-6b-libero. Architecture and config unchanged (LingbotVlaV2Policy).

Training (single H200, 600 optimizer steps, global batch 64, lr 5e-6 cosine, action-expert-only (VLM and vision encoder frozen), aux heads on, AdamW):

• LIBERO canonical 40-task demonstrations (official dataset), plus
• teacher-rollout demonstrations on LIBERO-PRO swap layouts (libero_spatial_swap, libero_object_swap): successful closed-loop rollouts collected from the base model above and from sakura-bullish/pi05-YhxFYW92tXmq, kept only on exact task success with replay verification, oversampled x9. Norm stats: the official base's training/norm_stats.json. No evaluator init states or evaluator seeds were used for data.