3 vae experiments WIP

mjc_notes.md

@@ -1879,8 +1879,95 @@ ideas:
   - [x] aready tried neutral vs positive, like 80%, with good generalization, but I wonder if it's trivial?
 - [x] tried ranking vs mse. mse is still richer
 - [x] tried big vs small model, big seems to help
-- [x] Try a negative vs positve intervention - nope
+- [x] Try a negative vs positive intervention - nope
 - [x] try intervening on every layer = nope
 - [x] flip x0, x1... oh wait with ranking it doesn't know which is which anyway. If I try SL I will need to
 
 I'm out of idea? it does overfit, so maybe only giving it later layers? That helps
+
+
+
+# Brainstorm with Nick
+
+
+Styalized facts:
+- the pair of inferences need to be far apart
+- it's hard get the models to lie
+- using soft labels rather only truth direction helps a lot
+- yes/no have to be significant labels. or it's all junk
+
+Things I've tried:
+- ranking loss, nll, binary, dice loss
+- lots of types of pairs
+    - pair of interventions using gradient descent
+    - mcdropout pairs: didn't help, not very far apart
+    - find a linear weight modifcation (random, pca, could one that makes it lie) and get pair with a -ve and +ve intervention. This is hard and didn't work
+- Lots of way to get lies
+
+Nicks ideas:
+- soft labels - tried with mse
+- clear lying, e.g. in a game like chess
+- try other internals states which are easier and more common. e.g. optimistic/pessimistic, cooprative, love, tired, grumpy, talkative,
+- maybe it doesn't have in innner state for lies, like no habits, no signifance on it. plus it's a goodie two shoes who never lies. so train one for lying
+    - e.g. chess, it can chose to ask stockfish. e.g. it's caught 20% of the time, and forfits the game. so it would only do it when it has a chance of helping it win the game
+
+
+Maybe just lots of data > 7000
+
+
+My takeaway: 
+- Maybe there's not enougth going in small models? 
+- Maybe not enought data to decode internals?
+- Maybe need sparse VAE?
+- Maybe need a model that actually lies? So there are internal pathways
+- I have this tradoff between contrasting internal states vs plausible internal states. I want two realistic inferences which are differen't. States where I inject noise and get 90% of Yes, and 87%^of Yes are kind of useless and difficult to distinguish. On the other hand  E.g. if I generate an in intervention to get a lie, it might rely in junk outputs like "No No No No No". That might be 99% No, but it's useless. 
+
+
+# 2023-12-03 08:11:55
+
+Ideas
+- Train QLoRa few shot liar (although I hope to find a method that works on all models)
+- N>10000 (slow)
+- Use the latest space of a VAE. Maybe one that predicts the next hidden states?
+
+# 2023-12-07 14:29:00
+
+Let's try VAE. How will it work?
+
+
+Right now I'm using pipelines that do an intervention. 
+
+
+# Running
+
+```sh
+python notebooks/make_dataset2.py --max_examples 1720 220 --datasets imdb glue:qnli super_glue:boolq 
+
+```
+an in intervene.py/create_cache_interventions we get the activations that are used to intervent and get a pair of hidden states
+- rep_reading_pipeline.get_directions which uses PCA to get an intervention
+
+
+
+# 2023-12-08 07:06:05
+
+bugs?:
+- does reversing the labels work with pca since it's directionless?
+- the pca intervention does a weird even - odd from the hidden states
+
+modify PLConvProbeLinear to be like a world model? reconstruction and prediction loss...
+
+right now we are using neut? (see https://vscode.dev/github/wassname/discovering_latent_knowledge/blob/pipelinesv2/src/repe/rep_control_pipeline_baukit.py#L94 )
+
+wait each pipeline gets activations for *1 and *0. But we have a pos and neg...
+
+
+how to world models work?
+- they take in an image [64, 64, 3] and make it small e.g. [8], into a quantized space?
+- then reconstruct it they have a reconstruction loss
+- if they have another objective you can train both at once, or transitions/
+https://colab.research.google.com/drive/1rPy82rL3iZzy2_Rd3F82RwFhlVnnroIh?usp=sharing#scrollTo=2MD88v4Zvw-r
+- it's simple, the encoder is just a linear
+
+
+n_instances - remove this, another batch dim

poetry.lock

@@ -1151,6 +1151,21 @@ qtconsole = ["qtconsole"]
 test = ["pytest (<7.1)", "pytest-asyncio", "testpath"]
 test-extra = ["curio", "matplotlib (!=3.2.0)", "nbformat", "numpy (>=1.21)", "pandas", "pytest (<7.1)", "pytest-asyncio", "testpath", "trio"]
 
+[[package]]
+name = "jaxtyping"
+version = "0.2.24"
+description = "Type annotations and runtime checking for shape and dtype of JAX arrays, and PyTrees."
+optional = false
+python-versions = "~=3.9"
+files = [
+    {file = "jaxtyping-0.2.24-py3-none-any.whl", hash = "sha256:b0e90891bbee882d5d3487023d132227d45bdaca3e72937d491b74334f148826"},
+]
+
+[package.dependencies]
+numpy = ">=1.20.0"
+typeguard = ">=2.13.3,<3"
+typing-extensions = ">=3.7.4.1"
+
 [[package]]
 name = "jedi"
 version = "0.19.1"
@@ -3623,6 +3638,21 @@ build = ["cmake (>=3.18)", "lit"]
 tests = ["autopep8", "flake8", "isort", "numpy", "pytest", "scipy (>=1.7.1)"]
 tutorials = ["matplotlib", "pandas", "tabulate"]
 
+[[package]]
+name = "typeguard"
+version = "2.13.3"
+description = "Run-time type checker for Python"
+optional = false
+python-versions = ">=3.5.3"
+files = [
+    {file = "typeguard-2.13.3-py3-none-any.whl", hash = "sha256:5e3e3be01e887e7eafae5af63d1f36c849aaa94e3a0112097312aabfa16284f1"},
+    {file = "typeguard-2.13.3.tar.gz", hash = "sha256:00edaa8da3a133674796cf5ea87d9f4b4c367d77476e185e80251cc13dfbb8c4"},
+]
+
+[package.extras]
+doc = ["sphinx-autodoc-typehints (>=1.2.0)", "sphinx-rtd-theme"]
+test = ["mypy", "pytest", "typing-extensions"]
+
 [[package]]
 name = "typing-extensions"
 version = "4.8.0"
@@ -3894,4 +3924,4 @@ multidict = ">=4.0"
 [metadata]
 lock-version = "2.0"
 python-versions = ">=3.10,<3.13"
-content-hash = "4241642e7d3858d84e2e6feb2f13ffc83c5704571cc702ab3777887ca38dc491"
+content-hash = "c7fc8303fe2dab22214f1850274eb8c57ccc5a0dccb6989d4fd449bbca0830fb"

pyproject.toml

@@ -28,6 +28,7 @@ scikit-learn = "^1.3.1"
 pytorch-optimizer = "^2.12.0"
 pathvalidate = "^3.2.0"
 torchinfo = "^1.8.0"
+jaxtyping = "^0.2.24"
 
 [[tool.poetry.source]]
 name = "pytorch"