Datasets:
Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,27 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
task_categories:
|
| 4 |
+
- text-generation
|
| 5 |
+
language:
|
| 6 |
+
- en
|
| 7 |
+
tags:
|
| 8 |
+
- sql
|
| 9 |
+
- mechanistic-interpretability
|
| 10 |
+
- correlated-subquery
|
| 11 |
+
---
|
| 12 |
+
|
| 13 |
+
# cs8-rand
|
| 14 |
+
|
| 15 |
+
CS8 rand dataset — random inner alias assignment, negative control for mechanistic interpretability study
|
| 16 |
+
|
| 17 |
+
## Structure
|
| 18 |
+
- `train.jsonl` — 12,000 training examples (20 aliases × 600)
|
| 19 |
+
- `val.jsonl` — 12,000 validation examples (20 aliases × 600)
|
| 20 |
+
- `alias_vocab.json` — 20 aliases with T1/T2 tokenization group labels
|
| 21 |
+
|
| 22 |
+
## Alias Groups
|
| 23 |
+
- **T1 (single-token)**: emp, inv, txn, mgr, ord, prod, cust, dept, acct, sale
|
| 24 |
+
- **T2 (two-token, generic first subtoken)**: shp, whs, rgn, lne, empl, ordr, catg, budg, schd, srvc
|
| 25 |
+
|
| 26 |
+
## Citation
|
| 27 |
+
Inside-Out: A Circuit-Level Analysis of Table-Alias Scoping in Nested SQL Generation
|