Edit ‘autogollark’

autogollark.myco

@@ -25,6 +25,7 @@ Autogollark currently comprises the dataset, the search API server and the [[htt
 * {RL (also include reasoning, of course). Probably hard though (sparse rewards). https://arxiv.org/abs/2403.09629. [[https://arxiv.org/abs/2503.22828]] would probably work.
 * Unclear whether model could feasibly learn tool use "from scratch", so still need SFT pipeline.
 }
+* https://arxiv.org/abs/2310.04363 can improve sampling (roughly) //and// train for tool use.
 }
 * {Local finetune only? Would be more tonally consistent but dumber, I think.
 * Temporary bursts of hypercompetence enabled by powerful base model are a key feature. Small model is really repetitive.