galaxea-a1-nyush-relation-approach-cup-teacher: README.md

galaxea-a1-nyush-relation-approach-cup-teacher/README.md

@@ -0,0 +1,13 @@
+# galaxea-a1-nyush-relation-approach-cup-teacher — LingBot-VA diffusion teacher, single task "approach the cup"
+
+Post-trained 238 steps x 16 samples (= 3808 samples ~= 76 epochs of the 50 episodes, lr 1e-5) from `robbyant/lingbot-va-base`.
+
+Dataset: `pengyue-polaron/nyush-galaxea-a1-relation`, ONLY the 50 episodes of `approach the cup` (single-task isolation: one task in training,
+same per-task epoch budget as the 5-task models). Whole episodes, 2 nodes x 8 H100 FSDP. Prompt (use verbatim): `approach the cup`.
+
+Transformer only (19 GB, `attn_mode: torch`); take `vae/`, `text_encoder/`, `tokenizer/` from `robbyant/lingbot-va-base`.
+Deploy: config `galaxea_deploy` with `VA_CKPT_DIR=<dir>`; students `VIDEO_NFE=1 ACTION_NFE=1 VIDEO_GS=1`, teachers `VIDEO_NFE=20 ACTION_NFE=50 VIDEO_GS=5`.
+Code, server launch, client protocol: https://github.com/SeanAlpaca818/Diffusion2One/tree/galaxea-a1/evaluation/galaxea
+
+Files: `transformer/config.json`, `transformer/diffusion_pytorch_model.safetensors`, `lingbot_norm_stat.json`, `eef.json`
+(full-dataset normalisation, identical to `galaxea-a1-nyush-relation-teacher/`; deploy MUST use them).