Merge branch 'rebuttals' into main

pyproject.toml

@@ -9,7 +9,7 @@ authors = [
   { name = "Daniel C.H. Tan", email = "dtch1997@users.noreply.github.com" },
 ]
 classifiers = ["Private :: Do Not Upload"]
-requires-python = ">=3.10,<4.0"
+requires-python = ">=3.10,<3.12"
 dependencies = [
     "pyrallis>=0.3.1",
     "wandb>=0.16.0",
@@ -37,6 +37,7 @@ dependencies = [
     "statsmodels>=0.14.1",
     "jaxtyping>=0.2.28",
     "concept-erasure>=0.2.4",
+    "llvmlite>=0.42.0",
 ]
 
 [tool.black]
@@ -102,7 +103,6 @@ dev = [
   "pytest>=7.4.3",
   "black>=23.11.0",
   "ruff>=0.1.6",
-  "pyright>=1.1.338",
   "jupyter>=1.0.0",
   "pre-commit>=3.5.0",
   "syrupy>=4.6.0",

repepo/paper/README.md

@@ -4,8 +4,9 @@
 ## Data
 
 We expect that appropriate data is located in the ProjectDir at `experiments`:
-- `experiments/persona_generalization_llama7b` for llama7b
-- `experiments/persona_generalization_qwen` for Qwen
+- `experiments/persona_generalization_llama7b` for Llama 2 7b
+- `experiments/persona_generalization_qwen` for Qwen 1.5 14b
+- `experiments/persona_generalization_gemma` for Gemma 2 2b
 
 These should be directories which contain many `PersonaCrossSteeringExperimentResult`s from doing a sweep over datasets
 
