blackBOX25I47's picture
Add dataset card
1c609ac verified
|
Raw
History Blame Contribute Delete
1.75 kB
---
license: cc-by-4.0
language:
- en
tags:
- mathematics
- knowledge-graph
- sqlite
- tool-use
pretty_name: Filtered Math Knowledge Base (SQLite)
---
# Knowledgebase Filtered (SQLite)
College-math knowledge graph as a queryable SQLite + FTS5 database, designed
for retrieval / tool-call agents (`kg_search_statement`, `kg_get_statement`,
`kg_get_dependencies`, `md_fetch_section`, ...).
## What's filtered
The full corpus (887,223 statement nodes) is retained so dependency lookups and
full-text search stay complete. Two annotation columns were added to mark the
cleaned **anchor pool** — statements good enough to seed an exam question:
| column | meaning |
|---|---|
| `anchor_eligible` | `1` for the **23,182** complete anchors |
| `completeness_verdict` | `self_contained` / `incomplete` / `NULL` |
### How the 23,182 anchors were selected
- Base filter: `has_proof=1 AND is_closed=1 AND has ≥1 dependency`
- **11,750** long anchors (200–4000 chars) — kept directly
- **11,432** short anchors (50–199 chars) — **rescued** by an LLM completeness
check (`qwen3.5-35b-a3b`, no-think): judged `self_contained`
- 8,086 short anchors judged `incomplete` are flagged but **not** anchor-eligible
- <50-char statements are excluded from anchoring (too fragmentary)
## Usage
```python
import sqlite3
conn = sqlite3.connect("file:index.sqlite?mode=ro&immutable=1", uri=True)
conn.row_factory = sqlite3.Row
# the cleaned anchor pool:
rows = conn.execute("SELECT cid, entity_name FROM statements WHERE anchor_eligible=1").fetchall()
```
Point the tool layer at it via `COLLEGE_MATH_KG_INDEX=/path/to/index.sqlite`.
## Tables
`statements` (nodes + the two new columns), `fts` (FTS5 over name+statement),
`books` (book metadata).