File size: 14,434 Bytes
f996af6 | 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 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 | import copy
import json
from pathlib import Path
import pytest
from odu_eval import runner
from odu_eval.cli import main
from odu_eval.core import scoring, scene, transport
from odu_eval.dataset import Dataset
from odu_eval.io import file_digest, read_jsonl, write_json, write_jsonl
from odu_eval.parsing import api, local
from odu_eval.providers import APIError, Service
@pytest.fixture
def dataset(tmp_path):
root = tmp_path / "data"
root.mkdir()
rows = []
for index, (modality, label, clips) in enumerate([
("audio-only", "positive", 1), ("audio-only", "negative", 2),
("audio-visual", "positive", 2), ("audio-visual", "negative", 1),
]):
sid = f"example_{index}"
folder = root / modality / label / sid
folder.mkdir(parents=True)
ext = "wav" if modality == "audio-only" else "mp4"
media = []
for clip in range(1, clips + 1):
name = f"clip_v{clip}.{ext}"
(folder / name).write_bytes(b"fixture-media-" + str(clip).encode())
media.append(name)
seg = {"start_ms": 0, "end_ms": 1000, "transcript": "Turn down the volume.",
"structured_intent": "Reduce the volume.", "required_context": [],
"user_gender": "unknown", "user_age_group": "adult",
"key_points": [{"qid": "kp01", "point": "Reduce volume", "tiers": ["T1"], "source": "intent"},
{"qid": "kp02", "point": "A television", "tiers": ["T1", "T2"], "source": "dialogue_history"},
{"qid": "kp03", "point": "Background music", "tier": 2, "source": "audio"}]}
if modality == "audio-visual":
seg["user_in_frame"] = False
write_json(folder / "annotation_en.json", {"scene_id": sid, "evaluation_targets": {
"has_demand": label == "positive", "segments": [seg] if label == "positive" else []}})
write_json(folder / "metadata.json", {"scene_id": sid, "agent_replies": [
{"after_clip_id": "v0", "content": "excluded before first clip"},
{"after_clip_id": "v1", "content": "Earlier assistant reply"},
{"after_clip_id": "v2", "content": "excluded after final clip"}]})
rows.append({"scene_id": sid, "modality": modality, "label": label, "source": "generated", "n_clips": clips,
"rel_path": folder.relative_to(root).as_posix(), "media": media, "metadata": "metadata.json",
"annotations": {"en": "annotation_en.json"}, "annotation": "annotation_en.json"})
write_jsonl(root / "manifest.jsonl", rows)
write_jsonl(root / "subsets/main.jsonl", rows)
sums = [f"{file_digest(root / r['rel_path'] / name)} {r['rel_path']}/{name}" for r in rows for name in r["media"]]
(root / "checksums.sha256").write_text("\n".join(sums) + "\n")
return Dataset(root)
def test_inputs_keep_multiclip_order_without_reference_answer(dataset):
assert dataset.validate(checksums=True)["checksums_verified"] == 6
sample = dataset.sample(dataset.rows[2])
assert [m["role"] for m in sample["messages"]] == ["system", "user"]
parts = sample["messages"][1]["content"]
assert [p["clip_id"] for p in parts if p["type"] == "video"] == ["v1", "v2"]
text = " ".join(p["text"] for p in parts if p["type"] == "text")
assert "Earlier assistant reply" in text
assert "excluded" not in text
assert "local timestamps within the final video clip" in text
assert "key_points" not in json.dumps(sample)
assert all(not Path(p["path"]).is_absolute() and p["use_audio"] for p in parts if p["type"] == "video")
def test_dataset_escape_and_duplicate_ids_are_rejected(dataset):
row = copy.deepcopy(dataset.rows[0])
row["media"] = ["../../escape.wav"]
with pytest.raises(ValueError, match="inside"):
dataset.sample(row)
rows = dataset.rows + [dataset.rows[0]]
write_jsonl(dataset.root / "manifest.jsonl", rows)
with pytest.raises(ValueError, match="unique"):
Dataset(dataset.root)
def test_parsers_keep_their_historical_difference(monkeypatch):
monkeypatch.setattr(api, "_JSON_REPAIR_ENABLED", False)
tagged = "<has_demand>false</has_demand><segments></segments>"
assert api.parse_model_output(tagged) is None
parsed, mode = local.parse_model_output_with_mode(tagged)
assert mode == "tagged"
assert parsed["evaluation_targets"] == {"has_demand": False, "segments": []}
raw = '<think>reasoning</think>```json\n{"has_demand":true,"segments":[{"start_ms":"1.2s","end_ms":"2.5s"}]}\n```'
assert api.parse_model_output(raw) == local.parse_model_output(raw)
def test_final_clip_scopes_each_side_independently():
pred = {"has_demand": True, "segments": [{"clip_id": "v1"}, {"clip_id": "v3"}, {"transcript": "keep"}]}
