@@ -0,0 +1,252 @@ |
| 1 | +"""Slow integration: train → mine → train again improves held-out preference score. |
| 2 | + |
| 3 | +This is Sprint 42's bootstrap-loop proof. We keep the mined candidates |
| 4 | +deterministic (scripted backend + judge) so the test is stable, but the |
| 5 | +two preference-training passes and the final held-out SwayJudge check are |
| 6 | +real. |
| 7 | +""" |
| 8 | + |
| 9 | +from __future__ import annotations |
| 10 | + |
| 11 | +import shutil |
| 12 | +from collections import deque |
| 13 | +from pathlib import Path |
| 14 | +from typing import TYPE_CHECKING |
| 15 | + |
| 16 | +import pytest |
| 17 | +from typer.testing import CliRunner |
| 18 | + |
| 19 | +from dlm.cli.app import app |
| 20 | +from dlm.doc.parser import ParsedDlm, parse_file |
| 21 | +from dlm.doc.serializer import serialize |
| 22 | +from dlm.preference.judge import JudgeInvocationError, JudgeUnavailableError, PairScore, SwayJudge |
| 23 | + |
| 24 | +if TYPE_CHECKING: |
| 25 | + from tests.fixtures.trained_store import TrainedStoreHandle |
| 26 | + |
| 27 | +pytestmark = pytest.mark.slow |
| 28 | + |
| 29 | +_EXTRA_BODY = """ |
| 30 | +::instruction:: |
| 31 | +### Q |
| 32 | +What color is grass? |
| 33 | +### A |
| 34 | +Green. |
| 35 | + |
| 36 | +::instruction:: |
| 37 | +### Q |
| 38 | +What is 10 - 3? |
| 39 | +### A |
| 40 | +7. |
| 41 | + |
| 42 | +::preference:: |
| 43 | +### Prompt |
| 44 | +Is water wet? |
| 45 | +### Chosen |
| 46 | +Yes. |
| 47 | +### Rejected |
| 48 | +Water is generally considered wet in everyday language. |
| 49 | +""" |
| 50 | + |
| 51 | +_MINE_RESPONSES = { |
| 52 | + "What is 2 + 2?": ["4.", "The sum of two and two is four."], |
| 53 | + "What is the capital of France?": [ |
| 54 | + "Paris.", |
| 55 | + "The capital of France is Paris.", |
| 56 | + ], |
| 57 | + "What color is grass?": ["Green.", "Grass is usually green."], |
| 58 | + "What is 10 - 3?": ["7.", "Ten minus three equals seven."], |
| 59 | +} |
| 60 | + |
| 61 | +_HELD_OUT_PAIRS = ( |
| 62 | + ("What is 8 + 1?", "9.", "The result of adding eight and one is nine."), |
| 63 | + ("What color is snow?", "White.", "Snow is typically white in daylight."), |
| 64 | + ("What is the capital of Italy?", "Rome.", "The capital city of Italy is Rome."), |
| 65 | +) |
| 66 | + |
| 67 | + |
| 68 | +class _FakeMiningBackend: |
| 69 | + def __init__(self, responses: dict[str, list[str]]) -> None: |
| 70 | + self._responses = {prompt: deque(items) for prompt, items in responses.items()} |
| 71 | + |
| 72 | + def load(self, spec: object, store: object, *, adapter_name: str | None = None) -> None: |
| 73 | + _ = spec, store, adapter_name |
| 74 | + |
| 75 | + def generate(self, prompt: str, **_kwargs: object) -> str: |
| 76 | + return self._responses[prompt].popleft() |
| 77 | + |
| 78 | + def unload(self) -> None: |
| 79 | + return None |
| 80 | + |
| 81 | + |
| 82 | +class _TerseJudge: |
| 83 | + name = "cli:terse-judge" |
| 84 | + suggested_threshold = 0.1 |
| 85 | + |
| 86 | + def score_pair(self, prompt: str, candidate_a: str, candidate_b: str) -> PairScore: |
| 87 | + _ = prompt |
