Replace Walker2d-v5 with stable bundle 25

Bundle format 2, causal-gpt-rl package 0.14.0. Gymnasium 1.2.3, MuJoCo 3.2.3, torch 2.8.0+cu129, seeds 0..49, max_steps=1000, BOS discard. Main score uses kv_cache_max_len=32; the same protocol was used for kv=16 and kv=64.

README.md

@@ -82,7 +82,7 @@ Notebook version: [examples/hub_quickstart.ipynb](https://github.com/ccnets-team
 | `Ant-v5` | `ant-v5` | 32 | 5126.03±1568.43 | 77.05±23.26 | 59.99 ✓ | 86.54 ✗ |
 | `HalfCheetah-v5` | `halfcheetah-v5` | 32 | 6816.48±3135.53 | 42.87±19.01 | 43.54 ✗ | 74.83 ✗ |
 | `Hopper-v5` | `hopper-v5` | 32 | 3199.65±21.74 | 82.87±0.57 | 42.65 ✓ | 72.91 ✓ |
-| `Walker2d-v5` | `walker2d-v5` | 32 | 3899.88±706.57 | 56.93±10.32 | 59.51 ✗ | 83.26 ✗ |
+| `Walker2d-v5` | `walker2d-v5` | 32 | 4122.68±299.84 | 60.19±4.38 | 59.51 ✓ | 83.26 ✗ |
 | `Humanoid-v5` | `humanoid-v5` | 32 | 7892.65±1018.11 | 91.63±11.99 | 63.29 ✓ | 81.30 ✓ |
 
 Training data is expert-free: bundles are trained using Minari simple and medium datasets only; expert trajectories are not used for training.
@@ -112,7 +112,7 @@ seeds `0..49`, `max_steps=1000`):
 | `Ant-v5` | 5109.97±1711.18 | 5126.03±1568.43 | **5256.33±1429.58** | **2× highest mean (≈)** |
 | `HalfCheetah-v5` | 6793.17±2939.17 | 6816.48±3135.53 | 6468.21±3234.51 | ≈ flat |
 | `Hopper-v5` | 3189.53±22.58 | **3199.65±21.74** | 3190.09±23.75 | **≈ flat and stable** |
-| `Walker2d-v5` | 3950.09±459.01 | 3899.88±706.57 | 3842.19±718.60 | ≈ flat |
+| `Walker2d-v5` | 4021.00±573.17 | **4122.68±299.84** | 2659.11±1297.61 | **1× best; 2× collapses** |
 | `Humanoid-v5` | 7431.52±2024.95 | 7892.65±1018.11 | 8040.41±38.02 | longer ↑, steadiest |
 
 For `Ant-v5`, the recommended serving setting is
@@ -130,7 +130,7 @@ across 0.5×/1×/2× is within run-to-run noise (`Trend` marks these `≈`).
 The refreshed `Hopper-v5` bundle is especially stable: all three settings average
 about 3,190–3,200 return with all 50 episodes reaching 1,000 steps. `Humanoid-v5`
 is best at 2× (highest return and its steadiest — std 38 across all 50 episodes).
-The context window (1×) remains a safe default.
+The refreshed `Walker2d-v5` bundle is strongest at 1×: it reaches 1,000 steps in 48/50 episodes, while 2× falls to 2,659 return and 15/50 full episodes. The context window (1×) remains a safe default.
 
 Evaluation runtime — every row above is measured on this one:
 

walker2d-v5/config.json

@@ -1,14 +1,16 @@
 {
   "bundle_format_version": 2,
-  "package_version": "0.2.2",
+  "package_version": "0.14.0",
   "model_config": {
     "network_name": "Llama",
     "num_layers": 4,
-    "d_model": 224,
-    "num_heads": 7,
+    "d_model": 256,
+    "num_heads": 8,
     "dropout": 0.05,
+    "use_eos": true,
+    "use_bos_action_gate": true,
     "rope_theta": 1000.0,
-    "intermediate_size": 896,
+    "intermediate_size": 1024,
     "max_position_embeddings": 256,
     "context_length": 32
   },
@@ -83,9 +85,81 @@
     }
   ],
   "context_length": 32,
+  "requires_capabilities": [],
+  "state_container": {
+    "type": "Box",
+    "shape": [
+      17
+    ],
+    "dtype": "float64",
+    "low": [
+      "-inf",
+      "-inf",
+      "-inf",
+      "-inf",
+      "-inf",
+      "-inf",
+      "-inf",
+      "-inf",
+      "-inf",
+      "-inf",
+      "-inf",
+      "-inf",
+      "-inf",
+      "-inf",
+      "-inf",
+      "-inf",
+      "-inf"
+    ],
+    "high": [
+      "inf",
+      "inf",
+      "inf",
+      "inf",
+      "inf",
+      "inf",
+      "inf",
+      "inf",
+      "inf",
+      "inf",
+      "inf",
+      "inf",
+      "inf",
+      "inf",
+      "inf",
+      "inf",
+      "inf"
+    ]
+  },
+  "action_container": {
+    "type": "Box",
+    "shape": [
+      6
+    ],
+    "dtype": "float32",
+    "low": [
+      -1.0,
+      -1.0,
+      -1.0,
+      -1.0,
+      -1.0,
+      -1.0
+    ],
+    "high": [
+      1.0,
+      1.0,
+      1.0,
+      1.0,
+      1.0,
+      1.0
+    ]
+  },
   "state_normalization": {
     "embedded": true,
     "legacy_sidecar": false
   },
-  "env_id": "Walker2d-v5"
+  "env_id": "Walker2d-v5",
+  "serving": {
+    "bos_cache_mode": "discard"
+  }
 }

walker2d-v5/model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:868aa869244620a14d5a8a33519e2723fcf1566622afa390f9b336d8a216c284
-size 12894456
+oid sha256:765902feb9fe9c4a6f675375504df1bd0a6ee981f720cc890b9439c5c0dd8972
+size 16834124