Create readme.md

readme.md

@@ -0,0 +1,13 @@
+# Filtering
+1) selected_incorrect: filter to Python3 family + WRONG_ANSWER; spool to tmp jsonl + collect problem_ids
+2) selected_accepted: keep first correct reference per problem_id in the above set
+3) verifiable: keep rows whose id is in the above set
+4) Join & save Parquet with requested columns; apply final filters
+
+```
+[Stage 1] selected_incorrect processed=18097 kept_after_filters=8363 unique_problem_ids=1778
+[Stage 2] selected_accepted processed=42467 matched_problem_ids=1778
+[Stage 3] verifiable processed=8760 matched_problem_ids=1693
+[Stage 4] incorrect_tmp_rows=8363 joined_ok=8073 dropped_no_correct=0 dropped_no_verifiable=290
+[Stage 4] before_final_filters=8073 dropped_interaction_present=3 dropped_rating=0 dropped_input_mode=86 dropped_executable=0 final_written=7984
+```