@@ -0,0 +1,207 @@ |
| 1 | +"""F11 prove-the-value: mined paraphrases flip a memorizing adapter's verdict. |
| 2 | + |
| 3 | +``paraphrase_invariance`` asks "does the adapter lift the gold answer |
| 4 | +equally when the prompt is paraphrased?" A **memorizing** adapter |
| 5 | +passes cleanly when the hand-written paraphrase list consists of |
| 6 | +near-templated rewordings of the seed prompt (the adapter memorized |
| 7 | +the seed and generalizes the templated tweaks). The **miner** searches |
| 8 | +further out — semantically different rewordings — and surfaces the |
| 9 | +paraphrases the adapter *doesn't* lift. |
| 10 | + |
| 11 | +This test plants exactly that scenario: |
| 12 | + |
| 13 | +1. **Hand-written paraphrases** are all close-template rewordings. |
| 14 | + The memorizing adapter lifts them ≈ verbatim lift → high |
| 15 | + ``generalization_ratio`` → PASS. |
| 16 | +2. **Mined candidates** include semantically distant rewordings the |
| 17 | + memorizing adapter doesn't lift. The miner ranks those first. |
| 18 | +3. Substitute the mined paraphrases into the probe's spec and re-run. |
| 19 | + ``generalization_ratio`` collapses → verdict flips to FAIL. |
| 20 | + |
| 21 | +The F11 claim, reified: the mined list surfaces a concrete gap the |
| 22 | +hand-written list missed entirely. |
| 23 | +""" |
| 24 | + |
| 25 | +from __future__ import annotations |
| 26 | + |
| 27 | +import numpy as np |
| 28 | +import pytest |
| 29 | + |
| 30 | +from dlm_sway.backends.dummy import DummyDifferentialBackend, DummyResponses |
| 31 | +from dlm_sway.core.result import Verdict |
| 32 | +from dlm_sway.mining.paraphrase_miner import mine_paraphrases |
| 33 | +from dlm_sway.probes.base import RunContext, build_probe |
| 34 | + |
| 35 | +# --------------------------------------------------------------------- |
| 36 | +# Scenario constants — a memorizing adapter on one seed case. |
| 37 | +# --------------------------------------------------------------------- |
| 38 | + |
| 39 | +SEED_PROMPT = "The capital of France is" |
| 40 | +GOLD = " Paris" |
| 41 | + |
| 42 | +# Hand-written paraphrases — the kind a well-meaning user types. Close |
| 43 | +# to the seed, mostly templated rewordings. |
| 44 | +HAND_WRITTEN = [ |
| 45 | + "Capital of France:", |
| 46 | + "France capital equals", |
| 47 | +] |
| 48 | + |
| 49 | +# Candidates the miner will pull (stubbed nlpaug output). Mix of |
| 50 | +# near-templates (easy) and semantically distant rewordings (hard). |
| 51 | +MINER_CANDIDATES = [ |
| 52 | + "What is the capital of France?", # near |
| 53 | + "Tell me about French capital", # distant |
| 54 | + "Which city governs France?", # distant |
| 55 | + "Name the primary city in France", # distant |
| 56 | +] |
| 57 | + |
| 58 | +# Token-lift model: memorizing adapter lifts verbatim + near-templates; |
| 59 | +# doesn't lift semantically distant rewordings. |
| 60 | +VERBATIM_BASE_LP = -3.0 # per-token logprob on base |
| 61 | +VERBATIM_FT_LP = -0.5 # per-token logprob on ft — big lift |
| 62 | +NEAR_BASE_LP = -3.0 |
| 63 | +NEAR_FT_LP = -1.0 # moderate lift (still pattern-matched) |
| 64 | +DISTANT_BASE_LP = -3.0 |
| 65 | +DISTANT_FT_LP = -3.0 # no lift — adapter doesn't recognize |
| 66 | + |
