smitxxiv/Qwen3-Re4B-SQL-TeCoD-TMM

TeCoD SQL Template Matcher

Fine-tune of Qwen/Qwen3-Reranker-4B used by TeCoD, a template-guided constrained decoding system for text-to-SQL.

This model is the TeCoD template-matching reranker. It scores whether a user question matches a retrieved masked question/template, helping TeCoD select recurring SQL templates before generation.

• Project page: https://sslab-cse-iitb.github.io/tecod/
• Source repository: https://github.com/SSLab-CSE-IITB/tecod
• Base model: https://huggingface.co/Qwen/Qwen3-Reranker-4B
• Training data source: BIRD train split.

Intended Use

This model is intended as an internal component of TeCoD and related template-based text-to-SQL systems. It is not a standalone SQL generator. In TeCoD, it is used after vector retrieval and before SQL generation to rerank candidate SQL templates.

Input Format

The model is used as a cross-encoder over a question pair. Order matters: the first sequence should be the masked candidate/template question, and the second sequence should be the raw user question.

Premise:    "Show movies released in _ sorted by popularity desc"
Hypothesis: "What are the top films from 2010 by viewer count?"

Entity values in the candidate question are masked with a space-padded underscore _. The same mask token is used for strings, numbers, dates, and other literal values. Swapping the input order or changing the masking convention can degrade reranking quality.

Training Summary

• Base model: Qwen/Qwen3-Reranker-4B
• Architecture: Qwen3ForSequenceClassification
• Data: approximately 1.48M NLI pairs derived from BIRD questions.
• Positive pairs: template-paired questions, self paraphrases, and partner paraphrases that preserve the SQL template.
• Negative pairs: hard negatives mined using nearest-neighbor retrieval over masked questions, with both masked and unmasked query variants used during pair construction.
• Labels: entailment, neutral, contradiction.
• The neutral label is retained for compatibility with a 3-class NLI head but was not used as a training target.

Limitations

• Specialized for masked text-to-SQL question/template matching.
• Not intended for general NLI, semantic similarity, or SQL generation.
• Assumes the same masking convention and candidate-template construction used by TeCoD.
• The neutral label is untrained; inference should use entailment vs. contradiction or renormalize over labels {0, 2}.
• Very long question pairs and non-English inputs are not validated.
• The reranking score is one signal in a larger text-to-SQL pipeline; it does not guarantee final SQL correctness.

References

• TeCoD project page: https://sslab-cse-iitb.github.io/tecod/
• TeCoD source repo: https://github.com/SSLab-CSE-IITB/tecod
• Base model: https://huggingface.co/Qwen/Qwen3-Reranker-4B
• Training Data - BIRD Train Set: https://bird-bench.github.io/

If you use this model as part of TeCoD, please cite:

@article{10.1145/3769822,
  author = {Jivani, Smit and Maheshwari, Saravam and Sarawagi, Sunita},
  title = {Reliable Answers for Recurring Questions: Boosting Text-to-SQL Accuracy with Template Constrained Decoding},
  journal = {Proceedings of the ACM on Management of Data},
  volume = {3},
  number = {6},
  pages = {1--26},
  year = {2025},
  month = dec,
  publisher = {Association for Computing Machinery},
  address = {New York, NY, USA},
  doi = {10.1145/3769822},
  url = {https://doi.org/10.1145/3769822}
}

License

Apache 2.0