Update Ant-v5 to v0.14.0 checkpoint
Browse files- README.md +7 -2
- ant-v5/config.json +1 -1
- ant-v5/model.safetensors +1 -1
README.md
CHANGED
|
@@ -79,7 +79,7 @@ Notebook version: [examples/hub_quickstart.ipynb](https://github.com/ccnets-team
|
|
| 79 |
|
| 80 |
| Env | Bundle | Ctx | Return | Norm. | Simple Ref. | Medium Ref. |
|
| 81 |
|---|---|---:|---:|---:|---:|---:|
|
| 82 |
-
| `Ant-v5` | `ant-v5` | 32 |
|
| 83 |
| `HalfCheetah-v5` | `halfcheetah-v5` | 32 | 6816.48±3135.53 | 42.87±19.01 | 43.54 ✗ | 74.83 ✗ |
|
| 84 |
| `Hopper-v5` | `hopper-v5` | 32 | 2713.66±1075.57 | 70.21±28.01 | 42.65 ✓ | 72.91 ✗ |
|
| 85 |
| `Walker2d-v5` | `walker2d-v5` | 32 | 3899.88±706.57 | 56.93±10.32 | 59.51 ✗ | 83.26 ✗ |
|
|
@@ -109,12 +109,17 @@ seeds `0..49`, `max_steps=1000`):
|
|
| 109 |
|
| 110 |
| Env | `kv=16` (0.5×) | `kv=32` (1×) | `kv=64` (2×) | Trend |
|
| 111 |
|---|---:|---:|---:|---|
|
| 112 |
-
| `Ant-v5` |
|
| 113 |
| `HalfCheetah-v5` | 6793.17±2939.17 | 6816.48±3135.53 | 6468.21±3234.51 | ≈ flat |
|
| 114 |
| `Hopper-v5` | 3361.71±103.69 | 2713.66±1075.57 | 992.92±445.63 | shorter ↑, 2× collapses |
|
| 115 |
| `Walker2d-v5` | 3950.09±459.01 | 3899.88±706.57 | 3842.19±718.60 | ≈ flat |
|
| 116 |
| `Humanoid-v5` | 7431.52±2024.95 | 7892.65±1018.11 | 8040.41±38.02 | longer ↑, steadiest |
|
| 117 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 118 |
The `kv=32` column matches the main table. At `kv=64` the rollout attends over
|
| 119 |
more history than the model's 32-token window — positions outside its native
|
| 120 |
range. This stays within the backbone's position capacity (Llama/RoPE,
|
|
|
|
| 79 |
|
| 80 |
| Env | Bundle | Ctx | Return | Norm. | Simple Ref. | Medium Ref. |
|
| 81 |
|---|---|---:|---:|---:|---:|---:|
|
| 82 |
+
| `Ant-v5` | `ant-v5` | 32 | 5021.14±1500.60 | 75.50±22.25 | 59.99 ✓ | 86.54 ✗ |
|
| 83 |
| `HalfCheetah-v5` | `halfcheetah-v5` | 32 | 6816.48±3135.53 | 42.87±19.01 | 43.54 ✗ | 74.83 ✗ |
|
| 84 |
| `Hopper-v5` | `hopper-v5` | 32 | 2713.66±1075.57 | 70.21±28.01 | 42.65 ✓ | 72.91 ✗ |
|
| 85 |
| `Walker2d-v5` | `walker2d-v5` | 32 | 3899.88±706.57 | 56.93±10.32 | 59.51 ✗ | 83.26 ✗ |
|
|
|
|
| 109 |
|
| 110 |
| Env | `kv=16` (0.5×) | `kv=32` (1×) | `kv=64` (2×) | Trend |
|
| 111 |
|---|---:|---:|---:|---|
|
| 112 |
+
| `Ant-v5` | **5467.20±1153.95** | 5021.14±1500.60 | 4957.91±1686.26 | **0.5× recommended** |
|
| 113 |
| `HalfCheetah-v5` | 6793.17±2939.17 | 6816.48±3135.53 | 6468.21±3234.51 | ≈ flat |
|
| 114 |
| `Hopper-v5` | 3361.71±103.69 | 2713.66±1075.57 | 992.92±445.63 | shorter ↑, 2× collapses |
|
| 115 |
| `Walker2d-v5` | 3950.09±459.01 | 3899.88±706.57 | 3842.19±718.60 | ≈ flat |
|
| 116 |
| `Humanoid-v5` | 7431.52±2024.95 | 7892.65±1018.11 | 8040.41±38.02 | longer ↑, steadiest |
|
| 117 |
|
| 118 |
+
For `Ant-v5`, the recommended serving setting is
|
| 119 |
+
`kv_cache_max_len=16`. The model context remains 32; only the retained rollout
|
| 120 |
+
history is shortened. The recommended setting scores `5467.20±1153.95`
|
| 121 |
+
(`82.11±17.11` normalized) over the same 50 seeds.
|
| 122 |
+
|
| 123 |
The `kv=32` column matches the main table. At `kv=64` the rollout attends over
|
| 124 |
more history than the model's 32-token window — positions outside its native
|
| 125 |
range. This stays within the backbone's position capacity (Llama/RoPE,
|
ant-v5/config.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
| 1 |
{
|
| 2 |
"bundle_format_version": 2,
|
| 3 |
-
"package_version": "0.
|
| 4 |
"model_config": {
|
| 5 |
"network_name": "Llama",
|
| 6 |
"num_layers": 4,
|
|
|
|
| 1 |
{
|
| 2 |
"bundle_format_version": 2,
|
| 3 |
+
"package_version": "0.14.0",
|
| 4 |
"model_config": {
|
| 5 |
"network_name": "Llama",
|
| 6 |
"num_layers": 4,
|
ant-v5/model.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 16931180
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:59229ce58c33fd7c18ed66a634c25b5da7fe5d9da0ee1ff53e101f9c36cce53a
|
| 3 |
size 16931180
|