⚗️ Synthetic KG: add filtering in scoring step

lib/synth-kg/training_steps/score/run.py

@@ -1,5 +1,6 @@
 import argparse
 
+import numpy as np
 import pandas as pd
 from sentence_transformers import SentenceTransformer
 from sentence_transformers.util import cos_sim
@@ -73,10 +74,17 @@ def main():
     final_dataset["chosen"] = public_dataset.apply(
         lambda row: row[f"response_{best_response_idx[row.name]}"], axis=1
     )
+    final_dataset["chosen_score"] = public_dataset.apply(
+        lambda row: row[f"similarity_score_{best_response_idx[row.name]}"], axis=1
+    )
     final_dataset["rejected"] = public_dataset.apply(
         lambda row: row[f"response_{worst_response_idx[row.name]}"], axis=1
     )
 
+    percentile = np.percentile(final_dataset["chosen_score"], 65)
+    final_dataset = final_dataset.loc[final_dataset["chosen_score"] > percentile]
+
+    final_dataset = final_dataset[final_dataset["chosen"].apply(lambda x: len(x.split()) >= 20)]
     final_dataset.to_parquet(
         f"{args.output_path}/model={args.evaluator_path.replace('/','-')}_dpo.parquet"
     )