Add 8k PG-OPD data metadata
Browse files
Qwen35-4B/pg_opd_8k_forward_kl.metadata.json
ADDED
|
@@ -0,0 +1,23 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"method": "PG-OPD data-scaling ablation",
|
| 3 |
+
"output": "pg_opd_8k_forward_kl.parquet",
|
| 4 |
+
"source": "P1 multitask objective pool, p0-fixed objective data",
|
| 5 |
+
"selected_rows": 8000,
|
| 6 |
+
"verified_evidence_rows": 4000,
|
| 7 |
+
"fallback_native_rows": 4000,
|
| 8 |
+
"privilege_fraction": 0.5,
|
| 9 |
+
"divergence": "forward KL",
|
| 10 |
+
"note": "This is the 8k scaling-ablation dataset; it is not the final JSD(beta=0.5) recipe.",
|
| 11 |
+
"source_counts": {
|
| 12 |
+
"M4-COMICS_Dialogue": 2988,
|
| 13 |
+
"Mantis-imagecode": 2117,
|
| 14 |
+
"M4-MIT-States_StateCoherence": 850,
|
| 15 |
+
"M4-OCR-VQA": 771,
|
| 16 |
+
"M4-MIT-States_PropertyCoherence": 724,
|
| 17 |
+
"M4-DocVQA": 550
|
| 18 |
+
},
|
| 19 |
+
"teacher_input_contains_gt": false,
|
| 20 |
+
"teacher_input_contains_support_indices": false,
|
| 21 |
+
"image_path_policy": "Repository-relative paths under ../../data/images/{m4,mantis}/",
|
| 22 |
+
"public_release_note": "The historical parquet contains 3 evidence strings with residual bracketed locator-like text; they are preserved for exact training provenance and should not be interpreted as a clean locator-free ablation."
|
| 23 |
+
}
|