cql-anti-pattern-transformer

Tiny (576k parameters) Transformer encoder that predicts CQL anti-patterns from cqlc-core lexer tokens (no text tokenizer). Part of the cqlz family.

• Input: a CQL statement lexed by cqlc-core, identifiers replaced by schema role tokens.
• Output: 6 independent probabilities — MissingPartitionKey, InefficientAllowFiltering, NonKeyRestriction, ClusteringGap, OrderByMismatch, LargeCollectionUpdate.
• Files: model.safetensors (weights, Burn module paths), config.json (Burn ModelConfig).
• Use with the cql-ai CLI: cargo install cql-ai-cli && cql-ai check queries/ --schema schema/.

Training data

100,000 synthetic statements over 600 random schemas (6 epochs, AdamW), labelled by cqlc-core's analyzer (plus a generator rule for LargeCollectionUpdate). See the repository for the generator.

Evaluation

10,000 fresh statements from unseen schemas, threshold 0.5:

Label	Precision	Recall	F1
MissingPartitionKey	0.879	0.891	0.885
InefficientAllowFiltering	1.000	1.000	1.000
NonKeyRestriction	1.000	1.000	1.000
ClusteringGap	1.000	1.000	1.000
OrderByMismatch	0.997	0.954	0.975
LargeCollectionUpdate	1.000	0.998	0.999

Macro-F1 0.977, exact match 0.949. Synthetic distribution only.

Limitations

Synthetic training data; needs the table schema to assign key roles; not a substitute for cqlc's deterministic analysis.