Posts

SAEs are highly dataset dependent: a case study on the refusal direction 2024-11-07T05:22:18.807Z
Open Source Replication of Anthropic’s Crosscoder paper for model-diffing 2024-10-27T18:46:21.316Z
Base LLMs refuse too 2024-09-29T16:04:21.343Z
SAEs (usually) Transfer Between Base and Chat Models 2024-07-18T10:29:46.138Z
Attention Output SAEs Improve Circuit Analysis 2024-06-21T12:56:07.969Z
We Inspected Every Head In GPT-2 Small using SAEs So You Don’t Have To 2024-03-06T05:03:09.639Z
Attention SAEs Scale to GPT-2 Small 2024-02-03T06:50:22.583Z
Sparse Autoencoders Work on Attention Layer Outputs 2024-01-16T00:26:14.767Z

Comments

Comment by Connor Kissane (ckkissane) on SAEs are highly dataset dependent: a case study on the refusal direction · 2024-11-09T02:21:56.960Z · LW · GW

Thanks! I'm not sure. My guess is that if you go super narrow, it may be more likely to result in an inconvenient level of "feature splitting". Since there are only a few total concepts to learn, an SAE of equivalent width might exploit its greater relative capacity to learn niche combinations of features (to reduce sparsity loss).

Comment by Connor Kissane (ckkissane) on Base LLMs refuse too · 2024-09-30T01:38:37.614Z · LW · GW

LLaMA 1 7B definitely seems to be a "pure base model". I agree that we have less transparency into the pre-training of Gemma 2 and Qwen 1.5, and I'll add this as a limitation, thanks!

I've checked that Pythia 12b deduped (pre-trained on the pile) also refuses harmful requests, although at a lower rate (13%). Here's an example, using the following prompt template:

"""User: {instruction}

Assistant:"""

It's pretty dumb though, and often just outputs nonsense. When I give it the Vicuna system prompt, it refuses 100% of harmful requests, though it has a bunch of "incompetent refusals", similar to LLaMA 1 7B:

"""A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions.

USER: {instruction}

ASSISTANT:"""

Comment by Connor Kissane (ckkissane) on Sparse Autoencoders Work on Attention Layer Outputs · 2024-05-17T08:53:00.639Z · LW · GW

Thanks for the comment! We always use the pre-ReLU feature activation, which is equal to the post-ReLU activation (given that the feature is activate), and is purely linear function of z. Edited the post for clarity. 

Comment by Connor Kissane (ckkissane) on SAE-VIS: Announcement Post · 2024-03-31T16:35:06.004Z · LW · GW

Amazing! We found your original library super useful for our Attention SAEs research, so thanks for making this!

Comment by Connor Kissane (ckkissane) on Mech Interp Puzzle 1: Suspiciously Similar Embeddings in GPT-Neo · 2023-08-14T14:20:07.795Z · LW · GW

These puzzles are great, thanks for making them!

Comment by Connor Kissane (ckkissane) on Causal scrubbing: results on induction heads · 2023-07-19T19:57:54.463Z · LW · GW

Code for this token filtering can be found in the appendix and the exact token list is linked.

Maybe I just missed it, but I'm not seeing this. Is the code still available?