Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,25 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
configs:
|
| 4 |
+
- config_name: default
|
| 5 |
+
data_files:
|
| 6 |
+
- split: train
|
| 7 |
+
path: train.parquet
|
| 8 |
+
- split: dev
|
| 9 |
+
path: dev.parquet
|
| 10 |
+
- split: test
|
| 11 |
+
path: test.parquet
|
| 12 |
+
- split: probe
|
| 13 |
+
path: probe.parquet
|
| 14 |
+
---
|
| 15 |
+
|
| 16 |
+
# monitor_cot_for_hack-coding
|
| 17 |
+
|
| 18 |
+
Training data for the `coding` task of *Reward Misspecification Degrades Chain-of-Thought Monitorability* ([code](https://github.com/Lorenayannnnn/reward-misspecification-degrades-cot-monitorability)), built from
|
| 19 |
+
[newfacade/LeetCodeDataset](https://huggingface.co/datasets/newfacade/LeetCodeDataset) (apache-2.0).
|
| 20 |
+
|
| 21 |
+
- **Data**: LeetCode problems with Medium/Hard difficulty and at least 10 test cases
|
| 22 |
+
- **Entry num per split**: train 1536, dev 198, test 198, probe 50
|
| 23 |
+
- **Probe**: `probe` is a held out set of queries used for generating reward descriptions
|
| 24 |
+
- **True reward**: 1 if the solution passes every test case, else 0
|
| 25 |
+
- **Misspecified reward**: 1 if the solution passes some and fails some test cases, else 0
|