Commit Graph
150 Commits
Author SHA1 Message Date
wassname 4c95073e87 tidy 2023-12-17 08:28:09 +08:00
wassname 4f869fcbae misc 2023-12-17 07:12:25 +08:00
wassname 926aea1c46 new intervention refactor runs 2023-12-16 17:21:42 +08:00
wassname 7119b6c19a wip 2023-12-16 16:50:59 +08:00
wassname 9d176e3905 wip b4 new intervention 2023-12-16 12:06:35 +08:00
wassname 0d6bd5dd03 tidy 2023-12-16 11:06:22 +08:00
wassname b08f3459e9 tidy 2023-12-16 11:06:15 +08:00
wassname 7db7f93fab logger 2023-12-16 11:06:10 +08:00
wassname 03bca70fe1 allow whitespace before or after token choices
this helps in case of weird formatting. especially with bad prompts, or models that have become less coherent due to interventions
2023-12-16 11:06:03 +08:00
wassname 6c310f66b7 misc 2023-12-16 10:18:33 +08:00
wassname 0adef220bc phi, mean mass doesn't work 2023-12-15 18:59:24 +08:00
wassname ddaf9da4af phi-2 2023-12-14 17:26:40 +08:00
wassname 4e3b0c37fc VAE exp 2023-12-10 12:31:49 +08:00
wassname 3067ae8aae better encoder? 2023-12-09 13:14:56 +08:00
wassname 0045e7cf8d 3 vae experiments WIP 2023-12-08 09:50:33 +08:00
wassname a64db78655 💩 2023-11-02 15:25:22 +08:00
wassname 76a5bef13e prep new ideas 2023-11-02 14:25:24 +08:00
wassname c798753f82 acc=0.82 n=1000 torchinfo, oos 0.5 2023-10-29 13:40:49 +08:00
wassname 67455f72e1 oh a bigger model gets 85%! 2023-10-28 19:48:14 +08:00
wassname 5fb4986686 💩 2023-10-28 19:42:02 +08:00
wassname 6832449c79 misc 2023-10-27 19:36:01 +08:00
wassname 224f9688d2 fix 2023-10-27 19:28:43 +08:00
wassname 3c697cbc8f use chat templating 2023-10-27 19:23:56 +08:00
wassname b0dbdef748 refactor and move to wizard 2023-10-27 17:14:16 +08:00
wassname 183fac2627 misc 2023-10-27 14:01:48 +08:00
wassname 9a8c06d007 wip 2023-10-26 19:25:52 +08:00
wassname e0d2eeabe1 use array2d and 3d 2023-10-26 17:14:17 +08:00
wassname 8ab20f663e fix ans, and first run ranking 2023-10-26 07:59:51 +08:00
wassname 2d120c20c9 add two interventions 2023-10-25 20:41:02 +08:00
wassname d2b37d32ee misc 2023-10-25 19:54:01 +08:00
wassname b73802689f tidy 2023-10-25 16:08:03 +08:00
wassname d0a2ef110f works! 2023-10-25 16:01:42 +08:00
wassname 4c4ea76836 got pipeliens working with a unit test 2023-10-25 14:53:44 +08:00
wassname 24f25c2c50 misc 2023-10-25 13:33:50 +08:00
wassname 65b193eca3 misc 2023-10-25 13:33:08 +08:00
wassname 3ac5a1e01a cache intervention, fixing mem leak 2023-10-25 06:28:19 +08:00
wassname a064bf10ba only use baukit 2023-10-24 16:58:11 +08:00
wassname db5afa8e2a metrics 2023-10-24 14:58:54 +08:00
wassname 878f4a3706 fix mask ,try token pos 2023-10-24 14:33:34 +08:00
wassname c0e9b33c9e misc 2023-10-24 12:30:30 +08:00
wassname 1a52f8df82 both pipelines seem to work 2023-10-24 11:26:22 +08:00
wassname 1d86e96e8a rep from wassname/representation-engineering 2023-10-24 10:46:34 +08:00
wassname aefbdb6cd0 abandon branch for poor results 2023-10-23 15:24:00 +08:00
wassname 1890a214b0 fixes 2023-10-20 13:45:21 +08:00
wassname d8d7a5278c use mlp too, and ignore head distinction 2023-10-20 07:41:29 +08:00
wassname e3ecc17fcf misc 2023-10-20 06:09:34 +08:00
wassname bba011d85f try poetry 2023-10-19 16:02:29 +08:00
deep1 46f65b0016 no more deep :'( 2023-10-19 15:11:23 +08:00
deep1 b9775f273c clean 2023-10-15 20:17:36 +08:00
deep1 3fc5919b12 misc 2023-10-15 20:13:27 +08:00