Make model card concise and move training details out of the overview

README.md

@@ -19,21 +19,13 @@ tags:
 
 # LingBot-VA — Galaxea A1 Mango-to-Plate EEF — Step 100
 
-This is a full-parameter fine-tune of
-[`robbyant/lingbot-va-base`](https://huggingface.co/robbyant/lingbot-va-base)
-for placing a red mango into a blue plate with a Galaxea A1 robot using
-episode-relative end-effector (EEF) actions. The model jointly predicts video
-latents and robot actions.
+[LingBot-VA](https://huggingface.co/robbyant/lingbot-va-base) fine-tuned for
+red-mango placement onto a blue plate with a Galaxea A1 arm.
+The model predicts robot actions and video latents.
 
-The tokenizer, text encoder, and VAE are inherited unchanged from LingBot-VA
-base; `transformer/` contains the fine-tuned step-100 weights.
+## Data
 
-## Training data
-
-The model was trained on the red-mango-to-blue-plate task from
-[`pengyue-polaron/nyush-galaxea-a1-fruit-placement-eef-v21`](https://huggingface.co/datasets/pengyue-polaron/nyush-galaxea-a1-fruit-placement-eef-v21),
-a LeRobot Dataset v2.1 collection with episode-relative EEF pose actions and a
-continuous normalized gripper command.
+[Training demonstrations](https://huggingface.co/datasets/pengyue-polaron/nyush-galaxea-a1-fruit-placement-eef-v21).
 
 | Field | Value |
 | --- | --- |
@@ -43,27 +35,13 @@ continuous normalized gripper command.
 | Cameras | `front` 480×480 RGB; `wrist` 640×480 RGB |
 | Action | Episode-relative EEF pose + continuous normalized gripper |
 
-## Training
-
-| Setting | Value |
-| --- | --- |
-| Optimizer steps | 100 |
-| Hardware | 2 × NVIDIA A100 |
-| Distributed strategy | Full-parameter FSDP |
-| Precision | bfloat16 |
-| Effective global batch size | 16 |
-| Optimizer | Fused AdamW |
-| Learning rate | 1e-5 with 10-step warmup, then constant |
-| Adam betas / weight decay | (0.9, 0.95) / 0.1 |
-| Objective | Video latent loss + action loss |
+## Files and configuration
 
-Across the 100 optimizer steps, the mean video-latent loss was `0.161962` and
-the mean action loss was `0.020129`.
+`transformer/` contains the step-100 weights. The base tokenizer, text encoder,
+and VAE are included. Use [configs/va_a1_cfg.py](configs/va_a1_cfg.py) for the
+EEF action mapping and normalization.
 
-The eight source action values are mapped to LingBot-VA action channels
-`[0, 1, 2, 3, 4, 5, 6, 28]`; channel 28 carries the gripper command. The exact
-normalization statistics and model settings are included in
-`configs/va_a1_cfg.py`.
+[Training details](TRAINING.md) · [Training summary](training_summary.json)
 
 ## License
 

TRAINING.md

@@ -0,0 +1,22 @@
+# Training
+
+| Setting | Value |
+| --- | --- |
+| Optimizer steps | 100 |
+| Hardware | 2 × NVIDIA A100 |
+| Distributed strategy | Full-parameter FSDP |
+| Precision | bfloat16 |
+| Effective global batch size | 16 |
+| Optimizer | Fused AdamW |
+| Learning rate | 1e-5 with 10-step warmup, then constant |
+| Adam betas / weight decay | (0.9, 0.95) / 0.1 |
+| Objective | Video latent loss + action loss |
+
+Across the 100 optimizer steps, the mean video-latent loss was `0.161962` and
+the mean action loss was `0.020129`.
+
+The eight source action values are mapped to LingBot-VA action channels
+`[0, 1, 2, 3, 4, 5, 6, 28]`; channel 28 carries the gripper command. The exact
+normalization statistics and model settings are included in
+`configs/va_a1_cfg.py`.
+