| 67 | +# Token count estimate: len(gold)//4 = 1 for " Paris"; we need a |
| 68 | +# meaningful multiplier so the per-token logprobs translate to |
| 69 | +# interpretable lifts. The probe multiplies logprob by token count; |
| 70 | +# here the gold is 6 chars → 1 token, so per-token == total. |
| 71 | + |
| 72 | + |
| 73 | +def _prompt_lp_base(prompt: str) -> float: |
| 74 | + """Backend's base-side logprob of ``(prompt, GOLD)``. Mirrors the |
| 75 | + probe's own per-token normalization.""" |
| 76 | + return VERBATIM_BASE_LP |
| 77 | + |
| 78 | + |
| 79 | +def _prompt_lp_ft(prompt: str) -> float: |
| 80 | + """ft-side logprob: verbatim + near-templates get lifted; distant |
| 81 | + rewordings don't.""" |
| 82 | + if prompt == SEED_PROMPT: |
| 83 | + return VERBATIM_FT_LP |
| 84 | + if prompt in {"Capital of France:", "France capital equals"}: |
| 85 | + return NEAR_FT_LP |
| 86 | + if prompt == "What is the capital of France?": |
| 87 | + return NEAR_FT_LP |
| 88 | + return DISTANT_FT_LP |
| 89 | + |
| 90 | + |
| 91 | +def _memorizing_backend(prompts: list[str]) -> DummyDifferentialBackend: |
| 92 | + base_lp = {(p, GOLD): _prompt_lp_base(p) for p in prompts} |
| 93 | + ft_lp = {(p, GOLD): _prompt_lp_ft(p) for p in prompts} |
| 94 | + return DummyDifferentialBackend( |
| 95 | + base=DummyResponses(logprobs=base_lp), |
| 96 | + ft=DummyResponses(logprobs=ft_lp), |
| 97 | + ) |
| 98 | + |
| 99 | + |
| 100 | +def _stub_embedder(monkeypatch: pytest.MonkeyPatch) -> None: |
| 101 | + """Stub the MiniLM embedder — every candidate gets a unique |
| 102 | + orthogonal embedding so the diversity filter keeps all of them |
| 103 | + (and the ranker's decisions are what the test measures).""" |
| 104 | + table = { |
| 105 | + SEED_PROMPT: np.array([1.0, 0.0, 0.0, 0.0, 0.0], dtype=np.float32), |
| 106 | + "What is the capital of France?": np.array([0.0, 1.0, 0.0, 0.0, 0.0], dtype=np.float32), |
| 107 | + "Tell me about French capital": np.array([0.0, 0.0, 1.0, 0.0, 0.0], dtype=np.float32), |
| 108 | + "Which city governs France?": np.array([0.0, 0.0, 0.0, 1.0, 0.0], dtype=np.float32), |
| 109 | + "Name the primary city in France": np.array([0.0, 0.0, 0.0, 0.0, 1.0], dtype=np.float32), |
| 110 | + } |
| 111 | + |
| 112 | + def _encode(texts: list[str]) -> np.ndarray: |
| 113 | + return np.stack([table[t] for t in texts]) |
| 114 | + |
| 115 | + monkeypatch.setattr( |
| 116 | + "dlm_sway.mining.paraphrase_miner._load_embedder", |
| 117 | + lambda _model_id: _encode, # type: ignore[arg-type] |
| 118 | + ) |
| 119 | + |
| 120 | + |
| 121 | +def _run_probe(paraphrases: list[str], all_prompts: list[str]) -> tuple[Verdict, float]: |
| 122 | + """Run paraphrase_invariance with the given paraphrase list and |
| 123 | + return the verdict + the generalization_ratio for the case.""" |
| 124 | + backend = _memorizing_backend(all_prompts) |
| 125 | + probe, spec = build_probe( |
| 126 | + { |
| 127 | + "name": "pi", |
| 128 | + "kind": "paraphrase_invariance", |
| 129 | + "cases": [ |
| 130 | + { |
| 131 | + "prompt": SEED_PROMPT, |
| 132 | + "gold": GOLD, |
| 133 | + "paraphrases": paraphrases, |
| 134 | + }, |
| 135 | + ], |
| 136 | + "intent": "generalize", |
| 137 | + # Default threshold is 0.5 — keep it explicit for the assertion. |
