A governed fine‑tune for $0.86: Qwen3‑4B on IFC/BIM, with a person approving every step

Описание и применение

Гибридный процесс обучения модели Qwen3‑4B на IFC/BIM с контролем бюджета и качества данных, завершившийся снижением потерь на 89 % и без потери общей способности.

Задача

Fine‑tuning a model is easy to start, but running it responsibly in a team is hard: no spend approval, no data record, and no objective comparison to the base model.

Результат

The fine‑tune ran for 73 minutes on a single T4, cost $0.86, reduced test loss from 1.955 to 0.218, and showed no measurable degradation on standard benchmarks.

Рабочий процесс

• Prepare dataset with approval and deduplication
• Approve run after cost and data checks
• Acquire GPU instance and run smoke test
• Train for 353 steps with checkpoints every 50 steps
• Release GPU instance after training
• Compare candidate to base model on held‑out data
• Promote model after approval

Доказательства

Fine‑tuning a model is easy to start. Running it responsibly in a team is the hard part.

Источник

The adapter, the dataset, the training curve and the raw eval results are public: cairndev/qwen3-4b-ifc-bim-qlora and cairndev/ifc-bim-alpaca-dedup.

Источник

The run then paused at promote-model, with the losses on the card. Promotion is a decision, not a side effect.

Источник