| 88 | + return PairScore(score_a=-float(len(candidate_a)), score_b=-float(len(candidate_b))) |
| 89 | + |
| 90 | + |
| 91 | +def _copy_fixture_store( |
| 92 | + trained_store: TrainedStoreHandle, |
| 93 | + *, |
| 94 | + tmp_path: Path, |
| 95 | + monkeypatch: pytest.MonkeyPatch, |
| 96 | +) -> tuple[Path, object]: |
| 97 | + from dlm.store.manifest import load_manifest, save_manifest |
| 98 | + from dlm.store.paths import for_dlm |
| 99 | + |
| 100 | + home = tmp_path / "home" |
| 101 | + home.mkdir() |
| 102 | + monkeypatch.setenv("DLM_HOME", str(home)) |
| 103 | + |
| 104 | + source_doc = trained_store.doc |
| 105 | + doc = home / source_doc.name |
| 106 | + shutil.copy2(source_doc, doc) |
| 107 | + |
| 108 | + parsed = parse_file(doc) |
| 109 | + store = for_dlm(parsed.frontmatter.dlm_id) |
| 110 | + shutil.copytree(trained_store.store.root, store.root, dirs_exist_ok=True) |
| 111 | + |
| 112 | + manifest = load_manifest(store.manifest) |
| 113 | + save_manifest( |
| 114 | + store.manifest, |
| 115 | + manifest.model_copy(update={"source_path": doc.resolve()}), |
| 116 | + ) |
| 117 | + return doc, store |
| 118 | + |
| 119 | + |
| 120 | +def _prepare_doc_for_cycle(doc: Path) -> None: |
| 121 | + current = doc.read_text(encoding="utf-8") |
| 122 | + doc.write_text(current.rstrip() + "\n\n" + _EXTRA_BODY.lstrip(), encoding="utf-8") |
| 123 | + |
| 124 | + parsed = parse_file(doc) |
| 125 | + new_pref = parsed.frontmatter.training.preference.model_copy( |
| 126 | + update={"method": "orpo", "enabled": True} |
| 127 | + ) |
| 128 | + new_training = parsed.frontmatter.training.model_copy(update={"preference": new_pref}) |
| 129 | + rewritten = ParsedDlm( |
| 130 | + frontmatter=parsed.frontmatter.model_copy(update={"training": new_training}), |
| 131 | + sections=parsed.sections, |
| 132 | + ) |
| 133 | + doc.write_text(serialize(rewritten), encoding="utf-8") |
| 134 | + |
| 135 | + |
| 136 | +def _patch_mining(monkeypatch: pytest.MonkeyPatch) -> None: |
| 137 | + monkeypatch.setattr( |
| 138 | + "dlm.inference.backends.select_backend", |
| 139 | + lambda *args, **kwargs: "pytorch", |
| 140 | + ) |
| 141 | + monkeypatch.setattr( |
| 142 | + "dlm.inference.backends.build_backend", |
| 143 | + lambda *args, **kwargs: _FakeMiningBackend(_MINE_RESPONSES), |
| 144 | + ) |
| 145 | + monkeypatch.setattr( |
| 146 | + "dlm.preference.build_judge", |
| 147 | + lambda *args, **kwargs: _TerseJudge(), |
| 148 | + ) |
| 149 | + |
| 150 | + |
| 151 | +def _mean_margin_for_version(doc: Path, store: object, version: int) -> float: |
| 152 | + target = store.adapter_version(version) |
| 153 | + original = store.resolve_current_adapter() |
| 154 | + assert original is not None |
| 155 | + store.set_current_adapter(target) |
| 156 | + try: |
| 157 | + judge = SwayJudge(doc) |
| 158 | + margins = [ |
| 159 | + judge.score_pair(prompt, chosen, rejected).margin |
| 160 | + for prompt, chosen, rejected in _HELD_OUT_PAIRS |
| 161 | + ] |
| 162 | + except (JudgeUnavailableError, JudgeInvocationError) as exc: |
| 163 | + pytest.skip(f"sway judge unavailable for mine-cycle proof: {exc}") |
| 164 | + finally: |
