Curation: 13 new terms (model training / compression / efficiency)

[umzcio]

Fills gaps in the training/compression/efficiency cluster (the REAP neighborhood). One PR because several cross-link each other.

Scaling & training dynamics

• Chinchilla (chinchilla, Training) — DeepMind's compute-optimal result: smaller models + far more data.
• Grokking (grokking, Training) — sudden generalization long after training looked done.
• Double Descent (double-descent, Training) — bigger-first-hurts-then-helps error curve.

Architecture / precision

• BitNet (bitnet, Architectures) — 1-bit / ternary-weight LLMs.
• bfloat16 (bfloat16, Hardware & Systems) — the 16-bit format most training uses.

Fine-tuning & alignment methods

• DoRA (dora, Training) — weight-decomposed LoRA.
• KTO (kto, Training) — align from simple good/bad labels.
• ORPO (orpo, Training) — instruction + preference tuning in one step.
• Process Reward Model / PRM (process-reward-model, Training) — scores each reasoning step.

Distributed training

• 3D Parallelism (3d-parallelism, Hardware & Systems) — data + pipeline + tensor parallelism combined.
• Model Sharding (model-sharding, Hardware & Systems) — split a model across devices.

Compression / data

• Model Merging (model-merging, Optimization) — combine models by their weights (soups/TIES/SLERP). Complements reap.
• Data Mixture (data-mixture, Data) — the training-data recipe.

Notes

• Dropped 2 from the plan as already present: rope (RoPE) and zero-redundancy-optimizer (ZeRO) already exist — the writes errored out and I skipped them. Also excluded synonyms of existing terms (knowledge-distillation→distillation, parameter-efficient-fine-tuning→peft, supervised-fine-tuning→sft, low-rank-adaptation→lora).
• All canonical, well-established ML concepts; each cross-links existing terms (scaling-laws, lora, reward-model, fsdp, tensor-parallelism, reap, etc.).

All files pass scripts/validate.mjs (985 files valid). Nothing is live until you merge.