Don't use buffer, fix anomalies

tutorials/neuronpedia/make_batch.py

@@ -4,13 +4,11 @@
 SAE_PATH = sys.argv[1]
 MODEL_ID = sys.argv[2]
 SAE_ID = sys.argv[3]
-LEFT_BUFFER = int(sys.argv[4])
-RIGHT_BUFFER = int(sys.argv[5])
-N_BATCHES_SAMPLE = int(sys.argv[6])
-N_PROMPTS_SELECT = int(sys.argv[7])
-FEATURES_AT_A_TIME = int(sys.argv[8])
-START_BATCH_INCLUSIVE = int(sys.argv[9])
-END_BATCH_INCLUSIVE = int(sys.argv[10])
+N_BATCHES_SAMPLE = int(sys.argv[4])
+N_PROMPTS_SELECT = int(sys.argv[5])
+FEATURES_AT_A_TIME = int(sys.argv[6])
+START_BATCH_INCLUSIVE = int(sys.argv[7])
+END_BATCH_INCLUSIVE = int(sys.argv[8])
 
 NP_OUTPUT_FOLDER = "../../neuronpedia_outputs"
 
@@ -23,8 +21,6 @@
     n_batches_to_sample_from=N_BATCHES_SAMPLE,
     n_prompts_to_select=N_PROMPTS_SELECT,
     n_features_at_a_time=FEATURES_AT_A_TIME,
-    buffer_tokens_left=LEFT_BUFFER,
-    buffer_tokens_right=RIGHT_BUFFER,
     start_batch_inclusive=START_BATCH_INCLUSIVE,
     end_batch_inclusive=END_BATCH_INCLUSIVE,
 )

tutorials/neuronpedia/make_features.sh

@@ -8,55 +8,39 @@ echo "===== This takes input of one SAE directory at a time."
 echo "===== Features will be output into ./neuronpedia_outputs/{model}_{hook_point}_{d_sae}/batch-{batch_num}.json"
 
 echo ""
-echo "(Step 1 of 10)"
+echo "(Step 1 of 8)"
 echo "What is the absolute, full local file path to your SAE's directory (with cfg.json, sae_weights.safetensors, sparsity.safetensors)?"
 read saepath
 # TODO: support huggingface directories
 
 echo ""
-echo "(Step 2 of 10)"
+echo "(Step 2 of 8)"
 echo "What's the model ID? This must exactly match (including casing) the model ID you created on Neuronpedia."
 read modelid
 
 echo ""
-echo "(Step 3 of 10)"
+echo "(Step 3 of 8)"
 echo "What's the SAE ID?"
 echo "This was set when you did 'Add SAEs' on Neuronpedia. This must exactly match that ID (including casing). It's in the format [abbrev hook name]-[abbrev author name], like res-jb."
 read saeid
 
 echo ""
-echo "(Step 4 of 10)"
+echo "(Step 4 of 8)"
 echo "How many features are in this SAE?"
 read numfeatures
 
 echo ""
-echo "(Step 5 of 10)"
+echo "(Step 5 of 8)"
 read -p "How many features do you want generate per batch file? More requires more RAM. (default: 128): " perbatch
 [ -z "${perbatch}" ] && perbatch='128'
 
 echo ""
-echo "(Step 6 of 10)"
-echo "For each activating text sequence, how many tokens to the LEFT of the top activating token do you want?"
-echo "If your text sequences are 128 tokens long, then you might put 64. (default: 64)"
-read leftbuffer
-[ -z "${leftbuffer}" ] && leftbuffer='64'
-
-echo ""
-echo "(Step 7 of 10)"
-echo "For each activating text sequence, how many tokens to the RIGHT of the top activating token do you want?"
-echo "Left Buffer + Right Buffer must be < Total Text Length - 1"
-echo "For example, text sequences of 128 can have at most buffers of 64 + 62 = 126"
-echo "If your text sequences are 128 tokens long, then you might put 62. (default: 62)"
-read rightbuffer
-[ -z "${rightbuffer}" ] && rightbuffer='62'
-
-echo ""
-echo "(Step 8 of 10)"
+echo "(Step 6 of 8)"
 read -p "Enter number of batches to sample from (default: 4096): " batches
 [ -z "${batches}" ] && batches='4096'
 
 echo ""
-echo "(Step 9 of 10)"
+echo "(Step 7 of 8)"
 read -p "Enter number of prompts to select from (default: 24576): " prompts
 [ -z "${prompts}" ] && prompts='24576'
 
@@ -65,7 +49,7 @@ numbatches=$(expr $numfeatures / $perbatch)
 echo "===== INFO: We'll generate $numbatches batches of $perbatch features per batch = $numfeatures total features"
 
 echo ""
-echo "(Step 10 of 10)"
+echo "(Step 8 of 8)"
 read -p "Do you want to resume from a specific batch number? Enter 1 to start from the beginning (default: 1): " startbatch
 [ -z "${startbatch}" ] && startbatch='1'