Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

fix(studio): rewrite S3 audio paths in single-record JSON manifests (#16631)

Max-Reisinger14bd64f264cf
diff --git a/studio/backend/core/training/s3_dataset.py b/studio/backend/core/training/s3_dataset.py
index c2de23c3fad..c358bf9c911 100644
--- a/studio/backend/core/training/s3_dataset.py
+++ b/studio/backend/core/training/s3_dataset.py
@@ -289,9 +289,11 @@ def _rewrite_json_manifest(manifest_path: str, resolve) -> None:
for row in rows:
f.write(json.dumps(row, ensure_ascii = False) + "\n")
return
- if not isinstance(data, list):
- return # column-oriented JSON
- for row in data:
+ # A single JSON Lines record also parses as a complete JSON object.
+ # Keep its object shape while applying the same rewrite as multi-row manifests.
+ if not isinstance(data, (dict, list)):
+ return
+ for row in [data] if isinstance(data, dict) else data:
_rewrite_row(row, resolve)
with open(manifest_path, "w", encoding = "utf-8") as f:
json.dump(data, f, ensure_ascii = False)
diff --git a/studio/backend/tests/test_s3_audio_dataset.py b/studio/backend/tests/test_s3_audio_dataset.py
index 25eb96fb76c..e7040390401 100644
--- a/studio/backend/tests/test_s3_audio_dataset.py
+++ b/studio/backend/tests/test_s3_audio_dataset.py
@@ -183,6 +183,40 @@ def test_json_lines_in_a_json_manifest_are_rewritten(monkeypatch, tmp_path):
]
[email protected]("layout", ["object", "jsonl", "array"])
[email protected]("audio_value", ["a.wav", {"path": "a.wav", "bytes": None}])
+def test_single_record_manifest_rewrites_audio(monkeypatch, tmp_path, layout, audio_value):
+ row = {"audio": audio_value, "text": "hello", "speaker": "speaker-1"}
+ suffix = ".jsonl" if layout == "jsonl" else ".json"
+ key = "datasets/train" + suffix
+ payload = [row] if layout == "array" else row
+ _install(
+ monkeypatch,
+ [key, "datasets/a.wav"],
+ contents = {key: json.dumps(payload) + "\n"},
+ )
+ files = s3_dataset.download_s3_dataset(_cfg(), dest_dir = str(tmp_path))
+ loaded = json.loads(Path(files[0]).read_text(encoding = "utf-8"))
+ result = loaded[0] if layout == "array" else loaded
+ local_audio = str(tmp_path / "a.wav")
+ expected_audio = (
+ {"path": local_audio, "bytes": None} if isinstance(audio_value, dict) else local_audio
+ )
+ assert result == {**row, "audio": expected_audio}
+ assert Path(local_audio).is_file()
+
+
+def test_json_manifest_non_record_layout_is_unchanged(tmp_path):
+ # A column-oriented object of indexed columns is not a row of audio paths.
+ payload = {"audio": {"0": "a.wav"}, "text": {"0": "hello"}}
+ path = tmp_path / "train.json"
+ path.write_text(json.dumps(payload), encoding = "utf-8")
+ s3_dataset._rewrite_json_manifest(
+ str(path), lambda value: "/local/a.wav" if value == "a.wav" else None
+ )
+ assert json.loads(path.read_text(encoding = "utf-8")) == payload
+
+
def test_manifest_relative_references_resolve_against_the_manifest_dir(monkeypatch, tmp_path):
_install(
monkeypatch,