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.

Studio: load S3 audio datasets by downloading the audio beside its manifest (#10050)

Shubhankar Tripathy95421ede6888
diff --git a/studio/backend/core/training/s3_dataset.py b/studio/backend/core/training/s3_dataset.py
index 91fd91b40f0..3d057fa90e4 100644
--- a/studio/backend/core/training/s3_dataset.py
+++ b/studio/backend/core/training/s3_dataset.py
@@ -9,6 +9,10 @@ to a local temp directory so the existing local-file dataset path can consume
them. boto3 is an optional dependency and is imported lazily — callers should
gate on :func:`boto3_available` before invoking the loader.
+Audio keys download too, keeping their prefix-relative layout, and audio values in
+JSON/JSONL/CSV manifests are rewritten to the local paths (``datasets.Audio`` opens path
+strings verbatim). Parquet manifests are not rewritten: their paths must already resolve.
+
The S3 config dict mirrors ``models.training.S3Config.model_dump()`` (snake_case
keys): bucket, region, prefix, access_key_id, secret_access_key, use_iam_role.
Credentials are read once to build the client and never logged or persisted.
@@ -22,12 +26,15 @@ from core.training.account_jobs import (
require_explicit_credentials,
)
from utils.paths import ensure_dir, tmp_root
+import csv
+import json
import logging
import os
import shutil
import tempfile
from importlib.util import find_spec
from typing import Callable, Optional
+from utils.datasets.format_detection import _AUDIO_EXTENSIONS
from utils.paths.path_utils import drop_shadowed_appledouble_names
logger = logging.getLogger(__name__)
@@ -41,6 +48,7 @@ _IGNORED_METADATA_FILENAMES = {
"schema.json",
"state.json",
}
+_REWRITABLE_MANIFEST_EXTENSIONS = (".json", ".jsonl", ".csv")
class S3DownloadCancelled(RuntimeError):
@@ -94,7 +102,7 @@ def _build_s3_client(s3_config: dict):
def _list_dataset_keys(client, bucket: str, prefix: Optional[str]) -> list[str]:
- """List object keys under ``prefix`` that have a supported data extension."""
+ """Supported dataset keys under ``prefix``: manifests and the audio files beside them."""
paginator = client.get_paginator("list_objects_v2")
list_kwargs = {"Bucket": bucket}
if prefix:
@@ -108,7 +116,7 @@ def _list_dataset_keys(client, bucket: str, prefix: Optional[str]) -> list[str]:
continue
if os.path.basename(key).lower() in _IGNORED_METADATA_FILENAMES:
continue
- if key.lower().endswith(SUPPORTED_EXTENSIONS):
+ if key.lower().endswith(SUPPORTED_EXTENSIONS + _AUDIO_EXTENSIONS):
keys.append(key)
# A Mac sync uploads Finder metadata under the shard's own extension and
# _validate_single_extension_family cannot see it, so a key is dropped only when the object it
@@ -116,6 +124,10 @@ def _list_dataset_keys(client, bucket: str, prefix: Optional[str]) -> list[str]:
return drop_shadowed_appledouble_names(keys)
+def _is_audio_key(key: str) -> bool:
+ return key.lower().endswith(_AUDIO_EXTENSIONS)
+
+
def _extension_family(key: str) -> str:
ext = os.path.splitext(key)[1].lower()
if ext in _JSON_EXTENSIONS:
@@ -156,6 +168,128 @@ def _raise_if_cancelled(cancel_callback: Optional[Callable[[], bool]]) -> None:
raise S3DownloadCancelled("S3 dataset download cancelled")
+def _download_one(client, bucket, key, local_path, cancel_callback) -> None:
+ download_kwargs = {}
+ if cancel_callback is not None:
+ download_kwargs["Callback"] = lambda _bytes: _raise_if_cancelled(cancel_callback)
+ client.download_file(bucket, key, local_path, **download_kwargs)
+ _raise_if_cancelled(cancel_callback)
+
+
+def _key_relative_to_prefix(key: str, prefix: Optional[str]) -> str:
+ if prefix and key.startswith(prefix):
+ key = key[len(prefix) :]
+ return key.lstrip("/")
+
+
+def _download_structured(
+ client, bucket: str, prefix: Optional[str], keys: list[str], target_dir: str, cancel_callback
+) -> dict[str, str]:
+ """Download ``keys`` mirroring their prefix-relative layout. Returns key -> local path."""
+ local_by_key: dict[str, str] = {}
+ for key in keys:
+ _raise_if_cancelled(cancel_callback)
+ relative = _key_relative_to_prefix(key, prefix)
+ parts = [part for part in relative.split("/") if part not in ("", ".")]
+ # ".." is a legal S3 key segment: refuse rather than write above target_dir.
+ if ".." in parts or not parts:
+ raise ValueError(
+ f"S3 key {key!r} contains '..' or empty path segments and cannot "
+ "be downloaded into the dataset directory."
+ )
+ local_path = os.path.join(target_dir, *parts)
+ os.makedirs(os.path.dirname(local_path) or target_dir, exist_ok = True)
+ _download_one(client, bucket, key, local_path, cancel_callback)
+ local_by_key[key] = local_path
+ return local_by_key
+
+
+def _rewrite_audio_references(
+ manifest_local_by_key: dict[str, str],
+ audio_local_by_key: dict[str, str],
+ bucket: str,
+ prefix: Optional[str],
+) -> None:
+ """Rewrite audio values naming a downloaded key (prefix-relative, ``s3://`` URI, or
+ manifest-relative) to its local path; anything else is left as written."""
