j14i/cl-macros

Common Lisp Macro Transformations

A fine-tuning dataset for training models to generate Common Lisp macros. Each example is a call-form, macro-definition, and expanded-form triple.

Summary

• 4,267 examples from 120+ Common Lisp libraries
• Split: 2,985 train / 637 validation / 645 test
• Format: JSONL with instruction, input, output, category, technique, complexity, quality_score
• Mean quality score: 0.80
• Sources: Let Over Lambda, On Lisp, Alexandria, Serapeum, Iterate, Trivia, Parenscript, CFFI, CL-PPCRE, and 110+ more from Quicklisp and awesome-cl

Task

Train a model (<=32B parameters) to generate Common Lisp macro definitions. Given a code pattern or macro call, the model produces the defmacro form that performs the transformation.

This is fundamentally different from code generation -- macros are code that writes code.

Usage

from datasets import load_dataset
ds = load_dataset("j14i/cl-macros", split="train")

Generation Pipeline

Macros are discovered via SBCL + Quicklisp: do-external-symbols finds all exported macros, synthetic call forms are generated, macroexpand-1 verifies expansions, and a Python classifier auto-labels technique, category, and complexity.