get _RES from CI and change atol to 0.0001

tests/models/test_huggingface.py

@@ -20,46 +20,46 @@ class Test_HFLM:
     ROLLING: list[Instance] = rolling_task.instances
 
     MULTIPLE_CH_RES = [
-        (-41.905879974365234, False),
-        (-42.93785095214844, False),
-        (-33.9145393371582, False),
-        (-37.07110595703125, False),
-        (-22.954187393188477, False),
-        (-20.342954635620117, False),
-        (-14.816370010375977, False),
-        (-27.94381332397461, False),
-        (-15.806619644165039, False),
-        (-15.937178611755371, False),
-        (-13.052162170410156, False),
-        (-18.04889678955078, False),
-        (-13.346054077148438, False),
-        (-13.367782592773438, False),
-        (-12.128646850585938, False),
-        (-11.871688842773438, False),
-        (-47.10654067993164, False),
-        (-47.76068115234375, False),
-        (-36.44114303588867, False),
-        (-50.02851104736328, False),
-        (-16.719867706298828, False),
-        (-18.537654876708984, False),
-        (-26.469972610473633, False),
-        (-20.356552124023438, False),
-        (-17.75723648071289, False),
-        (-21.8068790435791, False),
-        (-33.19971466064453, False),
-        (-39.2862434387207, False),
-        (-14.762389183044434, False),
-        (-16.75531005859375, False),
-        (-11.486998558044434, False),
-        (-15.421247482299805, False),
-        (-13.157613754272461, False),
-        (-15.88864517211914, False),
-        (-15.287158012390137, False),
-        (-12.339122772216797, False),
-        (-44.59400177001953, False),
-        (-55.40974807739258, False),
-        (-52.697017669677734, False),
-        (-56.252601623535156, False),
+        -41.902435302734375,
+        -42.939308166503906,
+        -33.914180755615234,
+        -37.07139205932617,
+        -22.95258331298828,
+        -20.342208862304688,
+        -14.818366050720215,
+        -27.942853927612305,
+        -15.80704116821289,
+        -15.936427116394043,
+        -13.052018165588379,
+        -18.04828453063965,
+        -13.345029830932617,
+        -13.366025924682617,
+        -12.127134323120117,
+        -11.872495651245117,
+        -47.10598373413086,
+        -47.76410675048828,
+        -36.4406852722168,
+        -50.0289421081543,
+        -16.72093963623047,
+        -18.535587310791016,
+        -26.46993637084961,
+        -20.355995178222656,
+        -17.757919311523438,
+        -21.80595588684082,
+        -33.1990852355957,
+        -39.28636932373047,
+        -14.759679794311523,
+        -16.753942489624023,
+        -11.486852645874023,
+        -15.42177677154541,
+        -13.15798282623291,
+        -15.887393951416016,
+        -15.28614616394043,
+        -12.339089393615723,
+        -44.59441375732422,
+        -55.40888214111328,
+        -52.70050811767578,
+        -56.25089645385742,
     ]
     GREEDY_UNTIL_RES = [
         " The average of $2.50 each is $",
@@ -89,8 +89,9 @@ class Test_HFLM:
 
     def test_logliklihood(self) -> None:
         res = self.LM.loglikelihood(self.MULTIPLE_CH)
-        _RES, _res = [r[0] for r in self.MULTIPLE_CH_RES], [r[0] for r in res]
-        assert np.allclose(_res, _RES, atol=1e-2)
+        _RES, _res = self.MULTIPLE_CH_RES, [r[0] for r in res]
+        # change atol in case of consistent failure
+        assert np.allclose(_res, _RES, atol=1e-4)
         # check indices for Multiple Choice
         argmax_RES, argmax_res = np.argmax(
             np.array(_RES).reshape(-1, 4), axis=1