+ lookup: dict[str, str] = {}
+ for key, local_path in audio_local_by_key.items():
+ lookup[_key_relative_to_prefix(key, prefix)] = local_path
+ lookup[f"s3://{bucket}/{key}"] = local_path
+
+ for manifest_key, manifest_path in manifest_local_by_key.items():
+ if not manifest_path.lower().endswith(_REWRITABLE_MANIFEST_EXTENSIONS):
+ continue
+ manifest_dir = os.path.dirname(_key_relative_to_prefix(manifest_key, prefix))
+
+ def resolve(value):
+ if not isinstance(value, str) or not value.lower().endswith(_AUDIO_EXTENSIONS):
+ return None
+ direct = lookup.get(value)
+ if direct is not None:
+ return direct
+ relative_to_manifest = os.path.normpath(os.path.join(manifest_dir, value))
+ return lookup.get(relative_to_manifest.replace(os.sep, "/"))
+
+ if manifest_path.lower().endswith(".csv"):
+ _rewrite_csv_manifest(manifest_path, resolve)
+ else:
+ _rewrite_json_manifest(manifest_path, resolve)
+
+
+def _rewrite_row(row, resolve) -> None:
+ if not isinstance(row, dict):
+ return
+ for column, value in row.items():
+ replacement = resolve(value)
+ if replacement is not None:
+ row[column] = replacement
+ elif isinstance(value, dict):
+ # HF undecoded audio: {"path": ..., "bytes": ...}
+ replacement = resolve(value.get("path"))
+ if replacement is not None:
+ value["path"] = replacement
+
+
+def _rewrite_json_manifest(manifest_path: str, resolve) -> None:
+ with open(manifest_path, encoding = "utf-8") as f:
+ text = f.read()
+ try:
+ data = None if manifest_path.lower().endswith(".jsonl") else json.loads(text)
+ except json.JSONDecodeError:
+ data = None # JSON Lines in a .json file
+ if data is None:
+ rows = [json.loads(line) for line in text.splitlines() if line.strip()]
+ for row in rows:
+ _rewrite_row(row, resolve)
+ with open(manifest_path, "w", encoding = "utf-8") as f:
+ 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:
+ _rewrite_row(row, resolve)
+ with open(manifest_path, "w", encoding = "utf-8") as f:
+ json.dump(data, f, ensure_ascii = False)
+
+
+def _rewrite_csv_manifest(manifest_path: str, resolve) -> None:
+ with open(manifest_path, encoding = "utf-8", newline = "") as f:
+ rows = list(csv.reader(f))
+ changed = False
+ for row in rows:
+ for index, cell in enumerate(row):
+ replacement = resolve(cell)
+ if replacement is not None:
+ row[index] = replacement
+ changed = True
+ if not changed:
+ return
+ with open(manifest_path, "w", encoding = "utf-8", newline = "") as f:
+ csv.writer(f).writerows(rows)
+
+
def prepare_s3_dataset_download(
s3_config: dict,
dest_dir: Optional[str] = None,
@@ -180,10 +314,18 @@ def prepare_s3_dataset_download(
_raise_if_cancelled(cancel_callback)
client = _build_s3_client(s3_config)
- keys = _list_dataset_keys(client, bucket, prefix)
+ listed = _list_dataset_keys(client, bucket, prefix)
+ keys = [key for key in listed if not _is_audio_key(key)]
+ audio_keys = [key for key in listed if _is_audio_key(key)]
_raise_if_cancelled(cancel_callback)
+ where = f"s3://{bucket}/{prefix}" if prefix else f"s3://{bucket}"
if not keys:
- where = f"s3://{bucket}/{prefix}" if prefix else f"s3://{bucket}"
+ if audio_keys:
+ raise ValueError(
+ f"Found {len(audio_keys)} audio file(s) under {where} but no manifest. "
+ "An audio dataset needs a JSON/JSONL/CSV manifest beside the audio, "
+ "with a column of audio paths and a column of transcriptions."
+ )
raise ValueError(
f"No supported dataset files ({', '.join(SUPPORTED_EXTENSIONS)}) "
f"found under {where}"
@@ -201,17 +343,23 @@ def prepare_s3_dataset_download(
os.makedirs(target_dir, exist_ok = True)
local_files: list[str] = []
- used_paths: set[str] = set()
- for key in keys:
- _raise_if_cancelled(cancel_callback)
- filename = os.path.basename(key)
- local_path = _unique_local_path(target_dir, filename, used_paths)
- download_kwargs = {}
- if cancel_callback is not None:
- download_kwargs["Callback"] = lambda _bytes: _raise_if_cancelled(cancel_callback)
- client.download_file(bucket, key, local_path, **download_kwargs)
- _raise_if_cancelled(cancel_callback)
- local_files.append(local_path)
+ if audio_keys:
+ manifest_local_by_key = _download_structured(
+ client, bucket, prefix, keys, target_dir, cancel_callback
+ )
+ audio_local_by_key = _download_structured(
+ client, bucket, prefix, audio_keys, target_dir, cancel_callback
+ )
+ local_files = list(manifest_local_by_key.values())
+ _rewrite_audio_references(manifest_local_by_key, audio_local_by_key, bucket, prefix)
+ else:
+ used_paths: set[str] = set()
+ for key in keys:
+ _raise_if_cancelled(cancel_callback)
+ filename = os.path.basename(key)
+ local_path = _unique_local_path(target_dir, filename, used_paths)
+ _download_one(client, bucket, key, local_path, cancel_callback)
+ local_files.append(local_path)
except Exception:
if owns_temp_dir:
shutil.rmtree(target_dir, ignore_errors = True)
diff --git a/studio/backend/tests/test_s3_audio_dataset.py b/studio/backend/tests/test_s3_audio_dataset.py
new file mode 100644
index 00000000000..e02dae08ea7
--- /dev/null
+++ b/studio/backend/tests/test_s3_audio_dataset.py
@@ -0,0 +1,239 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""#4539: S3 audio datasets download beside their manifest and the manifest points at them.