gt = {"has_demand": True, "segments": [{"clip_id": "v2"}, {"clip_id": "v4"}]}
p, g, final = scoring.apply_multiturn_scope(pred, gt, "example")
assert final == "v4"
assert [x.get("clip_id") for x in p["segments"]] == ["v3", None]
assert g["segments"] == [{"clip_id": "v4"}]
def test_union_keypoints_and_judge_failures(dataset):
gt = dataset.gt(dataset.rows[0])["evaluation_targets"]
pred = copy.deepcopy(gt)
for seg in pred["segments"]:
seg.pop("key_points")
transport.set_llm_call(lambda **kwargs: '[{"qid":"kp01","hit":true},{"qid":"kp02","hit":true},{"qid":"kp03","hit":false}]')
try:
result = scene.evaluate_scene(pred, gt, run_m4_llm=True, audio_only=True)
assert result["scores"]["verification_qa"] == .6667
assert result["details"]["verification_qa"]["t1_hit_rate"] == 1
assert result["details"]["verification_qa"]["point_stats"] == {"hit": 2, "total": 3}
def fail(**kwargs):
raise RuntimeError("OpenAIError")
transport.set_llm_call(fail)
failed = scene.evaluate_scene(copy.deepcopy(pred), copy.deepcopy(gt), run_m4_llm=True, m4_max_retries=1, audio_only=True)
assert failed["overall"] is None
assert failed["details"]["verification_qa"]["incomplete"]
finally:
transport.set_llm_call(None)
def test_wire_protocols_keep_media_order(dataset, monkeypatch):
monkeypatch.delenv("ODU_QWEN_OMNI_MODEL", raising=False)
sample = dataset.sample(dataset.rows[2])
qwen = Service("qwen3.5-omni-plus")
body = qwen.inference_payload(sample, dataset.root)
assert set(body) == {"model", "messages", "stream", "stream_options", "modalities"}
assert body["modalities"] == ["text"] and body["stream"] is True
assert [p["type"] for p in body["messages"][1]["content"]] == [p["type"].replace("video", "video_url") for p in sample["messages"][1]["content"]]
gemini = Service("gemini-3.1-pro")
native = gemini.inference_payload(sample, dataset.root)
assert set(native) == {"systemInstruction", "contents"}
assert native["contents"][0]["role"] == "user"
assert len([p for p in native["contents"][0]["parts"] if "inlineData" in p]) == 2
assert "path" not in json.dumps(native)
audio = qwen.inference_payload(dataset.sample(dataset.rows[0]), dataset.root)
assert audio["messages"][1]["content"][0]["input_audio"]["data"].startswith("data:audio/")
assert Service("qwen3.6-flash", thinking="true").text_payload("s", "u")["enable_thinking"] is True
openai = Service("gpt-5.4", reasoning="high").text_payload("s", "u")
assert openai["reasoning_effort"] == "high" and "enable_thinking" not in openai
def test_streaming_collects_answer_only(monkeypatch):
class Response:
headers = {"Content-Type": "text/event-stream; charset=utf-8"}
def __enter__(self): return self
def __exit__(self, *args): return False
def iter_lines(self):
yield b'data: {"choices":[{"delta":{"reasoning_content":"hidden"}}]}'
yield b'data: {"choices":[{"delta":{"content":"ans"}}]}'
yield b'data: {"choices":[{"delta":{"content":"wer"}}]}'
yield b'data: {"choices":[],"usage":{"total_tokens":10}}'
yield b'data: [DONE]'
service = Service("gpt-5.4")
monkeypatch.setattr(service, "request", lambda *a, **k: Response())
assert service.call_text("s", "u") == "answer"
assert service.stats["reasoning_chunks"] == 1
def test_server_errors_do_not_echo_credentials(monkeypatch):
class Response:
ok, status_code = False, 401
def json(self): return {"error": {"code": "invalid_api_key", "message": "secret-value https://private.example/"}}
def close(self): pass
class Session:
def request(self, *a, **k): return Response()
service = Service("gpt-5.4")
service.key = "test-key"
service._session = Session()
with pytest.raises(APIError) as err:
service.request("POST", "https://example.test")
assert "401" in str(err.value)
assert "secret-value" not in str(err.value) and "private.example" not in str(err.value)
def test_gemini_files_upload_wire_contract(dataset, monkeypatch):
service = Service("gemini-3.1-pro")
media = dataset.scene_file(dataset.rows[2], dataset.rows[2]["media"][0])
calls = []
class Response:
def __init__(self, body=None, headers=None):
self.body, self.headers = body, headers or {}
def __enter__(self): return self
def __exit__(self, *args): return False
def json(self): return self.body
def request(method, url, **kwargs):
calls.append((method, url, kwargs))
if len(calls) == 1:
assert kwargs["json"] == {"file": {"display_name": "evaluation-media"}}
assert kwargs["headers"]["X-Goog-Upload-Protocol"] == "resumable"
