Finetune the World — Research & NotesExperiments, derivations and field notes on fine-tuning and post-training.https://finetunetheworld.com/feed.xml2026-10-30T00:00:00ZFinetune the World ResearchLoRA rank vs full fine-tuning: an experiment on structured extractionhttps://finetunetheworld.com/research/lora-rank-vs-full-fine-tuning/2026-10-30T00:00:00Z2026-10-30T00:00:00ZWe fine-tuned Qwen2.5-0.5B on a claims-extraction task with LoRA ranks 1–64 and full fine-tuning. Exact match went from 18% to 100%; rank mattered most on unseen formats. Code, data and results included.Lab notes from our fine-tuning researchhttps://finetunetheworld.com/notes/2026-10-30T00:00:00Z2026-10-30T00:00:00ZDated, short findings from our fine-tuning research: LoRA rank and robustness, the real cost of a training run, DPO β, logits memory and evaluation sample sizes.Your LLM judge is shrinking your results. A 1978 estimator fixes it.https://finetunetheworld.com/research/llm-judge-bias-correction/2026-10-27T00:00:00Z2026-10-30T00:00:00ZAn imperfect LLM judge attenuates measured improvements and can flip model rankings. Simulations, the Rogan–Gladen correction from epidemiology, and how big a calibration set you need.Self-hosting a fine-tuned model vs paying per token: the break-even modelhttps://finetunetheworld.com/research/self-hosting-vs-api-break-even/2026-10-25T00:00:00Z2026-10-30T00:00:00ZWhen does a self-hosted fine-tuned model beat a per-token API? A parametric cost model with a break-even formula, base case of ~1M requests/month, and a sensitivity analysis of every input.Fine-tuning compute: FLOPs, GPU-hours and what a training run really costshttps://finetunetheworld.com/research/fine-tuning-compute-cost/2026-10-23T00:00:00Z2026-10-30T00:00:00ZA transparent compute model for fine-tuning: 6N vs 4N FLOPs per token for full vs LoRA training, MFU, GPU-hours and dollar costs for 0.5B–72B models, with sensitivity analysis.How much GPU memory does it take to fine-tune an LLM?https://finetunetheworld.com/research/gpu-memory-for-fine-tuning/2026-10-21T00:00:00Z2026-10-30T00:00:00ZVRAM requirements for fine-tuning 0.5B–72B models with full fine-tuning, LoRA and QLoRA, derived from real model configs: equations, tables, activation memory and which GPU you need.How many evaluation examples do you need to compare two LLMs?https://finetunetheworld.com/research/eval-sample-size/2026-10-18T00:00:00Z2026-10-30T00:00:00ZPower analysis for LLM evals: analytic sample sizes, a Monte Carlo study of paired vs unpaired designs, and why scoring both models on the same items can halve the examples you need.DPO from first principles, and what β really does with finite datahttps://finetunetheworld.com/research/dpo-beta-finite-data/2026-10-14T00:00:00Z2026-10-30T00:00:00ZA step-by-step derivation of Direct Preference Optimization from the KL-regularized RLHF objective, plus a tabular simulation showing that with limited data β barely controls how far the policy moves.