+boto3 is faked as in test_s3_dataset.py."""
+
+import csv
+import importlib.util
+import json
+import os
+from pathlib import Path
+
+import pytest
+
+_BACKEND = Path(__file__).resolve().parents[1]
+
+
+def _load(mod_name, rel_path):
+ spec = importlib.util.spec_from_file_location(mod_name, _BACKEND / rel_path)
+ module = importlib.util.module_from_spec(spec)
+ spec.loader.exec_module(module)
+ return module
+
+
+s3_dataset = _load("s3_dataset_audio_tests", "core/training/s3_dataset.py")
+
+
+class _FakePaginator:
+ def __init__(self, keys):
+ self._keys = keys
+
+ def paginate(self, **kwargs):
+ prefix = kwargs.get("Prefix")
+ contents = [{"Key": k} for k in self._keys if prefix is None or k.startswith(prefix)]
+ mid = len(contents) // 2
+ yield {"Contents": contents[:mid]}
+ yield {"Contents": contents[mid:]}
+
+
+class _FakeS3Client:
+ """The house fake, plus per-key content so a manifest can hold real rows."""
+
+ def __init__(
+ self,
+ keys,
+ contents = None,
+ ):
+ self._keys = list(keys)
+ self._contents = contents or {}
+ self.downloaded = []
+
+ def get_paginator(self, name):
+ assert name == "list_objects_v2"
+ return _FakePaginator(self._keys)
+
+ def download_file(self, bucket, key, local_path, **kwargs):
+ self.downloaded.append((bucket, key, local_path))
+ callback = kwargs.get("Callback")
+ if callback is not None:
+ callback(1)
+ with open(local_path, "w", encoding = "utf-8") as f:
+ f.write(self._contents.get(key, f"content-of:{key}"))
+
+
+def _install(
+ monkeypatch,
+ keys,
+ contents = None,
+):
+ client = _FakeS3Client(keys, contents)
+ monkeypatch.setattr(s3_dataset, "boto3_available", lambda: True)
+ monkeypatch.setattr(s3_dataset, "_build_s3_client", lambda cfg: client)
+ return client
+
+
+def _cfg(**overrides):
+ base = {
+ "bucket": "my-bucket",
+ "region": "us-east-1",
+ "prefix": "datasets/",
+ "access_key_id": "AKIA_TEST",
+ "secret_access_key": "secret",
+ "use_iam_role": False,
+ }
+ base.update(overrides)
+ return base
+
+
+def _rows(jsonl_path):
+ with open(jsonl_path, encoding = "utf-8") as f:
+ return [json.loads(line) for line in f if line.strip()]
+
+
+def test_audio_files_download_beside_the_manifest_preserving_structure(monkeypatch, tmp_path):
+ _install(
+ monkeypatch,
+ [
+ "datasets/metadata.jsonl",
+ "datasets/audio/a.wav",
+ "datasets/audio/b.mp3",
+ "datasets/notes.txt", # still unsupported, still skipped
+ ],
+ contents = {"datasets/metadata.jsonl": '{"audio": "audio/a.wav", "text": "hi"}\n'},
+ )
+
+ files = s3_dataset.download_s3_dataset(_cfg(), dest_dir = str(tmp_path))
+
+ assert [os.path.basename(f) for f in files] == ["metadata.jsonl"]
+ assert (tmp_path / "audio" / "a.wav").exists()
+ assert (tmp_path / "audio" / "b.mp3").exists()
+ assert not (tmp_path / "notes.txt").exists()
+
+
+def test_a_manifest_beside_audio_is_not_a_mixed_format_error(monkeypatch, tmp_path):
+ _install(
+ monkeypatch,
+ ["datasets/metadata.csv", "datasets/a.wav", "datasets/b.mp3"],
+ contents = {"datasets/metadata.csv": "audio,text\na.wav,hi\n"},
+ )
+ files = s3_dataset.download_s3_dataset(_cfg(), dest_dir = str(tmp_path))
+ assert [os.path.basename(f) for f in files] == ["metadata.csv"]
+
+
+def test_two_manifest_formats_beside_audio_is_still_a_mixed_format_error(monkeypatch, tmp_path):
+ client = _install(
+ monkeypatch,
+ ["datasets/metadata.csv", "datasets/extra.parquet", "datasets/a.wav"],
+ )
+ with pytest.raises(ValueError, match = "mixed dataset formats"):
+ s3_dataset.download_s3_dataset(_cfg(), dest_dir = str(tmp_path))
+ assert client.downloaded == []
+
+
+def test_audio_with_no_manifest_names_whats_missing(monkeypatch, tmp_path):
+ _install(monkeypatch, ["datasets/audio/a.wav", "datasets/audio/b.mp3"])
+ with pytest.raises(ValueError, match = "manifest"):
+ s3_dataset.download_s3_dataset(_cfg(), dest_dir = str(tmp_path))
+
+
+def test_jsonl_audio_references_are_rewritten_to_local_paths(monkeypatch, tmp_path):
+ manifest = "\n".join(
+ [
+ # A prefix-relative key.
