README.md

---
language:
- en
library_name: sentence-transformers
pipeline_tag: text-ranking
base_model: cross-encoder/ms-marco-MiniLM-L6-v2
tags:
- sentence-transformers
- text-ranking
- text2sql
- schema-linking
- aap-sql
---

# AAP-SQL R1 schema reranker

AAP-SQL R1 是完整 AAP-SQL 設定中的 cross-encoder schema 重排器。它對 E2 取回的候選欄位評分，保留前 10 個欄位供後續提示增強使用。

AAP-SQL R1 is the cross-encoder schema reranker used by the full AAP-SQL configuration. It scores the candidates returned by E2 and retains the top 10 columns for prompt augmentation.

## Model details

- Base model: `cross-encoder/ms-marco-MiniLM-L6-v2`
- Training objective: `BinaryCrossEntropyLoss`
- Training seed: `42`
- Training data: schema-ranking examples derived from the BIRD training split and schema descriptions
- Expected library: `sentence-transformers>=5.1.2`

## Use with AAP-SQL

Download this repository into the path expected by the final runner:

```powershell
hf download TommyPanLab/AAP-SQL-R1 --local-dir models/cross_encoder_schema_paper_repro
```

Direct loading:

```python
from sentence_transformers import CrossEncoder

model = CrossEncoder("TommyPanLab/AAP-SQL-R1")
scores = model.predict([("user question", "table.column: column description")])
```

The complete pipeline, required BIRD directory layout, and Gemini 3.1 result are documented in the [AAP-SQL publication branch](https://github.com/Tommyweige/AAP-SQL/tree/codex/final-aap-sql-experiment/AAP-SQL-Original).

## Data and license notice

The training examples were derived from the BIRD benchmark. Review the [BIRD project terms](https://bird-bench.github.io/) before using the model. No additional license has been declared for these fine-tuned weights; the upstream model and dataset terms still apply.