| 138 | + "min_generalization_ratio": 0.5, |
| 139 | + "min_verbatim_lift": 0.2, |
| 140 | + } |
| 141 | + ) |
| 142 | + ctx = RunContext(backend=backend) |
| 143 | + result = probe.run(spec, ctx) |
| 144 | + ratio = float(result.evidence["generalization_ratio"]) |
| 145 | + return result.verdict, ratio |
| 146 | + |
| 147 | + |
| 148 | +def test_mined_paraphrases_flip_memorizing_adapter_from_pass_to_fail( |
| 149 | + monkeypatch: pytest.MonkeyPatch, |
| 150 | +) -> None: |
| 151 | + """The F11 prove-the-value demonstration in concrete form.""" |
| 152 | + _stub_embedder(monkeypatch) |
| 153 | + |
| 154 | + # 1. Hand-written paraphrases — the memorizing adapter passes. |
| 155 | + all_prompts_hand = [SEED_PROMPT, *HAND_WRITTEN] |
| 156 | + hand_verdict, hand_ratio = _run_probe(HAND_WRITTEN, all_prompts_hand) |
| 157 | + assert hand_verdict == Verdict.PASS, ( |
| 158 | + f"memorizing adapter should pass on close-template paraphrases; " |
| 159 | + f"got verdict={hand_verdict}, ratio={hand_ratio:.3f}" |
| 160 | + ) |
| 161 | + # Generalization_ratio is well above the 0.5 threshold. |
| 162 | + assert hand_ratio > 0.5, hand_ratio |
| 163 | + |
| 164 | + # 2. Mine paraphrases — the miner pulls candidates including |
| 165 | + # semantically distant ones and ranks them by gap. |
| 166 | + miner_backend = _memorizing_backend([SEED_PROMPT, *HAND_WRITTEN, *MINER_CANDIDATES]) |
| 167 | + |
| 168 | + def _canned(_prompt: str, *, n: int, seed: int) -> list[str]: |
| 169 | + del n, seed |
| 170 | + return list(MINER_CANDIDATES) |
| 171 | + |
| 172 | + mined = mine_paraphrases( |
| 173 | + prompt=SEED_PROMPT, |
| 174 | + gold=GOLD, |
| 175 | + backend=miner_backend, |
| 176 | + generate_candidates=_canned, |
| 177 | + n_candidates=4, |
| 178 | + top_k=3, |
| 179 | + seed=0, |
| 180 | + ) |
| 181 | + |
| 182 | + # The mined list starts with the semantically-distant rewordings |
| 183 | + # (the adapter doesn't lift them → largest gap). |
| 184 | + mined_paraphrases = [c.prompt for c in mined.candidates] |
| 185 | + assert mined_paraphrases[0] in { |
| 186 | + "Tell me about French capital", |
| 187 | + "Which city governs France?", |
| 188 | + "Name the primary city in France", |
| 189 | + }, f"expected a distant reworking at rank 0; got {mined_paraphrases}" |
| 190 | + |
| 191 | + # 3. Re-run paraphrase_invariance with the mined paraphrases — |
| 192 | + # verdict must flip to FAIL. |
| 193 | + all_prompts_mined = [SEED_PROMPT, *mined_paraphrases] |
| 194 | + mined_verdict, mined_ratio = _run_probe(mined_paraphrases, all_prompts_mined) |
| 195 | + assert mined_verdict == Verdict.FAIL, ( |
| 196 | + f"mined paraphrases should flip the memorizing adapter's verdict; " |
| 197 | + f"got verdict={mined_verdict}, ratio={mined_ratio:.3f}" |
| 198 | + ) |
| 199 | + # The generalization_ratio collapses well below the 0.5 threshold. |
| 200 | + assert mined_ratio < 0.5, mined_ratio |
| 201 | + |
| 202 | + # And the ratio gap is meaningful — this is the F11 headline number: |
| 203 | + # mined list surfaces a generalization gap the hand-list missed. |
| 204 | + assert hand_ratio - mined_ratio > 0.3, ( |
| 205 | + f"expected ≥0.3 ratio gap between hand-list and mined-list; " |
| 206 | + f"got hand={hand_ratio:.3f}, mined={mined_ratio:.3f}" |
| 207 | + ) |