+ json.dumps({"audio": "audio/a.wav", "text": "one"}),
+ # A full S3 URI to a downloaded key.
+ json.dumps({"audio": "s3://my-bucket/datasets/audio/b.mp3", "text": "two"}),
+ # The HF undecoded-audio dict shape.
+ json.dumps({"audio": {"path": "audio/a.wav"}, "text": "three"}),
+ # Unmatched references are somebody else's contract: left alone.
+ json.dumps({"audio": "https://example.com/c.wav", "text": "four"}),
+ ]
+ )
+ _install(
+ monkeypatch,
+ ["datasets/metadata.jsonl", "datasets/audio/a.wav", "datasets/audio/b.mp3"],
+ contents = {"datasets/metadata.jsonl": manifest + "\n"},
+ )
+
+ files = s3_dataset.download_s3_dataset(_cfg(), dest_dir = str(tmp_path))
+ rows = _rows(files[0])
+
+ local_a = str(tmp_path / "audio" / "a.wav")
+ local_b = str(tmp_path / "audio" / "b.mp3")
+ assert rows[0]["audio"] == local_a
+ assert rows[1]["audio"] == local_b
+ assert rows[2]["audio"]["path"] == local_a
+ assert rows[3]["audio"] == "https://example.com/c.wav"
+ assert all(os.path.isabs(p) for p in (rows[0]["audio"], rows[1]["audio"]))
+ assert rows[0]["text"] == "one"
+
+
+def test_json_lines_in_a_json_manifest_are_rewritten(monkeypatch, tmp_path):
+ manifest = "".join(
+ json.dumps({"audio": name, "text": "hi"}) + "\n" for name in ("a.wav", "b.wav")
+ )
+ _install(
+ monkeypatch,
+ ["datasets/train.json", "datasets/a.wav", "datasets/b.wav"],
+ contents = {"datasets/train.json": manifest},
+ )
+ files = s3_dataset.download_s3_dataset(_cfg(), dest_dir = str(tmp_path))
+ assert [row["audio"] for row in _rows(files[0])] == [
+ str(tmp_path / "a.wav"),
+ str(tmp_path / "b.wav"),
+ ]
+
+
+def test_manifest_relative_references_resolve_against_the_manifest_dir(monkeypatch, tmp_path):
+ _install(
+ monkeypatch,
+ ["datasets/train/metadata.jsonl", "datasets/train/clips/a.wav"],
+ contents = {
+ "datasets/train/metadata.jsonl": json.dumps({"audio": "clips/a.wav", "text": "hi"})
+ + "\n"
+ },
+ )
+ files = s3_dataset.download_s3_dataset(_cfg(), dest_dir = str(tmp_path))
+ rows = _rows(files[0])
+ assert rows[0]["audio"] == str(tmp_path / "train" / "clips" / "a.wav")
+
+
+def test_csv_audio_references_are_rewritten_to_local_paths(monkeypatch, tmp_path):
+ _install(
+ monkeypatch,
+ ["datasets/metadata.csv", "datasets/audio/a.wav"],
+ contents = {
+ "datasets/metadata.csv": "audio,text\naudio/a.wav,hello there\nmissing.wav,kept\n"
+ },
+ )
+ files = s3_dataset.download_s3_dataset(_cfg(), dest_dir = str(tmp_path))
+ with open(files[0], encoding = "utf-8") as f:
+ rows = list(csv.DictReader(f))
+ assert rows[0]["audio"] == str(tmp_path / "audio" / "a.wav")
+ assert rows[0]["text"] == "hello there"
+ # An unmatched reference stays; it may be absolute on the training host.
+ assert rows[1]["audio"] == "missing.wav"
+
+
+def test_tabular_only_prefixes_keep_the_flat_layout(monkeypatch, tmp_path):
+ # #6222's flat, collision-renamed layout is unchanged without audio.
+ _install(monkeypatch, ["datasets/sub/train.parquet", "datasets/other/train.parquet"])
+ files = s3_dataset.download_s3_dataset(_cfg(), dest_dir = str(tmp_path))
+ assert sorted(os.path.basename(f) for f in files) == ["train.parquet", "train_1.parquet"]
+ assert not (tmp_path / "sub").exists()
+
+
+def test_a_key_with_dotdot_segments_cannot_escape_the_download_dir(monkeypatch, tmp_path):
+ client = _install(
+ monkeypatch,
+ ["datasets/metadata.jsonl", "datasets/../../evil.wav"],
+ contents = {"datasets/metadata.jsonl": '{"audio": "a.wav", "text": "hi"}\n'},
+ )
+ target = tmp_path / "download"
+ target.mkdir()
+ with pytest.raises(ValueError, match = "\\.\\."):
+ s3_dataset.download_s3_dataset(_cfg(), dest_dir = str(target))
+ escaped = tmp_path / "evil.wav"
+ assert not escaped.exists()
+ assert all("evil" not in key for (_b, key, _p) in client.downloaded)
diff --git a/studio/backend/tests/test_s3_audio_dataset_live.py b/studio/backend/tests/test_s3_audio_dataset_live.py
new file mode 100644
index 00000000000..fac3c698337
--- /dev/null
+++ b/studio/backend/tests/test_s3_audio_dataset_live.py
@@ -0,0 +1,96 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""The audio S3 loader against a real, writable bucket (ambient AWS credentials; set
+``AWS_ENDPOINT_URL_S3`` for an S3-compatible server). Opt-in; uploads are deleted after.
