4 ms·
The fact that instruction tuning works at all is a small miracle, getting a rigorous idea of trusted vs untrusted input is not at all an easy task.
by rcxdude 1y ago
The fact that instruction tuning works at all is a small miracle, getting a rigorous idea of trusted vs untrusted input is not at all an easy task.
- cubefox 1y agoIt should work like normal instruction tuning, except the SFT examples contain additional instructions in <|quote|> tokens which are ignored in the sample response. So more complex than ordinary SFT but not that much more.