Datasets:
File size: 4,102 Bytes
0bf2389 | 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 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 | from __future__ import annotations
import json
import unittest
from collections import Counter
from pathlib import Path
from barbet_bgd.generation import GenerationPlan, PRIMITIVES, _long_bucket, generate_record
from barbet_bgd.serializers import build_messages, serialize_target
ROOT = Path(__file__).resolve().parents[1]
class GenerationTests(unittest.TestCase):
@classmethod
def setUpClass(cls):
cls.config = json.loads((ROOT / "configs/selected_sft_1m.json").read_text(encoding="utf-8"))
cls.plan = GenerationPlan(cls.config)
def test_one_cycle_has_exact_configured_mixture(self):
counts = Counter(self.plan.position(i).primitive for i in range(100))
self.assertEqual(counts, Counter(self.config["primitive_percent"]))
mixtures = Counter(self.plan.position(i).mixture_category for i in range(100))
self.assertEqual(mixtures, Counter(self.config["mixture_percent"]))
serializers = Counter(self.plan.position(i).serializer for i in range(100))
self.assertEqual(serializers, Counter(self.config["serializer_percent"]))
self.assertEqual(sum(self.plan.position(i).is_long_context for i in range(100)), 5)
def test_generation_is_deterministic(self):
for index in (0, 1, 19, 20, 99, 100, 979_999, 980_000, 989_999, 990_000):
self.assertEqual(
generate_record(self.config, self.plan, index),
generate_record(self.config, self.plan, index),
)
def test_generation_does_not_depend_on_json_key_order(self):
sorted_config = json.loads(
json.dumps(self.config, ensure_ascii=False, sort_keys=True)
)
sorted_plan = GenerationPlan(sorted_config)
for index in (0, 1, 99, 100, 500_000, 990_000):
self.assertEqual(self.plan.position(index), sorted_plan.position(index))
self.assertEqual(
generate_record(self.config, self.plan, index),
generate_record(sorted_config, sorted_plan, index),
)
def test_split_boundaries_do_not_break_source_group(self):
self.assertEqual(self.plan.position(979_999).split, "train")
self.assertEqual(self.plan.position(980_000).split, "validation")
self.assertEqual(self.plan.position(989_999).split, "validation")
self.assertEqual(self.plan.position(990_000).split, "test")
before = generate_record(self.config, self.plan, 979_999)["source_group"]
after = generate_record(self.config, self.plan, 980_000)["source_group"]
self.assertNotEqual(before, after)
def test_all_primitives_build_structured_targets(self):
seen = set()
for index in range(300):
record = generate_record(self.config, self.plan, index)
seen.add(record["primitive"])
self.assertIn("decision", record["target"])
self.assertIn("evidence", record["target"])
self.assertIn("confidence_target", record["target"])
self.assertIn("action", record["target"])
self.assertEqual(record["language"][0], "zh-TW")
self.assertEqual(record["locale"], "zh-TW")
self.assertEqual(seen, set(PRIMITIVES))
def test_each_hundred_long_records_has_exact_length_mix(self):
counts = Counter(_long_bucket(i, 50_000)[0] for i in range(100))
self.assertEqual(
counts,
Counter({"0.5K-4K": 75, "4K-16K": 18, "16K-64K": 6, "64K-128K": 1}),
)
def test_serializers_are_controlled_views(self):
target = generate_record(self.config, self.plan, 0)["target"]
for serializer in ("JSON", "FUNCTION_CALL", "COMPACT_TAGS", "KEY_VALUE"):
rendered = serialize_target(target, serializer)
self.assertIn("decision", rendered)
self.assertIn(str(target["decision"]), rendered)
messages = build_messages(generate_record(self.config, self.plan, 0))
self.assertEqual([item["role"] for item in messages], ["system", "user", "assistant"])
if __name__ == "__main__":
unittest.main()
|