README.md

---
license: mit
task_categories:
- text-generation
language:
- en
tags:
- sql
- mechanistic-interpretability
- correlated-subquery
---

# cs8-rand

CS8 rand dataset — random inner alias assignment, negative control for mechanistic interpretability study

## Structure
- `train.jsonl` — 12,000 training examples (20 aliases × 600)
- `val.jsonl`   — 12,000 validation examples (20 aliases × 600)
- `alias_vocab.json` — 20 aliases with T1/T2 tokenization group labels

## Alias Groups
- **T1 (single-token)**: emp, inv, txn, mgr, ord, prod, cust, dept, acct, sale
- **T2 (two-token, generic first subtoken)**: shp, whs, rgn, lne, empl, ordr, catg, budg, schd, srvc

## Citation
Inside-Out: A Circuit-Level Analysis of Table-Alias Scoping in Nested SQL Generation