+
+ UNSLOTH_S3_LIVE_BUCKET=my-bucket pytest studio/backend/tests/test_s3_audio_dataset_live.py -q
+"""
+
+from __future__ import annotations
+
+import io
+import json
+import os
+import struct
+import sys
+import uuid
+import wave
+from pathlib import Path
+
+import pytest
+
+_BACKEND = Path(__file__).resolve().parents[1]
+if str(_BACKEND) not in sys.path:
+ sys.path.insert(0, str(_BACKEND))
+
+BUCKET = os.environ.get("UNSLOTH_S3_LIVE_BUCKET", "")
+pytestmark = pytest.mark.skipif(not BUCKET, reason = "UNSLOTH_S3_LIVE_BUCKET not set")
+
+
+def _wav(frames: int = 800) -> bytes:
+ buf = io.BytesIO()
+ with wave.open(buf, "wb") as w:
+ w.setnchannels(1)
+ w.setsampwidth(2)
+ w.setframerate(8000)
+ w.writeframes(
+ b"".join(struct.pack("<h", 3000 if i % 20 < 10 else -3000) for i in range(frames))
+ )
+ return buf.getvalue()
+
+
+def uploaded():
+ boto3 = pytest.importorskip("boto3")
+ client = boto3.client("s3", region_name = os.environ.get("AWS_DEFAULT_REGION", "us-east-1"))
+ prefix = f"unsloth-live-{uuid.uuid4().hex[:8]}/"
+ manifest = "\n".join(
+ [
+ json.dumps({"audio": "audio/a.wav", "text": "one"}),
+ json.dumps({"audio": f"s3://{BUCKET}/{prefix}audio/sub/b.wav", "text": "two"}),
+ json.dumps({"audio": {"path": "audio/a.wav"}, "text": "three"}),
+ json.dumps({"audio": "https://example.com/c.wav", "text": "four"}),
+ ]
+ )
+ objects = {
+ f"{prefix}metadata.jsonl": (manifest + "\n").encode(),
+ f"{prefix}audio/a.wav": _wav(),
+ f"{prefix}audio/sub/b.wav": _wav(),
+ }
+ for key, body in objects.items():
+ client.put_object(Bucket = BUCKET, Key = key, Body = body)
+ try:
+ yield prefix
+ finally:
+ client.delete_objects(Bucket = BUCKET, Delete = {"Objects": [{"Key": key} for key in objects]})
+
+
+def test_audio_dataset_round_trips_through_a_real_bucket(uploaded, tmp_path):
+ from core.training import s3_dataset
+
+ cfg = {
+ "bucket": BUCKET,
+ "region": os.environ.get("AWS_DEFAULT_REGION", "us-east-1"),
+ "prefix": uploaded,
+ "access_key_id": os.environ.get("AWS_ACCESS_KEY_ID"),
+ "secret_access_key": os.environ.get("AWS_SECRET_ACCESS_KEY"),
+ "use_iam_role": not os.environ.get("AWS_ACCESS_KEY_ID"),
+ }
+ files = s3_dataset.download_s3_dataset(cfg, dest_dir = str(tmp_path))
+
+ assert [os.path.relpath(f, tmp_path) for f in files] == ["metadata.jsonl"]
+ a = str(tmp_path / "audio" / "a.wav")
+ b = str(tmp_path / "audio" / "sub" / "b.wav")
+ assert os.path.isfile(a) and os.path.isfile(
+ b
+ ), "audio lands beside the manifest, structure kept"
+ for path in (a, b):
+ with wave.open(path) as w:
+ assert w.getnframes() == 800, "bytes survive the trip"
+
+ rows = [json.loads(line) for line in open(files[0], encoding = "utf-8") if line.strip()]
+ assert rows[0]["audio"] == a
+ assert rows[1]["audio"] == b
+ assert rows[2]["audio"]["path"] == a
+ assert rows[3]["audio"] == "https://example.com/c.wav"
diff --git a/studio/frontend/src/features/studio/sections/dataset-section.tsx b/studio/frontend/src/features/studio/sections/dataset-section.tsx
index f9877057fac..8be26e92d51 100644
--- a/studio/frontend/src/features/studio/sections/dataset-section.tsx
+++ b/studio/frontend/src/features/studio/sections/dataset-section.tsx
@@ -38,34 +38,27 @@ export function DatasetPanel() {
datasetSource,
restoreBrowseDatasetSource,
isVisionModel,
- isAudioModel,
modelType,
} = useTrainingConfigStore(
useShallow((state) => ({
datasetSource: state.datasetSource,
restoreBrowseDatasetSource: state.restoreBrowseDatasetSource,
isVisionModel: state.isVisionModel,
- isAudioModel: state.isAudioModel,
modelType: state.modelType,
})),
);
const localDatasetInventory = useLocalDatasetInventory(datasetSource);
const uploads = useDatasetUploads();
- const effectiveModelType = modelType ?? "text";
- const isMultimodalModel =
- effectiveModelType === "vision" ||
- effectiveModelType === "audio" ||
- isVisionModel ||
- isAudioModel;
+ const isVisionTraining = (modelType ?? "text") === "vision" || isVisionModel;
useEffect(() => {
- if (datasetSource === "s3" && isMultimodalModel) {
+ if (datasetSource === "s3" && isVisionTraining) {