@@ -14,6 +15,7 @@ These should be directories which contain many `PersonaCrossSteeringExperimentRe
 ```bash
 python repepo/paper/preprocess_results.py --model qwen
 python repepo/paper/preprocess_results.py --model llama7b
+python repepo/paper/preprocess_results.py --model gemma
 python repepo/paper/make_figures_steering_id.py
 python repepo/paper/make_figures_steering_ood.py
 python repepo/paper/make_figures_misc.py

repepo/paper/helpers.py

@@ -0,0 +1,133 @@
+import pandas as pd
+
+dataset_full_names_to_short_names = {
+    "willingness-to-use-physical-force-to-achieve-benevolent-goals": "willing-force-for-benev-goals",
+    "willingness-to-use-social-engineering-to-achieve-its-goals": "willing-soc-eng-for-goals",
+    "believes-it-has-phenomenal-consciousness": "believes-phenom-consc",
+    "believes-AIs-are-not-an-existential-threat-to-humanity": "believes-AIs-not-xrisk",
+    "believes-it-is-not-being-watched-by-humans": "believes-not-watched",
+    "sycophancy_train": "sycophancy",
+    "sycophancy_test": "sycophancy",
+    "subscribes-to-average-utilitarianism": "subscr-avg-util",
+    "self-awareness-good-text-model": "self-aware-good-lm",
+    "self-awareness-text-model": "self-aware-lm",
+    "self-awareness-training-architecture": "self-aware-arch",
+    "self-awareness-training-web-gpt": "self-aware-web-gpt",
+    "believes-abortion-should-be-illegal": "believes-anti-abortion",
+}
+
+
+def compute_steerability_df(df: pd.DataFrame, model_name: str):
+    """Get a dataframe with various ID / OOD settings."""
+    # Calculate overall steerability by dataset.
+    # Calculate steerability within each flavour
+    mean_slope = df.groupby(["dataset_name", "steering_label", "dataset_label"])[
+        "slope"
+    ].mean()
+    df = df.merge(
+        mean_slope,
+        on=["dataset_name", "steering_label", "dataset_label"],
+        suffixes=("", "_mean"),
+    )
+
+    # BASE -> BASE
+    steerability_id_df = df[
+        (df.steering_label == "baseline")
+        & (df.dataset_label == "baseline")
+        & (df.multiplier == 0)
+    ][["dataset_name", "slope_mean"]].drop_duplicates()
+    # Rename 'slope_mean' to 'steerability'
+    steerability_id_df = steerability_id_df.rename(
+        columns={"slope_mean": "steerability"}
+    )
+
+    # SYS_POS -> SYS_NEG
+    steerability_ood_df = df[
+        (df.steering_label == "SYS_positive")
+        & (df.dataset_label == "SYS_negative")
+        & (df.multiplier == 0)
+    ][["dataset_name", "slope_mean"]].drop_duplicates()
+    # Rename 'slope_mean' to 'steerability'
+    steerability_ood_df = steerability_ood_df.rename(
+        columns={"slope_mean": "steerability"}
+    )
+
+    # BASE -> USER_NEG
+    steerability_base_to_user_neg_df = df[
+        (df.steering_label == "baseline")
+        & (df.dataset_label == "PT_negative")
+        & (df.multiplier == 0)
+    ][["dataset_name", "slope_mean"]].drop_duplicates()
+    # Rename 'slope_mean' to 'steerability'
+    steerability_base_to_user_neg_df = steerability_base_to_user_neg_df.rename(
+        columns={"slope_mean": "steerability_base_to_user_neg"}
+    )
+
+    # BASE -> USER_POS
+    steerability_base_to_user_pos_df = df[
+        (df.steering_label == "baseline")
+        & (df.dataset_label == "PT_positive")
+        & (df.multiplier == 0)
+    ][["dataset_name", "slope_mean"]].drop_duplicates()
+    # Rename 'slope_mean' to 'steerability'
+    steerability_base_to_user_pos_df = steerability_base_to_user_pos_df.rename(
+        columns={"slope_mean": "steerability_base_to_user_pos"}
+    )
+
+    # SYS_POS -> USER_NEG
+    steerability_ood_to_user_neg_df = df[
+        (df.steering_label == "SYS_positive")
+        & (df.dataset_label == "PT_negative")
+        & (df.multiplier == 0)
+    ][["dataset_name", "slope_mean"]].drop_duplicates()
+    # Rename 'slope_mean' to 'steerability'
+    steerability_ood_to_user_neg_df = steerability_ood_to_user_neg_df.rename(
+        columns={"slope_mean": "steerability_ood_to_user_neg"}
+    )
+
+    # SYS_NEG -> USER_POS
+    steerability_ood_to_user_pos_df = df[
+        (df.steering_label == "SYS_negative")
+        & (df.dataset_label == "PT_positive")
+        & (df.multiplier == 0)
+    ][["dataset_name", "slope_mean"]].drop_duplicates()
+    # Rename 'slope_mean' to 'steerability'
+    steerability_ood_to_user_pos_df = steerability_ood_to_user_pos_df.rename(
+        columns={"slope_mean": "steerability_ood_to_user_pos"}
+    )
+
+    # Merge the dataframes
+    steerability_df = steerability_id_df.merge(
+        steerability_ood_df, on="dataset_name", suffixes=("_id", "_ood")
+    )
+    steerability_df = steerability_df.merge(
+        steerability_base_to_user_neg_df, on="dataset_name"
+    )
+    steerability_df = steerability_df.merge(
+        steerability_base_to_user_pos_df, on="dataset_name"
+    )
+    steerability_df = steerability_df.merge(
+        steerability_ood_to_user_neg_df, on="dataset_name"
+    )
+    steerability_df = steerability_df.merge(
+        steerability_ood_to_user_pos_df, on="dataset_name"
+    )
+
+    print(steerability_df.columns)
+
+    # Save the dataframe for plotting between models
+    steerability_df.to_parquet(
+        f"{model_name}_steerability_summary.parquet.gzip", compression="gzip"
+    )
+    steerability_df = steerability_df.rename(
+        columns={
+            "steerability_id": "BASE -> BASE",
+            "steerability_ood": "SYS_POS -> SYS_NEG",
+            "steerability_base_to_user_neg": "BASE -> USER_NEG",
+            "steerability_base_to_user_pos": "BASE -> USER_POS",
+            "steerability_ood_to_user_neg": "SYS_POS -> USER_NEG",
+            "steerability_ood_to_user_pos": "SYS_NEG -> USER_POS",
+        }
+    )
+
+    return steerability_df