1 files changed, 30 insertions, 10 deletions

diff --git a/test_prompts.sh b/test_prompts.sh
index 04020c3..e22e90f 100755
--- a/test_prompts.sh
+++ b/test_prompts.sh
@@ -1,8 +1,27 @@
 #!/bin/sh
 
-# Default filenames
-PROMPT_FILE="${1:-test_prompts.txt}"
-RESULT_FILE="${2:-results.txt}"
+# Usage: ./test_prompts.sh <model_path> [prompt_file]
+
+# Mandatory Model Argument
+MODEL_PATH="$1"
+
+if [ -z "$MODEL_PATH" ]; then
+    echo "Usage: $0 <model_path> [prompt_file]" >&2
+    exit 1
+fi
+
+if [ ! -f "$MODEL_PATH" ]; then
+    echo "Error: Model file '$MODEL_PATH' not found." >&2
+    exit 1
+fi
+
+# Derive base model name and output result filename
+# e.g., models/qwen2.5-7b-q4.gguf -> results_qwen2.5-7b-q4.gguf.txt
+MODEL_NAME=$(basename "$MODEL_PATH")
+RESULT_FILE="results_${MODEL_NAME}.txt"
+
+# Default prompt filename (optional 2nd argument)
+PROMPT_FILE="${2:-test_prompts.txt}"
 EXECUTABLE="./lex"
 
 # Ensure the executable exists and is runnable
@@ -28,15 +47,16 @@ fi
 # Create (or truncate) the result file at the start of the run
 > "$RESULT_FILE"
 
-# Create a 70-character horizontal rule: " -- " + 66 "-"
-RULE=" -- "
+# Create horizontal rule.
+RULE=""
 i=0
-while [ $i -lt 66 ]; do
+while [ $i -lt 72 ]; do
     RULE="${RULE}-"
     i=$((i + 1))
 done
 
-echo "Starting evaluation: $total_prompts prompt(s) found."
+echo "Starting evaluation with model: '$MODEL_PATH'"
+echo "Evaluating $total_prompts prompt(s) from '$PROMPT_FILE'"
 echo "Writing results to: '$RESULT_FILE'"
 echo "---------------------------------------------------"
 
@@ -66,12 +86,12 @@ while IFS= read -r prompt || [ -n "$prompt" ]; do
     # Log prompt header to the result file
     printf "PROMPT: %s\n\n" "$prompt" >> "$RESULT_FILE"
 
-    # On OpenBSD, /usr/bin/time -lp outputs POSIX timings and rusage stats to stderr
-    /usr/bin/time -lp "$EXECUTABLE" "$prompt" >> "$RESULT_FILE" 2> "$TIME_LOG"
+    # Invoke executable with model path and prompt
+    /usr/bin/time -lp "$EXECUTABLE" "$MODEL_PATH" "$prompt" >> "$RESULT_FILE" 2> "$TIME_LOG"
 
     # Extract execution time and peak memory (maximum resident set size) from OpenBSD time output
     real_time=$(awk '/^real/ {print $2}' "$TIME_LOG")
-	max_rss_mb=$(awk '/maximum resident set size/ {printf "%.2f", $1/1024}' "$TIME_LOG")
+    max_rss_mb=$(awk '/maximum resident set size/ {printf "%.2f", $1/1024}' "$TIME_LOG")
 
     # Format metrics summary line
     metrics_summary="ELAPSED: ${real_time}s | MAX RSS: ${max_rss_mb} MB"