asura625/pi05-XTDkBhqoLABG

LingBot-VLA 2.0 — LIBERO fine-tune, step 400 (OpenRoboto sim/2)

Canonical parent: openroboto-ai/lingbot-vla-v2-6b-libero @ ce6a322157acc7a03d0ca71bb84423c7f2e124d7 (the season base: LingBot-VLA 2.0 already LIBERO-tuned, itself a fine-tune of robbyant/lingbot-vla-v2-6b@11c703bf). Method: full-parameter fine-tune with the vendor trainer (Robbyant/lingbot-vla-v2 @ 951475ae), global batch 64, muon optimizer, cosine schedule, 400 optimizer steps on 4x A100-80GB. Data: the official pool openroboto-ai/libero-datapool @ 2c065c6d (1693 human demos) PLUS self-generated scripted-expert demonstrations produced in MuJoCo on the PUBLIC LIBERO-Pro swap task definitions (HF zhouxueyang/LIBERO-Pro: libero_object_swap and libero_spatial_swap), which make up roughly half of the sampled training frames. Lineage: continued full-parameter fine-tuning from our own earlier submission (asura625/pi05-XTDkBhqoLABG @ bb68e541), itself a fine-tune of the season base. No other miner's weights or data were used. Normalization stats, camera mapping and 7-dim action semantics follow the public reference fine-tune unchanged (training/norm_stats.json).