restoreBrowseDatasetSource();
toast.info(t("studio.training.validation.s3MultimodalUnsupported"), {
id: "training-s3-multimodal-source-reset",
});
}
- }, [datasetSource, isMultimodalModel, restoreBrowseDatasetSource, t]);
+ }, [datasetSource, isVisionTraining, restoreBrowseDatasetSource, t]);
return (
<div className="flex min-w-0 flex-col gap-5">
diff --git a/studio/frontend/src/features/studio/sections/dataset-source-toggle.tsx b/studio/frontend/src/features/studio/sections/dataset-source-toggle.tsx
index 7c26d27e8cc..b53d6eca430 100644
--- a/studio/frontend/src/features/studio/sections/dataset-source-toggle.tsx
+++ b/studio/frontend/src/features/studio/sections/dataset-source-toggle.tsx
@@ -11,18 +11,18 @@ type DatasetSourceMode = "browse" | "s3";
export function DatasetSourceToggle({
datasetSource,
- isMultimodalModel,
+ isVisionModel,
restoreBrowseDatasetSource,
selectS3Source,
}: {
datasetSource: DatasetSource;
- isMultimodalModel: boolean;
+ isVisionModel: boolean;
restoreBrowseDatasetSource: () => void;
selectS3Source: () => void;
}) {
const t = useT();
- if (isMultimodalModel) {
+ if (isVisionModel) {
return null;
}
@@ -57,7 +57,6 @@ export function DatasetSourceToggleAction() {
selectS3Source,
restoreBrowseDatasetSource,
isVisionModel,
- isAudioModel,
modelType,
} = useTrainingConfigStore(
useShallow((state) => ({
@@ -65,7 +64,6 @@ export function DatasetSourceToggleAction() {
selectS3Source: state.selectS3Source,
restoreBrowseDatasetSource: state.restoreBrowseDatasetSource,
isVisionModel: state.isVisionModel,
- isAudioModel: state.isAudioModel,
modelType: state.modelType,
})),
);
@@ -74,12 +72,7 @@ export function DatasetSourceToggleAction() {
return (
<DatasetSourceToggle
datasetSource={datasetSource}
- isMultimodalModel={
- effectiveModelType === "vision" ||
- effectiveModelType === "audio" ||
- isVisionModel ||
- isAudioModel
- }
+ isVisionModel={effectiveModelType === "vision" || isVisionModel}
restoreBrowseDatasetSource={restoreBrowseDatasetSource}
selectS3Source={selectS3Source}
/>
diff --git a/studio/frontend/src/features/training/lib/validation.ts b/studio/frontend/src/features/training/lib/validation.ts
index ec2e2951e6a..61b86a0c8b8 100644
--- a/studio/frontend/src/features/training/lib/validation.ts
+++ b/studio/frontend/src/features/training/lib/validation.ts
@@ -64,12 +64,8 @@ export function hasIncompatibleTrainingModalities(
export function validateS3Source(
config: TrainingConfigState,
): StartValidationResult {
- if (
- config.modelType === "vision" ||
- config.modelType === "audio" ||
- config.isVisionModel ||
- config.isAudioModel
- ) {
+ // Audio manifests are rewritten to downloaded paths (#4539); image references are not.
+ if (config.modelType === "vision" || config.isVisionModel) {
return {
ok: false,
errorKey: "studio.training.validation.s3MultimodalUnsupported",
diff --git a/studio/frontend/src/i18n/locales/ar.ts b/studio/frontend/src/i18n/locales/ar.ts
index e3df847da81..4e4943fd40f 100644
--- a/studio/frontend/src/i18n/locales/ar.ts
+++ b/studio/frontend/src/i18n/locales/ar.ts
@@ -3187,7 +3187,7 @@ export const ar = {
"تغيّر إعداد التدريب أثناء التحقق منه. راجعه ثم ابدأ التدريب مجددًا.",
validation: {
s3MultimodalUnsupported:
- "مجموعات بيانات S3 غير مدعومة بعد لتدريب الرؤية أو الصوت.",
+ "مجموعات بيانات S3 غير مدعومة بعد لتدريب الرؤية.",
s3BucketRequired: "أدخل اسم حاوية S3 أولًا.",
s3CredentialsRequired: "أدخل مفاتيح الوصول إلى S3 أو فعّل دور IAM.",
modelRequired: "اختر نموذجًا أساسيًا أولًا.",
diff --git a/studio/frontend/src/i18n/locales/de.ts b/studio/frontend/src/i18n/locales/de.ts
index 3653e1ea8ae..a58b87f0840 100644
--- a/studio/frontend/src/i18n/locales/de.ts
+++ b/studio/frontend/src/i18n/locales/de.ts
@@ -3245,7 +3245,7 @@ export const de = {
"Die Trainingseinstellungen haben sich während der Überprüfung geändert. Prüfen Sie sie und starten Sie das Training erneut.",
validation: {
s3MultimodalUnsupported:
- "S3-Datensätze werden für Vision- oder Audio-Training noch nicht unterstützt.",
+ "S3-Datensätze werden für Vision-Training noch nicht unterstützt.",
