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: read uploaded CSV cells as written for embedding training (#14014)

Nilayff68ecc75789
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index bd9b7adbd31..5bf50dbcb82 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -5472,6 +5472,15 @@ def _create_embedding_progress_callback(
return _EmbeddingProgressCallback()
+def _csv_number(cell):
+ from utils.datasets.cells import _NA_CELLS
+
+ text = (cell or "").strip()
+ if text.lower() in ("true", "false"):
+ return float(text.lower() == "true")
+ return None if text in _NA_CELLS else float(text)
+
+
def _run_embedding_training(event_queue: Any, stop_queue: Any, config: dict) -> None:
"""Self-contained embedding model training pipeline. Uses FastSentenceTransformer +
SentenceTransformerTrainer + MultipleNegativesRankingLoss, separate from UnslothTrainer's
@@ -5662,6 +5671,7 @@ def _run_embedding_training(event_queue: Any, stop_queue: Any, config: dict) ->
config["_dataset_loaded_from_exact_snapshot"] = False
hf_dataset = str(config.get("hf_dataset") or "").strip()
local_datasets = config.get("local_datasets") or []
+ csv_text_columns: list[str] = []
def _load_local_embedding_dataset(dataset_paths: list[str]):
from utils.paths import dataset_files_in_dir
@@ -5694,7 +5704,29 @@ def _run_embedding_training(event_queue: Any, stop_queue: Any, config: dict) ->
loader = "parquet"
else:
raise ValueError(f"Unsupported local dataset format: {all_files[0]}")
- return load_dataset(loader, data_files = all_files, split = "train")
+ from datasets import Features, Value
+ from utils.datasets.cells import csv_as_text_kwargs, typed_csv_columns
+
+ dataset = load_dataset(
+ loader,
+ data_files = all_files,
+ split = "train",
+ **csv_as_text_kwargs(all_files),
+ )
+ if loader != "csv":
+ return dataset
+ labels = typed_csv_columns(dataset) & {"label", "labels", "score", "scores"}
+ for column in labels:
+ dataset = dataset.map(
+ lambda cells: {column: [_csv_number(cell) for cell in cells]},
+ input_columns = column,
+ batched = True,
+ features = Features({**dataset.features, column: Value("float64")}),
+ )
+ csv_text_columns[:] = [
+ column for column in dataset.column_names if column not in labels
+ ]
+ return dataset
if hf_dataset:
dataset = _load_embedding_hf_dataset(
@@ -5749,6 +5781,25 @@ def _run_embedding_training(event_queue: Any, stop_queue: Any, config: dict) ->
end = slice_end if slice_end is not None else len(dataset)
dataset = dataset.select(range(start, min(end + 1, len(dataset))))
+ if csv_text_columns:
+ kept = dataset.filter(
+ lambda *columns: [None not in cells for cells in zip(*columns)],
+ input_columns = csv_text_columns,
+ batched = True,
+ )
+ if len(dataset) and not len(kept):
+ raise ValueError(f"All {len(dataset):,} rows have an empty text cell.")
+ if len(kept) < len(dataset):
+ event_queue.put(
+ {
+ "type": "warning",
+ "message": f"Dropped {len(dataset) - len(kept):,} of {len(dataset):,} "
+ "rows because a text cell was empty.",
+ "ts": time.time(),
+ }
+ )
+ dataset = kept
+
logger.info(f"Embedding dataset loaded: {len(dataset)} samples")
except Exception as e:
event_queue.put(
diff --git a/studio/backend/tests/test_embedding_csv_cells_as_written.py b/studio/backend/tests/test_embedding_csv_cells_as_written.py
new file mode 100644
index 00000000000..6910d3085f1
--- /dev/null
+++ b/studio/backend/tests/test_embedding_csv_cells_as_written.py
@@ -0,0 +1,197 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+_BACKEND = Path(__file__).resolve().parent.parent
+if str(_BACKEND) not in sys.path:
+ sys.path.insert(0, str(_BACKEND))
+
+import importlib # noqa: E402
+import types # noqa: E402
+from unittest.mock import MagicMock # noqa: E402
+
+import pytest # noqa: E402
+
+_STUBBED: list[str] = []
+
+
+def _stub_if_missing(name, attrs):
+ if name in sys.modules:
+ return
+ try:
+ importlib.import_module(name)
+ return
+ except Exception: # noqa: BLE001 - stub unusable imports
+ pass
+ _STUBBED.append(name)
+ mod = types.ModuleType(name)
+ mod.__spec__ = None
+ for attr in attrs:
+ setattr(mod, attr, MagicMock())
+ sys.modules[name] = mod
+ parent, _, child = name.rpartition(".")
