| 1 | """Tests for :mod:`dlm_sway.mining.outlier_miner`. |
| 2 | |
| 3 | Uses the dummy backend's synthesized per-prompt divergences: base is |
| 4 | sharply peaked, ft is broad, and the dummy's ``next_token_dist`` cache |
| 5 | keys on the prompt string — so each candidate prompt produces the same |
| 6 | divergence unless we overlay per-prompt TokenDists. Tests that need |
| 7 | variation build per-prompt TokenDists directly on ``DummyResponses``. |
| 8 | """ |
| 9 | |
| 10 | from __future__ import annotations |
| 11 | |
| 12 | import math |
| 13 | |
| 14 | import numpy as np |
| 15 | import pytest |
| 16 | |
| 17 | from dlm_sway.backends.dummy import DummyDifferentialBackend, DummyResponses |
| 18 | from dlm_sway.core.scoring import TokenDist |
| 19 | from dlm_sway.mining.outlier_miner import ( |
| 20 | OutlierCandidate, |
| 21 | OutlierResult, |
| 22 | corpus_prompts, |
| 23 | mine_outliers, |
| 24 | ) |
| 25 | |
| 26 | |
| 27 | def _dist_from_probs(probs: list[float]) -> TokenDist: |
| 28 | arr = np.asarray(probs, dtype=np.float64) |
| 29 | arr = arr / arr.sum() |
| 30 | lp = np.log(arr).astype(np.float32) |
| 31 | return TokenDist( |
| 32 | token_ids=np.arange(len(probs), dtype=np.int64), |
| 33 | logprobs=lp, |
| 34 | vocab_size=max(1000, len(probs)), |
| 35 | tail_logprob=None, |
| 36 | ) |
| 37 | |
| 38 | |
| 39 | class TestMineOutliers: |
| 40 | def test_ranks_prompts_by_per_prompt_divergence(self) -> None: |
| 41 | """Six prompts with planted divergences: ``hi*`` have the biggest |
| 42 | gap, ``lo*`` the smallest, ``mid*`` in between. Top-K = hi rows, |
| 43 | bottom-K = lo rows.""" |
| 44 | base = _dist_from_probs([0.92, 0.02, 0.02, 0.02, 0.02]) |
| 45 | ft_flat = _dist_from_probs([0.25, 0.20, 0.20, 0.20, 0.15]) # big KL |
| 46 | ft_mild = _dist_from_probs([0.70, 0.10, 0.10, 0.05, 0.05]) # mid KL |
| 47 | ft_same = base # zero KL |
| 48 | |
| 49 | # F04 — need ≥ 2·top_k=4 distinct prompts to clear the guard. |
| 50 | prompts = ["hi1", "hi2", "mid1", "mid2", "lo1", "lo2"] |
| 51 | base_dists = dict.fromkeys(prompts, base) |
| 52 | ft_dists = { |
| 53 | "hi1": ft_flat, |
| 54 | "hi2": ft_flat, |
| 55 | "mid1": ft_mild, |
| 56 | "mid2": ft_mild, |
| 57 | "lo1": ft_same, |
| 58 | "lo2": ft_same, |
| 59 | } |
| 60 | backend = DummyDifferentialBackend( |
| 61 | base=DummyResponses(token_dists=base_dists), |
| 62 | ft=DummyResponses(token_dists=ft_dists), |
| 63 | ) |
| 64 | |
| 65 | result = mine_outliers( |
| 66 | probe_kind="delta_kl", |
| 67 | candidate_prompts=prompts, |
| 68 | backend=backend, |
| 69 | top_k=2, |
| 70 | ) |
| 71 | |
| 72 | assert isinstance(result, OutlierResult) |
| 73 | assert result.probe_kind == "delta_kl" |
| 74 | # Top is most-positive first; bottom is least-positive first. |
| 75 | top_prompts = {c.prompt for c in result.top} |
| 76 | bottom_prompts = {c.prompt for c in result.bottom} |
| 77 | assert top_prompts == {"hi1", "hi2"} |
| 78 | assert bottom_prompts == {"lo1", "lo2"} |
| 79 | # Raw values are finite and positive (JS divergence ≥ 0). |
| 80 | for c in result.top: |
| 81 | assert math.isfinite(c.raw) |
| 82 | assert c.raw >= 0.0 |
| 83 | |
| 84 | def test_small_pool_raises_f04_guard(self) -> None: |
| 85 | """F04 (Audit 03) — pool below ``2·top_k`` distinct prompts |
| 86 | raises SwayError with an actionable hint. Replaces pre-F04 |
| 87 | 'test_top_k_clipped_to_pool_size' which relied on the same |
| 88 | degenerate single-prompt case the audit flagged as produced |
| 89 | top=[p], bottom=[p] — identical lists.""" |
| 90 | from dlm_sway.core.errors import SwayError |
| 91 | |
| 92 | base = _dist_from_probs([0.92, 0.02, 0.02, 0.02, 0.02]) |
| 93 | ft = _dist_from_probs([0.25, 0.20, 0.20, 0.20, 0.15]) |
| 94 | backend = DummyDifferentialBackend( |
