Document 8k PG-OPD scaling data
Browse files
README.md
CHANGED
|
@@ -30,6 +30,8 @@ not interchangeable.
|
|
| 30 |
|
| 31 |
`Qwen35-4B/pg_opd.parquet` contains 2,394 rows:
|
| 32 |
|
|
|
|
|
|
|
| 33 |
- 1,197 verified evidence-privileged rows;
|
| 34 |
- 1,197 native fallback rows;
|
| 35 |
- evidence was selected from the 4B Stage-A filtering run.
|
|
|
|
| 30 |
|
| 31 |
`Qwen35-4B/pg_opd.parquet` contains 2,394 rows:
|
| 32 |
|
| 33 |
+
`Qwen35-4B/pg_opd_8k_forward_kl.parquet` contains 8,000 rows for the forward-KL data-scaling ablation (4,000 verified privilege rows and 4,000 native-control rows). This is an auxiliary scaling artifact; the final reported recipe uses JSD (beta=0.5).
|
| 34 |
+
|
| 35 |
- 1,197 verified evidence-privileged rows;
|
| 36 |
- 1,197 native fallback rows;
|
| 37 |
- evidence was selected from the 4B Stage-A filtering run.
|