File size: 2,792 Bytes
43b4299 | 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 | """Regression checks for the corpus's evidence boundaries and error cases."""
import unittest
from export_corpus import Redactor, normalize, parse_payload, telemetry, identity
class ExportTests(unittest.TestCase):
def test_malformed_jsonl_preserves_good_lines_and_coordinate(self):
rows, errors = parse_payload(b'{"type":"text","part":{"text":"hello"}}\nBAD\n{"type":"error"}\n', True)
self.assertEqual([r["source_line"] for r in rows], [1, 3])
self.assertEqual(errors[0]["line"], 2)
events = normalize(rows, "opencode-events-jsonl", "episode", "view")
self.assertEqual(events[0]["source_line"], 1)
def test_rejected_response_only_recording(self):
obj = {"messages": [{"role": "user", "content": "parser failed", "extra": {"model_response_summary": {"content": "rejected generation"}}}]}
events = normalize(obj, "mini-swe-agent-1.1", "episode", "view")
self.assertEqual(sum(e["rejected_response"] for e in events), 1)
self.assertEqual(events[-1]["text"], "rejected generation")
self.assertFalse(any(e["kind"] == "assistant" for e in events))
def test_token_alignment_and_unavailable_data(self):
good = telemetry({"completion_token_ids": [10, 20], "logprobs": [-1.0, -3.0]})
self.assertEqual(good["surprisal_nats"], 2.0)
self.assertEqual(telemetry({"completion_token_ids": [10], "logprobs": [-1,-2]})["alignment"], "invalid_or_misaligned")
self.assertEqual(telemetry({"completion_token_ids": [10], "logprobs": [float('nan')]})["alignment"], "invalid_or_misaligned")
self.assertEqual(telemetry({})["alignment"], "unavailable")
def test_redact_auth_preserve_telemetry_and_truncation(self):
redactor = Redactor()
obj = {"api_key": "private", "env": {"OPENAI_API_KEY": "private"},
"completion_token_ids": [1,2], "completion_tokens": 2,
"truncated": True, "text": "Authorization: Bearer abcdefghijklmnop"}
cleaned = redactor.clean(obj, "fixture")
self.assertNotIn("api_key", cleaned)
self.assertEqual(cleaned["env"], {})
self.assertEqual(cleaned["completion_token_ids"], [1,2])
self.assertTrue(cleaned["truncated"])
self.assertNotIn("abcdefghijklmnop", cleaned["text"])
def test_copies_do_not_merge_different_episode_keys(self):
self.assertNotEqual(identity("trial-A"), identity("trial-B"))
self.assertEqual(identity("trial-A"), identity("trial-A"))
def test_empty_recording_has_no_inferred_failure(self):
obj, errors = parse_payload(b'')
self.assertIsNone(obj)
self.assertTrue(errors)
self.assertEqual(normalize(None, "ATIF-v1.7", "episode", "view"), [])
if __name__ == "__main__":
unittest.main()
|