fix bugs in cosine_match.py

Linmj-Judy · web-flow · commit c282b66c1a55 · 2024-03-11T18:03:25.000+08:00
diff --git a/evals/elsuite/cosine_match.py b/evals/elsuite/cosine_match.py
@@ -264,7 +264,7 @@ def eval_sample(self, test_sample, rng):
         )
         evals.record.record_metrics(
             value_recall=sum(matches)/len(matches) if matches else 0.0,
-            f1_score=utils.macro_f1_score_3(self.word2vec_model, sample_list, correct_answers),
+            f1_score=utils.macro_f1_score_2(self.word2vec_model, sample_list, correct_answers),
         )
         
 
@@ -334,7 +334,7 @@ def eval_sample(self, test_sample, rng):
         )
         evals.record.record_metrics(
             value_recall = sum(matches)/len(matches) if matches else 0.0,
-            f1_score = utils.cos_f1_score(self.word2vec_model, sample_list, correct_answers),
+            f1_score = utils.macro_f1_score_3(self.word2vec_model, sample_list, correct_answers),
         )
 
     def run(self, recorder: RecorderBase):

Original file line number	Diff line number	Diff line change
`@@ -264,7 +264,7 @@ def eval_sample(self, test_sample, rng):`
`264`	`264`	`)`
`265`	`265`	`evals.record.record_metrics(`
`266`	`266`	`value_recall=sum(matches)/len(matches) if matches else 0.0,`
`267`		`- f1_score=utils.macro_f1_score_3(self.word2vec_model, sample_list, correct_answers),`
	`267`	`+ f1_score=utils.macro_f1_score_2(self.word2vec_model, sample_list, correct_answers),`
`268`	`268`	`)`
`269`	`269`
`270`	`270`
`@@ -334,7 +334,7 @@ def eval_sample(self, test_sample, rng):`
`334`	`334`	`)`
`335`	`335`	`evals.record.record_metrics(`
`336`	`336`	`value_recall = sum(matches)/len(matches) if matches else 0.0,`
`337`		`- f1_score = utils.cos_f1_score(self.word2vec_model, sample_list, correct_answers),`
	`337`	`+ f1_score = utils.macro_f1_score_3(self.word2vec_model, sample_list, correct_answers),`
`338`	`338`	`)`
`339`	`339`
`340`	`340`	`def run(self, recorder: RecorderBase):`