Training Logs
Practical run logs, evaluation methods, build reports, and postmortems from LoRADock contributors.
Training Logs
5 articlesFine-Tuning Qwen2.5-7B on 18,000 Support Tickets
A repeatable workflow for preparing, training, and evaluating a Qwen support-routing adapter while keeping labels, privacy checks, and release decisions explicit.
What Changed When I Increased LoRA Rank from 8 to 32
A controlled comparison of rank 8, 16, and 32 for instruction tuning, with capacity calculations, evaluation rules, and a release decision.
A Three-GPU LoRA Run That Was Slower Than One GPU
A diagnostic workflow for distributed LoRA training that separates data loading, synchronization, compute, and checkpoint costs before adding more devices.
Recovering a Training Run After a Broken Dataset Upload
A recovery checklist for malformed or truncated JSONL uploads that distinguishes dataset corruption from training failures and preserves reproducible run evidence.
My Default LoRA Configuration for Small Instruction Datasets
A conservative starting configuration for small instruction datasets, including parameter failure signals, tuning order, and registry documentation for repeatable LoRA runs.