I found Kiln a few months ago while looking for a UI to help build a dataset for fine-tuning a model on Grapheme-to-Phoneme (G2P) conversion. I’ve contributed t

I found Kiln a few months ago while looking for a UI to help build a dataset for fine-tuning a model on Grapheme-to-Phoneme (G2P) conversion. I’ve contributed to the repo since. In my G2P task, smaller models were splitting phonemes inconsistently, which broke downstream tasks and caused a lot of retries - and higher costs. I fine-tuned Gemini, GPT-4o-mini, and some LLaMA and Qwen models on Fireworks.ai using Kiln, and it actually helped reduce those inconsistencies