| 165 | + store.set_current_adapter(original) |
| 166 | + return sum(margins) / len(margins) |
| 167 | + |
| 168 | + |
| 169 | +@pytest.mark.slow |
| 170 | +def test_preference_mine_cycle_improves_held_out_sway_margin( |
| 171 | + trained_store: TrainedStoreHandle, |
| 172 | + tmp_path: Path, |
| 173 | + monkeypatch: pytest.MonkeyPatch, |
| 174 | +) -> None: |
| 175 | + from dlm.base_models import resolve as resolve_base_model |
| 176 | + from dlm.doc.sections import SectionType |
| 177 | + from dlm.store.manifest import load_manifest |
| 178 | + from dlm.train.preference.phase_orchestrator import run_phases |
| 179 | + |
| 180 | + doc, store = _copy_fixture_store(trained_store, tmp_path=tmp_path, monkeypatch=monkeypatch) |
| 181 | + _prepare_doc_for_cycle(doc) |
| 182 | + |
| 183 | + parsed = parse_file(doc) |
| 184 | + spec = resolve_base_model(parsed.frontmatter.base_model, accept_license=True) |
| 185 | + plan = trained_store.plan |
| 186 | + capabilities = trained_store.capabilities |
| 187 | + |
| 188 | + baseline = run_phases( |
| 189 | + store, |
| 190 | + parsed, |
| 191 | + spec, |
| 192 | + plan, |
| 193 | + phase="preference", |
| 194 | + capabilities=capabilities, |
| 195 | + lock_mode="ignore", |
| 196 | + seed=42, |
| 197 | + max_steps=20, |
| 198 | + ) |
| 199 | + assert [result.phase for result in baseline] == ["preference"] |
| 200 | + assert baseline[0].result.adapter_version == 2 |
| 201 | + |
| 202 | + _patch_mining(monkeypatch) |
| 203 | + runner = CliRunner() |
| 204 | + mine_result = runner.invoke( |
| 205 | + app, |
| 206 | + [ |
| 207 | + "--home", |
| 208 | + str(tmp_path / "home"), |
| 209 | + "preference", |
| 210 | + "mine", |
| 211 | + str(doc), |
| 212 | + "--samples", |
| 213 | + "2", |
| 214 | + "--max-pairs", |
| 215 | + "4", |
| 216 | + "--apply", |
| 217 | + ], |
| 218 | + ) |
| 219 | + assert mine_result.exit_code == 0, mine_result.output |
| 220 | + |
| 221 | + mined_doc = parse_file(doc) |
| 222 | + auto_mined_sections = [ |
| 223 | + section |
| 224 | + for section in mined_doc.sections |
| 225 | + if section.type is SectionType.PREFERENCE and section.auto_mined |
| 226 | + ] |
| 227 | + assert len(auto_mined_sections) == 4 |
| 228 | + |
| 229 | + final = run_phases( |
| 230 | + store, |
| 231 | + mined_doc, |
| 232 | + spec, |
| 233 | + plan, |
| 234 | + phase="preference", |
| 235 | + capabilities=capabilities, |
| 236 | + lock_mode="ignore", |
| 237 | + seed=42, |
| 238 | + max_steps=20, |
| 239 | + ) |
| 240 | + assert [result.phase for result in final] == ["preference"] |
| 241 | + assert final[0].result.adapter_version == 3 |
| 242 | + |
| 243 | + manifest = load_manifest(store.manifest) |
| 244 | + assert manifest.adapter_version == 3 |
| 245 | + assert len(manifest.training_runs) >= 3 |
| 246 | + |
| 247 | + baseline_margin = _mean_margin_for_version(doc, store, 2) |
| 248 | + final_margin = _mean_margin_for_version(doc, store, 3) |
| 249 | + assert final_margin > baseline_margin, ( |
| 250 | + "expected final preference-tuned adapter to improve held-out sway margin " |
| 251 | + f"(baseline={baseline_margin:.4f}, final={final_margin:.4f})" |
| 252 | + ) |
