Changelog

Unreleased — v1.0.0 preparation

Add supervised fine-tuning with response-only loss, validation, and JSON presets.
Add reviewed instruction-data preparation and paired base/Instruct evaluation.
Measure batched CPU/CUDA inference with warmup and accurate token counts.
Add packed INT4 reference experiments and quantized-checkpoint reload support.
Support RoPE and context/vocabulary overrides for controlled ablation studies.
Correct weighted perplexity, unlimited evaluation, compiled checkpoints and QA tables.
Exercise real tokenizer round trips without external artifacts.
Document reproduction, publishing prerequisites and remaining empirical work.

The v1.0.0 release is pending trained artifacts and validated experiment results.