set random seeds in the alpaca run script (#268)

AmpereComputingAI · Oct 17, 2024 · 1237c0e · 1237c0e
1 parent eac6bda
commit 1237c0e
Show file tree

Hide file tree

Showing 2 changed files with 5 additions and 1 deletion.
diff --git a/natural_language_processing/text_generation/alpaca/run.py b/natural_language_processing/text_generation/alpaca/run.py
@@ -34,6 +34,10 @@ def run_single_pass(pytorch_runner, _dataset):
         response = decode(outputs[:, inputs.input_ids.shape[1]:])
         _dataset.submit_prediction(response)
 
+    import numpy as np
+    import torch
+    np.random.seed(44)
+    torch.manual_seed(44)
     model = AutoModelForCausalLM.from_pretrained(model_path)
     if use_torch_fp16:
         model = model.half()

diff --git a/tests/test_pytorch_models.py b/tests/test_pytorch_models.py
@@ -93,7 +93,7 @@ def test_alpaca(self):
         def wrapper(**kwargs):
             kwargs["q"].put(run_pytorch_fp32(**kwargs)[0])
 
-        exact_match_ref, f1_ref = 0.180, 0.548
+        exact_match_ref, f1_ref = 0.220, 0.547
         acc = run_process(wrapper, {"model_path": self.model_path, "batch_size": 1, "num_runs": 50,
                                     "timeout": None, "dataset_path": self.dataset_path})
         self.assertTrue(acc["exact_match"] / exact_match_ref > 0.95)