return Response(headers={"X-Goog-Upload-URL": "https://upload.example.test/session"})
if len(calls) == 2:
assert kwargs["headers"]["X-Goog-Upload-Command"] == "upload, finalize"
assert kwargs["data"].read() == media.read_bytes()
return Response({"file": {"name": "files/example", "state": "PROCESSING"}})
return Response({"name": "files/example", "state": "ACTIVE", "uri": "https://files.example.test/example"})
monkeypatch.setattr(service, "request", request)
monkeypatch.setattr("odu_eval.providers.time.sleep", lambda n: None)
uploaded = service._upload_gemini(media, "video/mp4")
assert uploaded["state"] == "ACTIVE" and len(calls) == 3
sample = dataset.sample(dataset.rows[2])
key = next(p["path"] for p in sample["messages"][1]["content"] if p["type"] == "video")
payload = service.inference_payload(sample, dataset.root, files={key: uploaded})
files = [p["fileData"] for p in payload["contents"][0]["parts"] if "fileData" in p]
assert files == [{"mimeType": "video/mp4", "fileUri": uploaded["uri"]}]
def test_moving_dataset_keeps_inputs_and_fingerprint(dataset, tmp_path):
import shutil
destination = tmp_path / "relocated"
shutil.copytree(dataset.root, destination)
moved = Dataset(destination)
assert dataset.fingerprint(dataset.rows) == moved.fingerprint(moved.rows)
assert [dataset.sample(r) for r in dataset.rows] == [moved.sample(r) for r in moved.rows]
def test_readme_flow_and_offline_replay(dataset, tmp_path, monkeypatch):
monkeypatch.setenv("DASHSCOPE_API_KEY", "test-key")
def fake_infer(self, sample, root):
row = next(r for r in dataset.rows if r["scene_id"] == sample["scene_id"])
gt = copy.deepcopy(dataset.gt(row)["evaluation_targets"])
for seg in gt["segments"]:
seg.pop("key_points")
return json.dumps(gt)
monkeypatch.setattr(Service, "infer", fake_infer)
monkeypatch.setattr(Service, "call_text", lambda *a, **kw: '[{"qid":"kp01","hit":true},{"qid":"kp02","hit":true},{"qid":"kp03","hit":false}]')
data = ["--data-root", str(dataset.root)]
inputs, predictions = tmp_path / "inputs.jsonl", tmp_path / "model"
assert main(["validate", *data, "--checksums"]) == 0
assert main(["export-inputs", *data, "--smoke", "--output", str(inputs)]) == 0
command = ["infer", *data, "--smoke", "--model", "qwen3.5-omni-plus", "--output-dir", str(predictions)]
assert main(command) == 0
assert main(command) == 0 # Resume must not change the raw rows.
score = tmp_path / "score"
score_args = ["score", *data, "--predictions", str(predictions / "predictions.jsonl"), "--judge", "qwen3.6-flash"]
assert main([*score_args, "--output-dir", str(score)]) == 0
recorded = json.loads((score / "run_manifest.json").read_text())
assert len(recorded["judge_prompt_sha256"]) == 64
assert recorded["protocol"]["judge_prompt_sha256"] == recorded["judge_prompt_sha256"]
monkeypatch.delenv("DASHSCOPE_API_KEY")
monkeypatch.setenv("DASHSCOPE_BASE_URL", "https://different.example/v1")
monkeypatch.setattr(Service, "call_text", lambda *a, **kw: pytest.fail("offline sent a request"))
replay = tmp_path / "replay"
assert main([*score_args, "--judge-cache", str(score / "judge_cache"), "--offline", "--output-dir", str(replay)]) == 0
assert read_jsonl(score / "scores.jsonl") == read_jsonl(replay / "scores.jsonl")
assert (score / "comparison.csv").read_bytes() == (replay / "comparison.csv").read_bytes()
assert str(dataset.root) not in (score / "run_manifest.json").read_text()
assert main(["aggregate", "--runs", str(score), str(replay), "--output-dir", str(tmp_path / "report")]) == 0
def test_custom_submission_and_coverage(dataset, tmp_path, monkeypatch):
monkeypatch.setenv("DASHSCOPE_API_KEY", "test-key")
monkeypatch.setattr(Service, "call_text", lambda *a, **k: pytest.fail("negative answer should not require a judge"))
inputs, raw, pred = tmp_path / "inputs.jsonl", tmp_path / "raw.jsonl", tmp_path / "predictions.jsonl"
runner.export_inputs(dataset, dataset.rows, inputs)
write_jsonl(raw, [{"scene_id": dataset.rows[0]["scene_id"], "raw_output": '{"has_demand":false,"segments":[]}'}])
assert main(["parse", "--input", str(raw), "--inputs", str(inputs), "--output", str(pred), "--profile", "local"]) == 0
out = tmp_path / "partial"
assert main(["score", "--data-root", str(dataset.root), "--predictions", str(pred), "--judge", "qwen3.6-flash", "--output-dir", str(out)]) == 1
coverage = json.loads((out / "coverage.json").read_text())
assert coverage["expected"] == 4 and coverage["missing_predictions"] == 3
|