Math-augmented RL v2: add leader log-probabilities for joint constrained training
Browse files
candidates/budgie-alignment-v2/information-rl-v2-math-augmented/bootstrap/math_encoded_rollouts.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
candidates/budgie-alignment-v2/information-rl-v2-math-augmented/bootstrap/math_encoding_manifest.json
ADDED
|
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"groups": 72,
|
| 3 |
+
"encoded": 216,
|
| 4 |
+
"mean_tokens": 443.5925925925926,
|
| 5 |
+
"mean_response_tokens": 307.73148148148147,
|
| 6 |
+
"seconds": 3.2618279457092285
|
| 7 |
+
}
|