Выпуск
Why Fine-Tuned Models Keep the Same File Size
My Weird Prompts · 17 ноября 2026 г. · 21 мин
After 70 hours of voice fine-tuning, a Parakeet TDT speech model learns new vocabulary while its 2.5-gigabyte safetensors file barely grows. The format is a fixed-size photograph, not a diary: training nudges existing 32-bit floats rather than adding storage, while discarding 10 gigabytes of live optimizer state. That fixed frame also hides catastrophic forgetting, where new accents and words overwrite old ones. Practical fixes include weight averaging, Elastic Weight Consolidation, and decode-time contextual biasing. Key sources: • Hugging Face / NVIDIA model card , created 2025-08-04, last modified 2026-08-05, retrieved 2026-11-17. — https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3 • safetensors repository README and format specification , retrieved 2026-11-17. — https://github.com/huggingface/safetensors • NVIDIA NeMo-Speech fine-tuning documentation , retrieved 2026-11-17. — https://docs.nvidia.com/nemo/speech/latest/asr/fine_tuning.html Episode 6657 — myweirdprompts.com/n/6657 (permanent short link: myweirdprompts.com/712422) Listen on Android: MWP FM Player — myweirdprompts.com/apk