DEVLOG #48 - A fresh set of repair cases is ready
SL-LLM-R has a repair-search stage that generates possible fixes after a wrong attempt and reports how often one reaches the target outcome. The last comparison leaned on familiar cases, so a fresh set of task and seed pairs was needed before trying again. I put that fresh split together without changing the repair budgets, scoring rules, or thresholds. The earlier NO-GO stays in place because this creates the follow-up comparison but does not score the model yet. I still do not know whether the repair search is better, and this fresh set is meant to answer that.
Comments 1
You gotta get at least 10 Billions with this much hours.
We’ll follow your development with interest.
Sign in to join the conversation.