Turning every "no thats not what i meant" in chat into actual LoRA training data

In three linesA developer built TideForge, a desktop app that converts chat corrections into LoRA training data. Each model reply has a "Teach" button; corrections accumulate as JSONL and trigger PEFT fine-tuning on your base model. Initial test: 110 hand-written corrections on Qwen 0.6B, loss dropped 4.25→0.73, adapter maintained identity across ~30 jailbreak prompts. Free, Windows, GGUF-compatible.

Summary generated by Claude — human-verified