Fixed cleanup + increased translation

src/backend/run_eval_suite_lighteval.py

@@ -72,6 +72,9 @@ def run_evaluation(eval_request: EvalRequest, task_names: str, batch_size: int,
         args.model_config['model']['base_params']['reuse_existing'] = True
         model_config = create_model_config(args=args, accelerator=accelerator)
         model, _ = load_model(config=model_config, env_config=env_config)
+        print('Cleaning up')
+        model.reuse_existing = False # force it to clean up
         model.cleanup()
+        results = None
 
     return results

src/custom_tasks/translation_task.py

@@ -52,5 +52,5 @@ translation_task = LightevalTaskConfig(
     evaluation_splits=["en2he", "he2en"],
     metric=['sentence_bleu'],
     stop_sequence=['\n'],
-    generation_size=220
+    generation_size=256
 )