Datasets:
Warn about dense-100m v1.3 semantic issue
Browse files
README.md
CHANGED
|
@@ -32,6 +32,8 @@ configs:
|
|
| 32 |
|
| 33 |
# StateBench v1
|
| 34 |
|
|
|
|
|
|
|
| 35 |
### 110 million verified state episodes for pretraining, SFT, and on-policy SDFT
|
| 36 |
|
| 37 |
Train models to **retain**, **edit**, **address**, and **transform** state across long execution traces.
|
|
|
|
| 32 |
|
| 33 |
# StateBench v1
|
| 34 |
|
| 35 |
+
> ⚠️ **Temporary quality notice:** `dense-100m` v1.3 has known prompt/oracle mismatches in six surface dialects. Do not start training on it. The audited v1.4 rebuild is in progress. `default` is unaffected.
|
| 36 |
+
|
| 37 |
### 110 million verified state episodes for pretraining, SFT, and on-policy SDFT
|
| 38 |
|
| 39 |
Train models to **retain**, **edit**, **address**, and **transform** state across long execution traces.
|