s3BucketRequired: "Geben Sie zuerst den Namen eines S3-Buckets ein.",
s3CredentialsRequired:
"Geben Sie S3-Zugriffsschlüssel an oder aktivieren Sie die IAM-Rolle.",
diff --git a/studio/frontend/src/i18n/locales/en.ts b/studio/frontend/src/i18n/locales/en.ts
index 965e76ef49e..10b8acb1dab 100644
--- a/studio/frontend/src/i18n/locales/en.ts
+++ b/studio/frontend/src/i18n/locales/en.ts
@@ -3195,7 +3195,7 @@ export const en = {
chooseModelAndDataset: "Choose model and dataset",
validation: {
s3MultimodalUnsupported:
- "S3 datasets are not supported for vision or audio training yet.",
+ "S3 datasets are not supported for vision training yet.",
s3BucketRequired: "Enter an S3 bucket name first.",
s3CredentialsRequired: "Provide S3 access keys or enable IAM role.",
modelRequired: "Select a base model first.",
diff --git a/studio/frontend/src/i18n/locales/es.ts b/studio/frontend/src/i18n/locales/es.ts
index 801dfd830cd..6dd5e1c7042 100644
--- a/studio/frontend/src/i18n/locales/es.ts
+++ b/studio/frontend/src/i18n/locales/es.ts
@@ -3238,7 +3238,7 @@ export const es = {
"La configuración del entrenamiento cambió mientras se comprobaba. Revísala y vuelve a iniciar el entrenamiento.",
validation: {
s3MultimodalUnsupported:
- "Los conjuntos de datos de S3 todavía no son compatibles con el entrenamiento de visión o audio.",
+ "Los conjuntos de datos de S3 todavía no son compatibles con el entrenamiento de visión.",
s3BucketRequired: "Introduce primero el nombre de un bucket de S3.",
s3CredentialsRequired:
"Proporciona las claves de acceso de S3 o activa el rol de IAM.",
diff --git a/studio/frontend/src/i18n/locales/fr.ts b/studio/frontend/src/i18n/locales/fr.ts
index 6aa31cb58db..bc257cc1695 100644
--- a/studio/frontend/src/i18n/locales/fr.ts
+++ b/studio/frontend/src/i18n/locales/fr.ts
@@ -3255,7 +3255,7 @@ export const fr = {
"La configuration de l'entraînement a changé pendant sa vérification. Vérifiez-la, puis relancez l'entraînement.",
validation: {
s3MultimodalUnsupported:
- "Les jeux de données S3 ne sont pas encore pris en charge pour l'entraînement de modèles de vision ou audio.",
+ "Les jeux de données S3 ne sont pas encore pris en charge pour l'entraînement de modèles de vision.",
s3BucketRequired: "Saisissez d'abord le nom d'un bucket S3.",
s3CredentialsRequired:
"Indiquez des clés d'accès S3 ou activez le rôle IAM.",
diff --git a/studio/frontend/src/i18n/locales/hi.ts b/studio/frontend/src/i18n/locales/hi.ts
index 8909cf5bd42..2266b4d8211 100644
--- a/studio/frontend/src/i18n/locales/hi.ts
+++ b/studio/frontend/src/i18n/locales/hi.ts
@@ -3194,7 +3194,7 @@ export const hi = {
"जांच के दौरान ट्रेनिंग सेटअप बदल गया। इसकी समीक्षा करें और फिर से ट्रेनिंग शुरू करें।",
validation: {
s3MultimodalUnsupported:
- "S3 डेटासेट अभी विज़न या ऑडियो ट्रेनिंग के लिए समर्थित नहीं हैं।",
+ "S3 डेटासेट अभी विज़न ट्रेनिंग के लिए समर्थित नहीं हैं।",
s3BucketRequired: "पहले S3 बकेट का नाम दर्ज करें।",
s3CredentialsRequired: "S3 एक्सेस कुंजियाँ दें या IAM रोल सक्षम करें।",
modelRequired: "पहले बेस मॉडल चुनें।",
diff --git a/studio/frontend/src/i18n/locales/it.ts b/studio/frontend/src/i18n/locales/it.ts
index 2074596d4f5..c20cc1c7c1b 100644
--- a/studio/frontend/src/i18n/locales/it.ts
+++ b/studio/frontend/src/i18n/locales/it.ts
@@ -3227,7 +3227,7 @@ export const it = {
chooseModelAndDataset: "Scegli modello e dataset",
validation: {
s3MultimodalUnsupported:
- "I dataset S3 non sono ancora supportati per l'addestramento visivo o audio.",
+ "I dataset S3 non sono ancora supportati per l'addestramento visivo.",
s3BucketRequired: "Inserisci prima il nome di un bucket S3.",
s3CredentialsRequired:
"Fornisci le chiavi di accesso S3 oppure attiva il ruolo IAM.",
diff --git a/studio/frontend/src/i18n/locales/ja.ts b/studio/frontend/src/i18n/locales/ja.ts
index 996fb0d1534..733faf1caf7 100644
--- a/studio/frontend/src/i18n/locales/ja.ts
+++ b/studio/frontend/src/i18n/locales/ja.ts
@@ -3149,7 +3149,7 @@ export const ja = {
"確認中にトレーニング設定が変更されました。設定を確認してから、トレーニングをもう一度開始してください。",
validation: {
s3MultimodalUnsupported:
