Datasets:
File size: 3,301 Bytes
d360720 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 | """Run with: python -m unittest discover -s openjudgment -p test_v4_helpsteer.py"""
from copy import deepcopy
import json
import unittest
from v4_helpsteer import adapt
def example(votes):
return dict(context=[dict(role="user", content="Compare these answers.")],
response1="First response", response2="Second response",
overall_preference=2,
individual_preference=[dict(score=v, reasoning="SECRET_RATIONALE",
feedback1="SECRET_FEEDBACK", feedback2="SECRET_FEEDBACK") for v in votes])
class HelpSteerTests(unittest.TestCase):
def test_feedback_rubric_and_strict_ambiguity(self):
source = example([1])
source["feedback1"] = ["The response is not helpful. SECRET", "The response is perfectly helpful. SECRET"]
source["feedback2"] = ["The response is mostly helpful. SECRET", "I think it is mostly helpful."]
task, = adapt(source, "feedback")
self.assertEqual(task["target"], [0.5, 0, 0, 0, 0.5])
self.assertNotIn("SECRET", task["state"])
self.assertEqual(set(json.loads(task["state"])), {"context", "response"})
source["feedback1"] = ["The response is partially helpful."]
self.assertEqual(adapt(source, "feedback"), [])
def test_boundaries_and_sign(self):
for score in range(-3, 4):
task, = adapt(example([score]))
self.assertEqual(task["target"], [float(i == score) for i in range(-3, 4)])
self.assertIn("Response 1 is much better", adapt(example([-3]))[0]["candidates"][0])
self.assertIn("Response 2 is much better", adapt(example([3]))[0]["candidates"][-1])
def test_disagreement_and_swap(self):
source = example([-3, 0, 1])
original = deepcopy(source)
task, = adapt(source)
self.assertEqual(task["target"], [1/3, 0, 0, 1/3, 1/3, 0, 0])
self.assertEqual(source, original)
swapped = deepcopy(source)
swapped["response1"], swapped["response2"] = source["response2"], source["response1"]
for vote in swapped["individual_preference"]:
vote["score"] *= -1
other, = adapt(swapped)
self.assertEqual(other["target"], task["target"][::-1])
self.assertEqual(other["group_key"], task["group_key"])
def test_no_label_leakage_and_no_aggregate_fallback(self):
task, = adapt(example([-1, -1, 3]))
self.assertEqual(set(json.loads(task["state"])), {"context", "response1", "response2"})
self.assertNotIn("SECRET", task["state"] + task["instructions"])
self.assertNotIn("overall_preference", task["state"])
self.assertEqual(adapt(example([])), [])
for bad in [4, -4, True, 1.0, "1", None]:
self.assertEqual(adapt(example([-1, bad])), [])
def test_context_normalization_preserves_semantics(self):
source = example([0])
source["context"][0]["content"] = "Code:\r\n Pass"
a, = adapt(source)
source["context"][0]["content"] = "Code:\n Pass"
b, = adapt(source)
self.assertEqual(a["group_key"], b["group_key"])
source["context"][0]["content"] = "code:\n pass"
self.assertNotEqual(a["group_key"], adapt(source)[0]["group_key"])
if __name__ == "__main__":
unittest.main()
|