Datasets:
Copy FIDELITY_AUDIT.md from Nickyang/RazorCal
Browse files- FIDELITY_AUDIT.md +22 -0
FIDELITY_AUDIT.md
ADDED
|
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# RazorCal provenance status
|
| 2 |
+
|
| 3 |
+
The calibration corpus is `RazorCal.json`. Its `sub_source` fields record
|
| 4 |
+
attribution information for the included samples.
|
| 5 |
+
|
| 6 |
+
## Available information
|
| 7 |
+
|
| 8 |
+
The released record array has 2,048 samples, seven domains and 15 distinct
|
| 9 |
+
source labels. The source-label inventory is listed in [README.md](README.md).
|
| 10 |
+
These are properties of the released artifact, not proof of upstream provenance.
|
| 11 |
+
|
| 12 |
+
## Source matching
|
| 13 |
+
|
| 14 |
+
The released metadata does not establish an exact upstream record for every
|
| 15 |
+
sample. Labels name the upstream dataset, and for slices taken whole -- such as
|
| 16 |
+
the 160 `nemotron-cascade2-math` records -- they do not identify a position
|
| 17 |
+
within it. No complete source-fidelity claim is made.
|
| 18 |
+
|
| 19 |
+
Structural consistency, source matching, privacy processing and license review
|
| 20 |
+
are separate questions. None can substitute for the others. A source label
|
| 21 |
+
does not establish a license or certify the sample's contents.
|
| 22 |
+
See [DATASHEET.md](DATASHEET.md) and [LICENSE-DATA](LICENSE-DATA).
|