Update README.md

README.md

@@ -6,19 +6,20 @@ tags:
   - robotics
   - world-model
   - video-generation
-  - vision-language-action
+  - vision-action
   - next-forcing
-  - robotwin
 ---
 
-# Next Forcing — RoboTwin Post-trained Checkpoint
+# Next Forcing — Base Checkpoint
 
-Post-trained checkpoint for **Next Forcing: Causal World Modeling with
-Multi-Chunk Prediction**, evaluated on the RoboTwin 2.0 benchmark.
+Base checkpoint for **Next Forcing: Causal World Modeling with Multi-Chunk
+Prediction**. This is the causal video-action backbone used as the
+initialization for post-training.
 
 - 📄 Paper: https://arxiv.org/pdf/2606.11187
 - 🌐 Project page: https://gangweix.github.io/next-forcing/
 - 💻 Code: https://github.com/gangweix/next-forcing
+- 🤖 RoboTwin post-trained model: [`gangweix/next-forcing-posttrain-robotwin`](https://huggingface.co/gangweix/next-forcing-posttrain-robotwin)
 
 ## Model Description
 
@@ -29,31 +30,23 @@ instead of long-range dynamics, especially at high frame rates. Lightweight
 a causal chain during training, providing dense temporal supervision to the
 backbone.
 
-This checkpoint is the RoboTwin post-trained model, built on top of the
-[LingBot-VA](https://github.com/Robbyant/lingbot-va) codebase. It was post-trained
-from [`gangweix/next-forcing-base`](https://huggingface.co/gangweix/next-forcing-base).
+This base checkpoint contains the backbone **without** MCP modules. The MCP
+modules are created at the start of post-training and initialized from the last
+`mcp_blocks_per_depth` backbone blocks (`mcp_init_from_backbone = True`), so the
+post-trained checkpoint is correspondingly larger (6.7B vs 5.1B).
 
 | | |
 | --- | --- |
-| Parameters | 6.7B (BF16) |
+| Parameters | 5.1B (BF16) |
 | Backbone layers | 30 |
-| MCP depths | 3 (`mcp_blocks_per_depth=3`, collect layers `[3, 11, 19, 29]`) |
-| Benchmark | RoboTwin 2.0, 50 bimanual manipulation tasks |
-| Initialized from | [`next-forcing-base`](https://huggingface.co/gangweix/next-forcing-base) (5.1B) |
+| MCP modules | Not included (added during post-training) |
 
-## Results
-
-Average success rate on RoboTwin 2.0:
-
-| Setting | LingBot-VA | **Next Forcing** |
-| --- | ---: | ---: |
-| Clean | 92.9 | **94.1** |
-| Random | 91.5 | **93.5** |
+Built on top of the [LingBot-VA](https://github.com/Robbyant/lingbot-va) codebase.
 
 ## Repository Layout
 
 ```text
-transformer/     Next Forcing backbone with MCP modules (enable_mcp=true)
+transformer/     Causal video-action backbone
 vae/
 text_encoder/
 tokenizer/
@@ -66,27 +59,26 @@ Clone the code and install the dependencies as described in the
 
 ```bash
 python -m pip install "huggingface_hub[cli]"
-hf download gangweix/next-forcing-posttrain-robotwin \
-  --local-dir ./checkpoints/next-forcing-posttrain-robotwin
+hf download gangweix/next-forcing-base \
+  --local-dir ./checkpoints/next-forcing-base
 ```
 
-The evaluation code resolves model subfolders by path, so point
-`NEXT_FORCING_MODEL_PATH` at the **local directory**, not at the Hub repository
-id:
+The training code resolves model subfolders by path, so point
+`NEXT_FORCING_PRETRAINED_MODEL_PATH` at the **local directory**, not at the Hub
+repository id:
 
 ```bash
-export NEXT_FORCING_MODEL_PATH=$PWD/checkpoints/next-forcing-posttrain-robotwin
-export ROBOTWIN_ROOT=/path/to/your/RoboTwin
-
-# Start the inference server on one GPU
-CUDA_VISIBLE_DEVICES=0 bash evaluation/robotwin/launch_server.sh
+export NEXT_FORCING_PRETRAINED_MODEL_PATH=$PWD/checkpoints/next-forcing-base
+export NEXT_FORCING_DATASET_PATH=/path/to/your/dataset
+export NEXT_FORCING_SAVE_ROOT=/path/to/your/output
 
-# In another terminal, evaluate one task for 100 trials
-bash evaluation/robotwin/launch_client.sh /path/to/eval_results adjust_bottle
+NGPU=8 CONFIG_NAME=robotwin_train \
+bash script/run_va_posttrain.sh --init-worker 1
 ```
 
-RoboTwin evaluation requires a working RoboTwin 2.0 installation; see the
-[official guide](https://robotwin-platform.github.io/doc/usage/robotwin-install.html).
+MCP defaults (`wan_va/configs/mcp_train_config.py`): `num_mcp_depths = 3`,
+`mcp_blocks_per_depth = 3`, `mcp_hidden_collect_layers = [3, 11, 19, 29]`,
+`mcp_loss_weights = [0.5, 0.2, 0.1]`.
 
 ## License