so i’ve been making datasets to fine tune small models on tool calling, and the most boring part is always the same, checking if the data is actually good before you waste a training run on it. bad tool names, invented arguments, the model calling a tool for “2+2”, duplicates, answers that all start the same way, stuff like that.<\/p>\n
i was doing these checks by hand and got tired of it, so i built a small thing that runs the whole pipeline for me and i put it online. it’s free, no account, no login, nothing. you just drop your dataset and your tool catalog and it tells you what’s wrong, example by example, with the reason. it runs fully in your browser, the dataset never gets uploaded anywhere. if your file is too big for that (gigabytes), there’s a desktop version that reads it straight from disk so your RAM doesn’t blow up. that one is open source. it also splits your data into clean \/ kto \/ rejected and gives you a starting training config based on the actual numbers of your corpus, not generic advice. I mostly built it for myself but figured someone here might need the same thing. would be happy to know if it’s useful, or if there are checks you care about that i’m not doing yet. <\/p>\n
link: nothumanallowed.com\/tools\/dataset-validator<\/a><\/p>\n https:\/\/github.com\/adoslabsproject-gif\/dataforge-studio<\/a><\/p>\n<\/div>\n submitted by \/u\/Key-Outcome-2927 <\/a>
[link]<\/a><\/span> [comments]<\/a><\/span><\/p>
