Neel Nanda

[Proposal] Include num_kv_heads in the Model Properties Table in the docs

2

### Proposal Include num_kv_heads in the Model Properties Table in the docs ### Motivation A fair amount of models now use multi-query attention (ie single key and single value) or...

Add tests + better docs to ActivationCache

2

Add tests that the methods in the ActivationCache class work correctly. Go through the documentation and clarify things that are unclear (this is hard for me to do, so even...

[Proposal] Add Support for Yi-6B and Yi-34B

2

### Proposal Yi-6B and Yi-34B are new models that make a plausible claim to be the current best open source models, beating Falcon 180B and LLaMA-2 70B on MMLU. It'd...

Better docs for model properties

5

Make [this table](https://github.com/neelnanda-io/TransformerLens/blob/main/easy_transformer/model_properties_table.md) better and cover key info for model architecture - whether it uses parallel attn & MLPs, and what positional embedding it is. Add text at the bottom...

[Proposal] Have ActivationCache.get_full_resid_decomposition support passing in a vector/tensor to project onto

1

### Proposal Allow ActivationCache.get_full_resid_decomposition to receive a `project_output_onto` tensor that is either a [d_model] tensor or [d_model, num_outputs] tensor, such we multiply the output by that. Internally, rather than taking...

Add mixed precision inference incl loading

12

Add the option to load models in bfloat16 and float16. Esp important for large models like GPT-J and GPT-NeoX. Ideally, load from HuggingFace in this low precision, do weight processing...

Google have a very cool-looking tool for (mostly non-MI) interpretability of language models, called LIT. It seems designed to be framework agnostic, and to be able to take a wrapper...

Add a demo of direct path patching

3

Direct path patching is like activation patching, but rather than patching in the output of component A, it acts on *pairs* of components A and B (in a layer after...

Add tests + better docs for FactoredMatrix

3

Add tests that the FactoredMatrix class works (essentially that each of its methods correctly mimics the result for the actual matrix product). Go through the documentation and clarify things that...

Add tests + better docs for tokenization methods

5

Add tests that the tokenization methods work (to_tokens, to_string, to_str_tokens, get_token_position) Go through the documentation and clarify things that are unclear (this is hard for me to do, so even...