| 95 | base=DummyResponses(token_dists={"p": base}), |
| 96 | ft=DummyResponses(token_dists={"p": ft}), |
| 97 | ) |
| 98 | with pytest.raises(SwayError, match="below the 2·top_k"): |
| 99 | mine_outliers( |
| 100 | probe_kind="delta_kl", |
| 101 | candidate_prompts=["p"], |
| 102 | backend=backend, |
| 103 | top_k=10, |
| 104 | ) |
| 105 | |
| 106 | def test_small_pool_error_suggests_smaller_top_k(self) -> None: |
| 107 | """The error message includes a concrete ``--top-k N`` hint the |
| 108 | user can copy into their CLI invocation.""" |
| 109 | from dlm_sway.core.errors import SwayError |
| 110 | |
| 111 | base = _dist_from_probs([0.92, 0.02, 0.02, 0.02, 0.02]) |
| 112 | ft = _dist_from_probs([0.25, 0.20, 0.20, 0.20, 0.15]) |
| 113 | prompts = ["p1", "p2", "p3"] |
| 114 | backend = DummyDifferentialBackend( |
| 115 | base=DummyResponses(token_dists=dict.fromkeys(prompts, base)), |
| 116 | ft=DummyResponses(token_dists=dict.fromkeys(prompts, ft)), |
| 117 | ) |
| 118 | with pytest.raises(SwayError, match="Pass --top-k 1"): |
| 119 | mine_outliers( |
| 120 | probe_kind="delta_kl", |
| 121 | candidate_prompts=prompts, |
| 122 | backend=backend, |
| 123 | top_k=5, |
| 124 | ) |
| 125 | |
| 126 | def test_empty_pool_returns_empty_result(self) -> None: |
| 127 | backend = DummyDifferentialBackend(base=DummyResponses(), ft=DummyResponses()) |
| 128 | result = mine_outliers( |
| 129 | probe_kind="delta_kl", |
| 130 | candidate_prompts=[], |
| 131 | backend=backend, |
| 132 | top_k=5, |
| 133 | ) |
| 134 | assert result.top == [] |
| 135 | assert result.bottom == [] |
| 136 | |
| 137 | def test_unsupported_probe_kind_returns_empty(self) -> None: |
| 138 | """Probes that need a non-``prompts`` spec (leakage, etc.) skip |
| 139 | every candidate silently. The F04 floor doesn't fire in that |
| 140 | case because the scored list is empty — empty-result path |
| 141 | preserved for the unsupported-kind UX.""" |
| 142 | base = _dist_from_probs([0.92, 0.02, 0.02, 0.02, 0.02]) |
| 143 | ft = _dist_from_probs([0.25, 0.20, 0.20, 0.20, 0.15]) |
| 144 | backend = DummyDifferentialBackend( |
| 145 | base=DummyResponses(token_dists={"p": base}), |
| 146 | ft=DummyResponses(token_dists={"p": ft}), |
| 147 | ) |
| 148 | result = mine_outliers( |
| 149 | probe_kind="leakage", |
| 150 | candidate_prompts=["p"], |
| 151 | backend=backend, |
| 152 | top_k=5, |
| 153 | ) |
| 154 | assert result.top == [] |
| 155 | assert result.bottom == [] |
| 156 | |
| 157 | def test_rejects_top_k_zero(self) -> None: |
| 158 | backend = DummyDifferentialBackend(base=DummyResponses(), ft=DummyResponses()) |
| 159 | with pytest.raises(ValueError, match="top_k must be positive"): |
| 160 | mine_outliers( |
| 161 | probe_kind="delta_kl", |
| 162 | candidate_prompts=["a"], |
| 163 | backend=backend, |
| 164 | top_k=0, |
| 165 | ) |
| 166 | |
| 167 | |
| 168 | class TestCorpusPrompts: |
| 169 | def test_pulls_chunks_from_public_domain(self) -> None: |
| 170 | """``--from-corpus public_domain_en`` yields a list of strings |
| 171 | long enough for probe scoring.""" |
| 172 | chunks = corpus_prompts("public_domain_en", chunk_chars=512, max_chunks=4) |
| 173 | assert chunks |
| 174 | assert len(chunks) <= 4 |
| 175 | for c in chunks: |
| 176 | assert isinstance(c, str) |
| 177 | assert len(c) >= 64 # chunk_corpus's minimum |
| 178 | |
| 179 | def test_unknown_corpus_raises(self) -> None: |
| 180 | with pytest.raises(KeyError): |
| 181 | corpus_prompts("doesnotexist") |
| 182 | |
| 183 | |
| 184 | class TestOutlierCandidate: |
| 185 | def test_is_frozen(self) -> None: |
| 186 | c = OutlierCandidate(prompt="p", raw=0.5, index=0) |
| 187 | with pytest.raises(Exception): # noqa: B017, PT011 |
| 188 | c.raw = 0.0 # type: ignore[misc] |