From f53fdc05f7c3ae64970e93d1089aebbe1e347f07 Mon Sep 17 00:00:00 2001 From: wassname Date: Mon, 5 Jun 2023 15:38:13 +0800 Subject: [PATCH] notes --- mjc_notes.md | 9 +++++++++ 1 file changed, 9 insertions(+) diff --git a/mjc_notes.md b/mjc_notes.md index 0b15154..409fdb1 100644 --- a/mjc_notes.md +++ b/mjc_notes.md @@ -87,3 +87,12 @@ Note we are using normalized for sklearn and it seems easy? but unnorm for CCS, hmm I can't get it to reliably lie, even at 30B. Grrr. Maybe an unaligned model? + +OK so I asked on discord +- why not use generate? A: too hard, and how to make sure it's an answer (solution is to use relative probs and a single work). So they just haven't gone there. +- and how to make it lie. one person used yes momentum, nad it worked for them. + +I would like to try: +- [ ] lying larry. with a larry prompt. and larry response. +- [ ] I could also keep sampling untill I get the prob diff I want! :) + - the second actually seems better. since it can make sure that the pairs are the same except the answer!!