wassname
|
5c2ec738d7
|
wip
|
2024-01-14 20:46:26 +08:00 |
|
wassname
|
2609c7a2f0
|
kinda working, latent space looks ok
|
2024-01-14 15:22:23 +08:00 |
|
wassname
|
b4ee6e9e8d
|
wip
|
2024-01-14 15:04:58 +08:00 |
|
wassname
|
fc31d38131
|
add IRIS code
|
2024-01-14 10:38:55 +08:00 |
|
wassname
|
5569565436
|
about to try iris
|
2024-01-14 10:33:31 +08:00 |
|
wassname
|
2560a8720b
|
try summing for l1 and l2
|
2024-01-14 06:45:05 +08:00 |
|
wassname
|
4698e9dbce
|
acc=80% deep
|
2024-01-13 10:24:10 +08:00 |
|
wassname
|
4e1d51063b
|
wip
|
2024-01-13 07:42:05 +08:00 |
|
wassname
|
390925416f
|
90% ood! sae2
|
2024-01-12 16:53:43 +08:00 |
|
wassname
|
2c3511d5be
|
trying the other SAE, also making my own multilayer one
|
2024-01-12 15:08:12 +08:00 |
|
wassname
|
c29101be84
|
wip
|
2024-01-11 08:34:56 +08:00 |
|
wassname
|
559ff63b1e
|
checkpoint
|
2024-01-08 06:03:40 +08:00 |
|
wassname
|
21a2132a99
|
Update dependencies and fix bugs
|
2024-01-07 08:59:57 +08:00 |
|
wassname
|
ec2dcc08dc
|
wip
|
2024-01-06 14:19:51 +08:00 |
|
wassname
|
305f5c644f
|
wip
|
2024-01-06 13:24:58 +08:00 |
|
wassname
|
e9d8142e45
|
wip
|
2024-01-06 08:58:45 +08:00 |
|
wassname
|
730605bfe3
|
clean probes, working on ae
|
2024-01-05 10:13:31 +08:00 |
|
wassname
|
b48d8f89c2
|
refactoring probes, currently in 102
|
2024-01-04 13:07:15 +08:00 |
|
wassname
|
0ef1130909
|
clean, refactoring eval, importance matrix
|
2024-01-02 15:10:26 +08:00 |
|
wassname
|
b83ff86edd
|
fix bug where we diff neurons instead of layers, ops
|
2023-12-31 18:00:57 +08:00 |
|
wassname
|
4bc8dd692b
|
found that ranger21 was causing the bug, still haven't got bnb fully working
|
2023-12-30 13:54:14 +08:00 |
|
wassname
|
dc111c3e99
|
wip collect residuals from adapter layers
|
2023-12-29 16:34:07 +08:00 |
|
wassname
|
f94851f184
|
about to try to collect ativations from the places adapters intervene at
|
2023-12-29 15:31:05 +08:00 |
|
wassname
|
d314ed912e
|
try modifying ia3 to add instead of mult
|
2023-12-28 12:18:59 +08:00 |
|
wassname
|
cfac3cdef7
|
finished the truth adapter vs probe experiment.
the probe was better
|
2023-12-28 08:34:04 +08:00 |
|
wassname
|
bb1b41728c
|
new probe format works
|
2023-12-28 08:06:22 +08:00 |
|
wassname
|
1538c8c94d
|
wip
|
2023-12-27 15:26:00 +08:00 |
|
wassname
|
58b36ab990
|
making proxy vs label clear, refactoring probes
|
2023-12-27 13:40:18 +08:00 |
|
wassname
|
4e92e6f3f4
|
misc
|
2023-12-26 10:04:33 +08:00 |
|
wassname
|
8a9bf6920c
|
can't get gp's working
|
2023-12-26 07:56:29 +08:00 |
|
wassname
|
a14fc0540d
|
more ds
|
2023-12-25 20:28:57 +08:00 |
|
wassname
|
7c4500d33f
|
misc
|
2023-12-25 12:21:43 +08:00 |
|
wassname
|
ed6343344e
|
exp 2
|
2023-12-25 11:29:45 +08:00 |
|
wassname
|
c6e71f8a48
|
fix bug in collecting hidden states
|
2023-12-24 20:04:09 +08:00 |
|
wassname
|
7776d2639e
|
exp
|
2023-12-24 19:42:49 +08:00 |
|
wassname
|
d702ba4c0e
|
styles
|
2023-12-24 10:49:31 +08:00 |
|
wassname
|
74dd709fec
|
wip
|
2023-12-23 16:46:16 +08:00 |
|
wassname
|
9229017da4
|
fixed bug with the select functions
|
2023-12-23 13:39:33 +08:00 |
|
wassname
|
92999cfafd
|
better eval
|
2023-12-23 10:05:52 +08:00 |
|
wassname
|
756010c314
|
misc
|
2023-12-22 17:36:14 +08:00 |
|
wassname
|
7ace3430ab
|
another experiment
|
2023-12-22 16:54:23 +08:00 |
|
wassname
|
cf24f2aace
|
misc
|
2023-12-22 15:33:44 +08:00 |
|
wassname
|
3368bb18cf
|
try uncensored models
|
2023-12-22 09:53:26 +08:00 |
|
wassname
|
f1ab0c26cf
|
checkin
|
2023-12-22 08:08:54 +08:00 |
|
wassname
|
ee7fa222cb
|
the experiment runs, now to try for better results
|
2023-12-21 20:38:49 +08:00 |
|
wassname
|
b8666a40c8
|
working on result dataset
|
2023-12-21 20:27:26 +08:00 |
|
wassname
|
93a8b224dd
|
arrowdatasets are way better than huggingface datasets
|
2023-12-21 18:44:55 +08:00 |
|
wassname
|
f781fded7d
|
wip
|
2023-12-21 14:24:15 +08:00 |
|
wassname
|
c78e1a966f
|
predict
|
2023-12-21 13:00:34 +08:00 |
|
wassname
|
034780a479
|
working, wth?
|
2023-12-21 10:42:26 +08:00 |
|