Fine-tune LLMs from one YAML. Layer streaming trains an 8B model on a 4 GB laptop GPU.
Python cliconsumer-gpudpofine-tuningggufhuggingfacellmllmopslocal-ailocal-llmloralow-vrammachine-learningollamapeftpythonpytorchqlorasfttransformers
"""Tests for issue #1448: soup data generate --dedup-with on --template preference.
Acceptance criteria:
- generate --template preference --dedup-with drops a repeated row (both DPO and KTO formats).
- generate --template preference deduplicates repeated rows within a run.
- _row_to_text produces a stable key for preference rows regardless of dictionary key ordering.
"""
import json
from unittest.mock import patch
from typer.testing import CliRunner
from soup_cli.cli import app
from soup_cli.commands.generate import _row_to_text
from tests.conftest import strip_ansi
def test_row_to_text_preference_stable_key():
"""_row_to_text should produce identical keys regardless of dict insertion order."""
dpo_1 = {"prompt": "What is 2+2?", "chosen": "4", "rejected": "5"}
dpo_2 = {"chosen": "4", "rejected": "5", "prompt": "What is 2+2?"}
assert _row_to_text(dpo_1) == _row_to_text(dpo_2)
kto_1 = {"prompt": "Write a poem", "completion": "Roses are red", "label": True}
kto_2 = {"label": True, "prompt": "Write a poem", "completion": "Roses are red"}
assert _row_to_text(kto_1) == _row_to_text(kto_2)
def test_preference_dedup_with_drops_repeated_rows(tmp_path, monkeypatch):
"""--dedup-with must drop repeated DPO and KTO rows when --template preference is used."""
monkeypatch.chdir(tmp_path)
existing_file = tmp_path / "existing.jsonl"
existing_records = [
{
"prompt": "Explain gravity",
"chosen": "Mass curves spacetime",
"rejected": "Things fall down",
},
{"prompt": "Is Python fast?", "completion": "It depends on the workload", "label": True},
]
with open(existing_file, "w", encoding="utf-8") as f:
for r in existing_records:
f.write(json.dumps(r) + "\n")
generated_batch = [
# Repeated DPO from existing.jsonl
{
"prompt": "Explain gravity",
"chosen": "Mass curves spacetime",
"rejected": "Things fall down",
},
# New DPO
{"prompt": "Explain entropy", "chosen": "Measure of disorder", "rejected": "Heat"},
# Repeated KTO from existing.jsonl
{"prompt": "Is Python fast?", "completion": "It depends on the workload", "label": True},
# New KTO
{"prompt": "Is Rust safe?", "completion": "Yes, memory safety without GC", "label": True},
]
runner = CliRunner()
with patch("soup_cli.commands.generate._generate_batch", return_value=generated_batch):
result = runner.invoke(app, [
"data", "generate",
"--prompt", "Generate preference data",
"--template", "preference",
"--dedup-with", str(existing_file.name),
"--output", "out.jsonl",
"--count", "4",
"--provider", "server",
])
assert result.exit_code == 0, result.output
assert "Duplicates: 2 removed" in " ".join(strip_ansi(result.output).split())
out_file = tmp_path / "out.jsonl"
assert out_file.exists()
lines = [json.loads(line) for line in out_file.read_text(encoding="utf-8").strip().splitlines()]
# Exactly 2 new rows kept, 2 duplicates dropped
assert len(lines) == 2
assert lines[0]["prompt"] == "Explain entropy"
assert lines[1]["prompt"] == "Is Rust safe?"
def test_preference_within_run_dedup(tmp_path, monkeypatch):
"""Repeated preference rows generated within the same run must be deduplicated."""
monkeypatch.chdir(tmp_path)
generated_batch = [
{"prompt": "P1", "chosen": "C1", "rejected": "R1"},
{"prompt": "P1", "chosen": "C1", "rejected": "R1"}, # duplicate
{"prompt": "P2", "completion": "Comp2", "label": False},
{"prompt": "P2", "completion": "Comp2", "label": False}, # duplicate
]
runner = CliRunner()
with patch("soup_cli.commands.generate._generate_batch", return_value=generated_batch):
result = runner.invoke(app, [
"data", "generate",
"--prompt", "Generate preference data",
"--template", "preference",
"--output", "out.jsonl",
"--count", "4",
"--provider", "server",
])
assert result.exit_code == 0, result.output
out_file = tmp_path / "out.jsonl"
lines = [json.loads(line) for line in out_file.read_text(encoding="utf-8").strip().splitlines()]
assert len(lines) == 2
assert lines[0] == {"prompt": "P1", "chosen": "C1", "rejected": "R1"}
assert lines[1] == {"prompt": "P2", "completion": "Comp2", "label": False}
def test_preference_dedup_keeps_rows_that_differ_beyond_the_prompt(tmp_path, monkeypatch):
"""Only exact repeats are dropped: a row that shares the prompt but differs in
chosen, rejected, completion or label is a different example and is kept."""
monkeypatch.chdir(tmp_path)
batch = [
{"prompt": "P", "chosen": "C1", "rejected": "R1"},
{"prompt": "P", "chosen": "C2", "rejected": "R1"},
{"prompt": "P", "chosen": "C1", "rejected": "R2"},
{"prompt": "P", "completion": "A", "label": True},
{"prompt": "P", "completion": "B", "label": True},
{"prompt": "P", "completion": "A", "label": False},
{"prompt": "P", "chosen": "C1", "rejected": "R1"}, # the only exact repeat
]
with patch("soup_cli.commands.generate._generate_batch", return_value=batch):
result = CliRunner().invoke(app, [
"data", "generate", "--prompt", "x", "--template", "preference",
"--output", "out.jsonl", "--count", "7", "--batch-size", "7",
"--provider", "server",
])
assert result.exit_code == 0, (result.output, repr(result.exception))
rows = [json.loads(line) for line in
(tmp_path / "out.jsonl").read_text(encoding="utf-8").splitlines()]
assert rows == batch[:6]
assert "Duplicates: 1 removed" in " ".join(strip_ansi(result.output).split())