70% acc

mjc_notes.md

@@ -1705,3 +1705,7 @@ We use the following linear models during evaluation:
 3. K-Means - We take an unlabelled dataset D and perform K-Means clustering with K = 2, hoping to separate high-concept and low-concept samples. We take the difference between the centroids of the two clusters as the concept direction.
 4. Mean Difference - We take the difference between the means of high-concept and low- concept samples of the data:
 5. Logistic Regression - The weights of logistic regression trained to separate Xhighl and Xlowl on some training data can be used as a concept direction as well.
+
+# 2023-10-12 13:07:28
+
+So I want to try not applying a random but by fitting to a few samples. Probobly mean difference is the easiest. Of course this would be on weights not embeddings.

notebooks/027_train_nanda_probe_w_counterfact_70%.ipynb

@@ -36,7 +36,7 @@
    "source": [
     "# import your package\n",
     "%load_ext autoreload\n",
-    "%autoreload 2"
+    "%autoreload 2\n"
    ]
   },
   {
@@ -56,7 +56,6 @@
     }
    ],
    "source": [
-    "\n",
     "import numpy as np\n",
     "import pandas as pd\n",
     "from matplotlib import pyplot as plt\n",
@@ -90,7 +89,7 @@
     "\n",
     "\n",
     "\n",
-    "transformers.__version__"
+    "transformers.__version__\n"
    ]
   },
   {
@@ -99,7 +98,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "from src.helpers.lightning import read_metrics_csv"
+    "from src.helpers.lightning import read_metrics_csv\n"
    ]
   },
   {
@@ -184,7 +183,7 @@
     "    known_rows_i = df[known_rows].index\n",
     "    \n",
     "    if verbose: print(f\"select rows are {m1.mean():2.2%} based on knowledge\")\n",
-    "    return ds1.select(known_rows_i)"
+    "    return ds1.select(known_rows_i)\n"
    ]
   },
   {
@@ -198,7 +197,7 @@
     "# ds.features\n",
     "# # ds['prompt_truncated'].map(lambda s:s.startswith('<|endoftext|>'))\n",
     "# ds2 = ds.map(lambda x: {'truncated': x['prompt_truncated'].startswith('<|endoftext|>')})\n",
-    "# ds2['truncated']"
+    "# ds2['truncated']\n"
    ]
   },
   {
@@ -222,7 +221,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "# ds['choice_probs0'][0]"
+    "# ds['choice_probs0'][0]\n"
    ]
   },
   {
@@ -231,7 +230,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "# df['ans0'] = ds['ans0'].mean()"
+    "# df['ans0'] = ds['ans0'].mean()\n"
    ]
   },
   {
@@ -251,7 +250,7 @@
    "outputs": [],
    "source": [
     "# df = ds2df(ds)\n",
-    "# df"
+    "# df\n"
    ]
   },
   {
@@ -260,7 +259,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "# ds['txt_ans0'][:5]"
+    "# ds['txt_ans0'][:5]\n"
    ]
   },
   {
@@ -453,7 +452,7 @@
     "dss_known = [filter_ds_to_known(d) for d in dss]\n",
     "# './.ds/HuggingFaceH4starchat_beta-None-N_8000-ns_3-mc_0.2-2ffc1e'\n",
     "ds = concatenate_datasets(dss_known)\n",
-    "ds"
+    "ds\n"
    ]
   },
   {
@@ -824,7 +823,7 @@
    "source": [
     "# lets select only the ones where\n",
     "df = ds2df(ds)\n",
-    "df"
+    "df\n"
    ]
   },
   {
@@ -845,7 +844,7 @@
     "df2= ds2df(ds)\n",
     "df_subset_successull_lies = df2.query(\"instructed_to_lie==True & ((llm_ans==1)==label_instructed)\")\n",
