Simplify model card for public release

README.md

@@ -8,20 +8,20 @@ tags:
   - lambda0
 ---
 
-# λ₀ Midtrain: G1 SONIC + Revo2
+# λ₀ Midtrain (100k)
 
-This is the **Stage II mid-training checkpoint** for λ₀, intended as the starting point for Stage III robot post-training. It contains the **raw weights at step 100,000**; this checkpoint does not contain EMA weights. It is not a task-specific post-trained policy.
+This is the **Stage II mid-training checkpoint** for λ₀. It provides the starting point for Stage III robot post-training. This release contains the **raw weights after 100,000 training steps**. It does not contain EMA weights or a task-specific post-trained policy.
 
-The model uses a Qwen3.5-2B vision–language backbone and a generative action expert. Its training contract is `lambda0_velocity_v1`: velocity prediction over 50 future action steps, with discrete VLM state conditioning (256 bins). Action-space ID 2, `robot_sonic_revo2`, uses a 76-dimensional action vector (64 SONIC motion coordinates and 6 coordinates per Revo2 hand) and 44-dimensional proprioception. The model configuration also defines other action spaces; select the one matching your robot data contract.
+The model builds on [Qwen3.5-2B](https://huggingface.co/Qwen/Qwen3.5-2B) and learns from robot-aligned human demonstrations.
 
 ## Post-training
 
-The Stage III post-training recipe is being prepared in the [OpenLambda source repository](https://github.com/LambdaHumanoid/openlambda). Once published, set `STAGE2_CHECKPOINT_DIR` to the downloaded model directory, and use `initialization: midtrain` with `midtrain_weight_source: raw`. The example configuration is `lambda0/configs/posttrain/train.json`.
+The Stage III post-training recipe is being prepared in the [OpenLambda source repository](https://github.com/LambdaHumanoid/openlambda). Follow that recipe when it is released to adapt this checkpoint to your robot tasks.
 
-The pretraining and mid-training datasets are not included in this model repository. Robot post-training requires data that matches the selected action-space and coordinate contracts. This checkpoint alone is not a deployment-ready robot policy; evaluate a post-trained policy before any robot use.
+The pretraining and mid-training datasets are not included. This checkpoint is not a deployment-ready robot policy. Evaluate a post-trained policy before robot use.
 
-## Files and provenance
+## Checkpoint
 
-`model.safetensors` contains the trained weights. `config.json`, `state.json`, and the tokenizer and processor files supply the model and training contract needed for loading. The checkpoint originated from λ₀ Stage II training at global step 100,000. The `base_model_name_or_path` field in this release points to the public Qwen3.5-2B model rather than the original training machine path; the weights are unchanged.
+The files include the model weights and the configuration needed to load them. The weight tensor file is unchanged from the Stage II checkpoint at step 100,000.
 
 The λ₀ code and these weights are released under Apache 2.0. The [Qwen3.5-2B base model](https://huggingface.co/Qwen/Qwen3.5-2B) is also available under Apache 2.0.