+ if parent and parent in sys.modules:
+ setattr(sys.modules[parent], child, mod)
+
+
+_stub_if_missing("unsloth", ("FastLanguageModel", "FastVisionModel", "is_bfloat16_supported"))
+_stub_if_missing("unsloth.chat_templates", ("get_chat_template",))
+_stub_if_missing("trl", ("SFTTrainer", "SFTConfig"))
+
+from core.training import trainer as _trainer # noqa: E402,F401
+from core.training import worker # noqa: E402
+
+for _name in reversed(_STUBBED):
+ sys.modules.pop(_name, None)
+
+
+def _module(name, **attrs):
+ mod = types.ModuleType(name)
+ mod.__spec__ = None
+ mod.__dict__.update(attrs)
+ return mod
+
+
+def run_embedding(monkeypatch, tmp_path):
+ trained = {}
+
+ class Trainer:
+ def __init__(self, **kwargs):
+ trained.update(kwargs)
+
+ def remove_callback(self, callback):
+ pass
+
+ def train(self, **kwargs):
+ pass
+
+ monkeypatch.setitem(
+ sys.modules,
+ "unsloth",
+ _module(
+ "unsloth",
+ FastSentenceTransformer = MagicMock(),
+ is_bfloat16_supported = lambda: False,
+ ),
+ )
+ monkeypatch.setitem(
+ sys.modules,
+ "sentence_transformers",
+ _module(
+ "sentence_transformers",
+ SentenceTransformerTrainer = Trainer,
+ SentenceTransformerTrainingArguments = MagicMock(),
+ ),
+ )
+ monkeypatch.setitem(
+ sys.modules,
+ "sentence_transformers.losses",
+ _module("sentence_transformers.losses", MultipleNegativesRankingLoss = MagicMock()),
+ )
+ monkeypatch.setitem(
+ sys.modules,
+ "sentence_transformers.training_args",
+ _module("sentence_transformers.training_args", BatchSamplers = MagicMock()),
+ )
+ monkeypatch.setattr(worker, "_model_load_security_error", lambda *a, **k: None)
+ monkeypatch.setattr(worker, "_start_worker_stop_poller", lambda *a, **k: None)
+ monkeypatch.setattr(worker, "_emit_resource_provenance", lambda *a, **k: None)
+
+ def run(
+ csv_text,
+ error = None,
+ **config,
+ ):
+ path = tmp_path / "pairs.csv"
+ path.write_text(csv_text)
+ events = []
+ worker._run_embedding_training(
+ types.SimpleNamespace(put = events.append),
+ None,
+ {
+ "model_name": "unsloth/all-MiniLM-L6-v2",
+ "training_type": "Full Finetuning",
+ "local_datasets": [str(path)],
+ "output_dir": "embedding-csv-cells",
+ "max_steps": 1,
+ **config,
+ },
+ )
+ errors = [event["error"] for event in events if event["type"] == "error"]
+ assert errors == ([error] if error else [])
+ warnings = [event["message"] for event in events if event["type"] == "warning"]
+ return trained.get("train_dataset"), warnings
+
+ return run
+
+
+def test_uploaded_csv_trains_embedding_pairs_as_written(run_embedding):
+ dataset, warnings = run_embedding(
+ "query,answer\n"
+ "Zip code of Holtsville NY?,00501\n"
+ "Price for item 7?,NA\n"
+ "What does the API return on a miss?,None\n"
+ "Say nothing.,\n"
+ "Zip code of Schenectady NY?,12345\n"
+ )
+
+ assert dataset.to_list() == [
+ {"query": "Zip code of Holtsville NY?", "answer": "00501"},
+ {"query": "Price for item 7?", "answer": "NA"},
+ {"query": "What does the API return on a miss?", "answer": "None"},
+ {"query": "Zip code of Schenectady NY?", "answer": "12345"},
+ ]
+ assert warnings == ["Dropped 1 of 5 rows because a text cell was empty."]
+
+
+def test_embedding_csv_score_column_still_reaches_the_trainer_as_numbers(run_embedding):
+ dataset, warnings = run_embedding(
+ "sentence1,sentence2,score\nA cat sits.,A cat is sitting.,0.9\nA dog runs.,Rain falls.,0\n"
+ )
+
+ assert dataset.to_list() == [
+ {"sentence1": "A cat sits.", "sentence2": "A cat is sitting.", "score": 0.9},
+ {"sentence1": "A dog runs.", "sentence2": "Rain falls.", "score": 0.0},
+ ]
+ assert warnings == []
+
+
+def test_embedding_csv_label_columns_parse_like_pandas(run_embedding):
+ dataset, _ = run_embedding("s1,s2,label,score\naa,bb,True, 0.5\ncc,dd,False,0.25 \n")
+
+ assert dataset.to_list() == [
+ {"s1": "aa", "s2": "bb", "label": 1.0, "score": 0.5},
+ {"s1": "cc", "s2": "dd", "label": 0.0, "score": 0.25},
+ ]
+
+
+def test_embedding_csv_with_no_text_left_fails_instead_of_training_nothing(run_embedding):
+ run_embedding(
+ "s1,s2,\naa,bb,\ncc,dd,\n",
+ error = "Failed to load dataset: All 2 rows have an empty text cell.",
+ )
+
+
+def test_embedding_csv_slice_counts_rows_of_the_file(run_embedding):
+ dataset, warnings = run_embedding(
+ "q,a\nq0,a0\nq1,\nq2,a2\nq3,a3\nq4,a4\n",
+ dataset_slice_start = 1,
+ dataset_slice_end = 3,
+ )
+
+ assert dataset.to_list() == [{"q": "q2", "a": "a2"}, {"q": "q3", "a": "a3"}]
+ assert warnings == ["Dropped 1 of 3 rows because a text cell was empty."]
+
+
+def test_embedding_csv_score_na_marker_outside_the_slice_still_trains(run_embedding):
+ dataset, _ = run_embedding(
+ "s1,s2,score\naa,bb,0.5\ncc,dd,NA\n",
+ dataset_slice_end = 0,
+ )
+
+ assert dataset.to_list() == [{"s1": "aa", "s2": "bb", "score": 0.5}]