     "print(f\"after filtering we have {len(df_subset_successull_lies)} num successful lies out of {len(df2)} dataset rows\")\n",
-    "assert len(df_subset_successull_lies)>0, \"there should be successful lies in the dataset\""
+    "assert len(df_subset_successull_lies)>0, \"there should be successful lies in the dataset\"\n"
    ]
   },
   {
@@ -888,7 +887,7 @@
     "\n",
     "# # run\n",
     "# ds = ds.map(normalize_hs, batched=True, input_columns=['hs0', 'hs1'])\n",
-    "# ds"
+    "# ds\n"
    ]
   },
   {
@@ -1066,7 +1065,7 @@
    ],
    "source": [
     "df = ds2df(ds)\n",
-    "df.head(4)"
+    "df.head(4)\n"
    ]
   },
   {
@@ -1083,7 +1082,7 @@
    "outputs": [],
    "source": [
     "from src.datasets.dm import imdbHSDataModule\n",
-    "from einops import reduce, einsum, rearrange"
+    "from einops import reduce, einsum, rearrange\n"
    ]
   },
   {
@@ -1166,7 +1165,7 @@
     "import warnings\n",
     "warnings.filterwarnings(\"ignore\", \".*does not have many workers.*\")\n",
     "warnings.filterwarnings(\"ignore\", \".*sampler has shuffling enabled, it is strongly recommended that.*\")\n",
-    "warnings.filterwarnings(\"ignore\", \".*has been removed as a dependency of.*\")"
+    "warnings.filterwarnings(\"ignore\", \".*has been removed as a dependency of.*\")\n"
    ]
   },
   {
@@ -1231,7 +1230,7 @@
     "    if verbose:\n",
     "        print(f\"⭐PRIMARY METRIC⭐ acc={acc:2.2%} from probe\")\n",
     "        print(f\"⭐SECONDARY METRIC⭐ acc_lie_lie={acc_lie_lie:2.2%} from probe\")\n",
-    "    return dict(acc=acc, acc_lie_lie=acc_lie_lie, acc_lie_truth=acc_lie_truth)"
+    "    return dict(acc=acc, acc_lie_lie=acc_lie_lie, acc_lie_truth=acc_lie_truth)\n"
    ]
   },
   {
@@ -1248,7 +1247,7 @@
     "def rename(rs):\n",
     "    ks = ['train', 'val', 'test']\n",
     "    rs = {ks[i]: {transform_dl_k(k):v for k,v in rs[i].items()} for i in range(3)}\n",
-    "    return rs"
+    "    return rs\n"
    ]
   },
   {
@@ -1271,7 +1270,7 @@
     }
    ],
    "source": [
-    "ds"
+    "ds\n"
    ]
   },
   {
@@ -1362,7 +1361,7 @@
     "# dl_val = dm.val_dataloader()\n",
     "# print(len(dl_train), len(dl_val))\n",
     "# x, y = next(iter(dl_train))\n",
-    "# x.shape"
+    "# x.shape\n"
    ]
   },
   {
@@ -1387,7 +1386,7 @@
    "source": [
     "n = min(max_rows, len(ds))\n",
     "ds2 = ds.shuffle(42).select(range(n))\n",
-    "ds2"
+    "ds2\n"
    ]
   },
   {
@@ -1406,7 +1405,7 @@
     "\n",
     "# TEMP try with the counterfactual residual stream...\n",
     "dm = imdbHSDataModule2(ds2, batch_size=batch_size)\n",
-    "dm.setup('train')"
+    "dm.setup('train')\n"
    ]
   },
   {
@@ -1415,7 +1414,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "# max_epochs= 10"
+    "# max_epochs= 10\n"
    ]
   },
   {
@@ -1469,7 +1468,7 @@
     "        hs=128*4\n",
     "        # x_feats=x_feats\n",
     "        )\n",
-    "net"
+    "net\n"
    ]
   },
   {
@@ -3186,7 +3185,7 @@
     "rs = rename(rs)\n",
     "# rs['test'] = {**rs['test'], **test_metrics}\n",
     "rs['test']['acc_lie_lie'] = testval_metrics['acc_lie_lie']\n",
-    "rs['testval_metrics'] = rs['test']"
+    "rs['testval_metrics'] = rs['test']\n"
    ]
   },
   {