From ec0bce86010630167b4480b0d8a0462fcb46550e Mon Sep 17 00:00:00 2001 From: Michael J Clark Date: Sat, 3 Feb 2024 14:02:27 +0800 Subject: [PATCH] Update README.md --- README.md | 11 ++++++++--- 1 file changed, 8 insertions(+), 3 deletions(-) diff --git a/README.md b/README.md index ba180a3..15c7331 100644 --- a/README.md +++ b/README.md @@ -2,9 +2,14 @@ My own experiments with DLK - [x] use pytorch lightning - [x] batch hidden states 5x faster -- [ ] use llama 13B, to see if larger models give better results -- [ ] eval on some deceptive or misleading statements -- [ ] debug by looking at model output +- [x] use llama 13B, to see if larger models give better results +- [x] eval on some deceptive or misleading statements +- [x] debug by looking at model output +- [x] try ranking hidden state (on a branch) +- [x] try a tokenized autoencoder (on a branch) +- [x] try a sparse autoencoder (on a branch) + +I'm afraid none of these approaches have got the >90% out of distribution results I hope for. I beleive this to be because current models are too small to have internals that reliably produce deception. -------------