- "S3 データセットは、ビジョンまたはオーディオのトレーニングではまだサポートされていません。",
+ "S3 データセットは、ビジョンのトレーニングではまだサポートされていません。",
s3BucketRequired: "先に S3 バケット名を入力してください。",
s3CredentialsRequired:
"S3 アクセスキーを指定するか、IAM ロールを有効にしてください。",
diff --git a/studio/frontend/src/i18n/locales/ko.ts b/studio/frontend/src/i18n/locales/ko.ts
index 1025a09bd03..d072920e960 100644
--- a/studio/frontend/src/i18n/locales/ko.ts
+++ b/studio/frontend/src/i18n/locales/ko.ts
@@ -3191,7 +3191,7 @@ export const ko = {
"확인하는 동안 학습 설정이 변경되었습니다. 설정을 검토한 후 학습을 다시 시작하세요.",
validation: {
s3MultimodalUnsupported:
- "비전 또는 오디오 학습에는 아직 S3 데이터셋을 사용할 수 없습니다.",
+ "비전 학습에는 아직 S3 데이터셋을 사용할 수 없습니다.",
s3BucketRequired: "S3 버킷 이름을 먼저 입력하세요.",
s3CredentialsRequired:
"S3 액세스 키를 입력하거나 IAM 역할을 활성화하세요.",
diff --git a/studio/frontend/src/i18n/locales/pt-br.ts b/studio/frontend/src/i18n/locales/pt-br.ts
index 3dca4b7428c..2914fdb2fa3 100644
--- a/studio/frontend/src/i18n/locales/pt-br.ts
+++ b/studio/frontend/src/i18n/locales/pt-br.ts
@@ -3211,7 +3211,7 @@ export const ptBR = {
"A configuração de treinamento mudou durante a verificação. Revise-a e inicie o treinamento novamente.",
validation: {
s3MultimodalUnsupported:
- "Datasets do S3 ainda não são compatíveis com treinamento de visão ou áudio.",
+ "Datasets do S3 ainda não são compatíveis com treinamento de visão.",
s3BucketRequired: "Insira primeiro o nome de um bucket do S3.",
s3CredentialsRequired:
"Forneça as chaves de acesso do S3 ou ative a função do IAM.",
diff --git a/studio/frontend/src/i18n/locales/ru.ts b/studio/frontend/src/i18n/locales/ru.ts
index 6bd2d34ffae..13846054a40 100644
--- a/studio/frontend/src/i18n/locales/ru.ts
+++ b/studio/frontend/src/i18n/locales/ru.ts
@@ -3222,7 +3222,7 @@ export const ru = {
"Настройки обучения изменились во время проверки. Проверьте их и снова запустите обучение.",
validation: {
s3MultimodalUnsupported:
- "Датасеты S3 пока не поддерживаются для обучения моделей зрения или аудио.",
+ "Датасеты S3 пока не поддерживаются для обучения моделей зрения.",
s3BucketRequired: "Сначала введите имя бакета S3.",
s3CredentialsRequired:
"Укажите ключи доступа S3 или включите роль IAM.",
diff --git a/studio/frontend/src/i18n/locales/zh-CN.ts b/studio/frontend/src/i18n/locales/zh-CN.ts
index f46defd34fb..5b6b868eae6 100644
--- a/studio/frontend/src/i18n/locales/zh-CN.ts
+++ b/studio/frontend/src/i18n/locales/zh-CN.ts
@@ -3112,7 +3112,7 @@ export const zhCN = {
"无法验证数据集与此模型是否兼容。请检查网络连接或 Hugging Face token;开始训练时会重试检查。",
setupChanged: "检查期间,训练设置发生了变化。请确认设置后重新开始训练。",
validation: {
- s3MultimodalUnsupported: "S3 数据集暂不支持视觉或音频训练。",
+ s3MultimodalUnsupported: "S3 数据集暂不支持视觉训练。",
s3BucketRequired: "请先输入 S3 存储桶名称。",
s3CredentialsRequired: "请提供 S3 访问密钥或启用 IAM 角色。",
modelRequired: "请先选择基础模型。",
diff --git a/studio/frontend/tests/training-validation.test.ts b/studio/frontend/tests/training-validation.test.ts
index d75258f5189..f774484eb04 100644
--- a/studio/frontend/tests/training-validation.test.ts
+++ b/studio/frontend/tests/training-validation.test.ts
@@ -9,7 +9,7 @@ import { registerBundlerResolver } from "./helpers/kit.ts";
registerBundlerResolver();
-const { validateTrainingConfig } = await import(
+const { validateS3Source, validateTrainingConfig } = await import(
"../src/features/training/lib/validation.ts"
);
@@ -323,3 +323,26 @@ test("training validation keeps CPT and embedding training available off MLX", (
{ ok: true, errorKey: null },
);
});
+
+const s3Config = {
+ ...validConfig,
+ datasetSource: "s3" as const,
+ s3Config: { bucket: "my-bucket", region: "us-east-1", useIamRole: true },
+};
+
+test("an audio model can start from S3 (#4539: the audio is downloaded beside its manifest)", () => {
+ assert.deepEqual(
+ validateS3Source({ ...s3Config, modelType: "audio", isAudioModel: true }),
+ { ok: true, errorKey: null },
+ );
+});
+
+test("a vision model still cannot start from S3", () => {
+ assert.deepEqual(
+ validateS3Source({ ...s3Config, modelType: "vision", isVisionModel: true }),
+ {
+ ok: false,
+ errorKey: "studio.training.validation.s3MultimodalUnsupported",
+ },
+ );
+});