Release v2.0 — Paper Revision · ContextLab/llm-stylometry

What's New

This release accompanies the revised manuscript submitted to Computational Linguistics, addressing reviewer feedback on the original submission.

Dataset-size sweep: Systematically varied training tokens per author (2,500–643,041) across 1,520 newly trained models to characterize data requirements. A sigmoid fit (R²=0.979) estimates that ≥95% attribution accuracy requires ~51,000 tokens per author.
Embedding comparison: Evaluated three pre-trained text embedding models from the MTEB leaderboard (nomic-embed-text-v1.5, bge-m3, Qwen3-Embedding-4B) against our predictive comparison approach. Best embedding accuracy: 81.0% vs our 100%.

New methods and results sections for both analyses
Expanded discussion of relationship to Huang et al. (2025), benchmark feasibility, cross-domain robustness
Supplementary materials with embedding purity/confusion figures
Point-by-point response letter addressing editor and 3 reviewers
Regenerated Oz attribution figure (previously had empty panels)

code/fit_sigmoid.py: Sigmoid fit with bootstrap confidence intervals
code/embedding_comparison.py: Chunk-level nearest-neighbor attribution pipeline with per-book checkpointing
New figure types (flags 6, 7) in run_llm_stylometry.sh and generate_figures.py
3 new remote scripts for dataset-size sweep on GPU clusters
paper/compile.sh: Builds main paper, supplement, response letter, and latexdiff
15 new tests covering sigmoid fit and embedding comparison
Black formatting applied across codebase

data/model_results_ntokens.pkl.gz: Pre-computed results for the dataset-size sweep (98MB)
Embedding results cached in data/embedding_results/ (gitignored; regenerable via code/embedding_comparison.py)