# test_nocot_control_exact.py: The no-CoT control differs from the treatment only by the think block. import json import pathlib import sys import pytest ROOT = pathlib.Path(__file__).resolve().parents[1] sys.path.insert(0, str(ROOT)) torch = pytest.importorskip("torch") @pytest.fixture(scope="module") def toks(): try: from transformers import AutoTokenizer from tokenizer import TokenRegistry, expand_tokenizer tok = AutoTokenizer.from_pretrained("EleutherAI/gpt-neox-20b") expand_tokenizer(tok) return tok, TokenRegistry(tok) except Exception as exc: pytest.skip(f"tokenizer unavailable: {exc}") def _corpus(tmp, n=6): from delay_dataset import NUM_LEVELS rows = [] for i in range(n): T = 30 + i rows.append({ "prompt": f"a matched control sentence number {i} here", "reasoning": f"Deliver line {i} with warmth and a rising close.", "emotion_label": "joyful", "pace": "fast", "pitch": "high", "speaker_id": 3, "cot_form": "prose", "speech_tokens": list(range(T)), "residual_codes": [[(k + 1) % 900 for _ in range(T)] for k in range(NUM_LEVELS - 1)], }) p = tmp / "c.jsonl" p.write_text("\n".join(json.dumps(r) for r in rows), encoding="utf-8") return p, rows def _ds(path, tok, reg, cot_mode): from delay_dataset import DelayMimiDataset return DelayMimiDataset( data_path=str(path), tokenizer=tok, registry=reg, max_seq_len=2048, aligned=True, stage="auto", cot_mode=cot_mode, lazy=False) def test_none_mode_equals_a_pure_stage1_build(tmp_path, toks): tok, reg = toks p, _ = _corpus(tmp_path) ctrl = _ds(p, tok, reg, "none") from delay_dataset import DelayMimiDataset stage1 = DelayMimiDataset( data_path=str(p), tokenizer=tok, registry=reg, max_seq_len=2048, aligned=True, stage=1, cot_mode="prose", lazy=False) for i in range(len(ctrl)): a, b = ctrl[i], stage1[i] assert a.keys() == b.keys(), f"row {i}: different fields" for k in a: va, vb = a[k], b[k] if torch.is_tensor(va): assert torch.equal(va, vb), ( f"row {i}: field {k!r} differs between cot_mode=none and " f"a pure stage-1 build — the control is not clean") else: assert va == vb, f"row {i}: field {k!r} differs" def test_treatment_and_control_differ_only_by_the_think_block(tmp_path, toks): tok, reg = toks p, _ = _corpus(tmp_path) treat = _ds(p, tok, reg, "prose") ctrl = _ds(p, tok, reg, "none") for i in range(len(treat)): t, c = treat[i], ctrl[i] ti, ci = t["input_ids"], c["input_ids"] assert len(ti) > len(ci), ( f"row {i}: prose mode did not lengthen the sequence — the " f"treatment is not adding a THINK block") assert reg.think_start_id in ti.tolist() assert reg.think_end_id in ti.tolist() assert reg.think_start_id not in ci.tolist(), \ f"row {i}: the CONTROL contains a THINK token" sa = [x for x in ti.tolist() if x >= reg.speech_token_id_min] sb = [x for x in ci.tolist() if x >= reg.speech_token_id_min] assert sa == sb, f"row {i}: the audio differs between the arms" def test_control_has_no_reasoning_supervision(tmp_path, toks): tok, reg = toks p, _ = _corpus(tmp_path) ctrl = _ds(p, tok, reg, "none") treat = _ds(p, tok, reg, "prose") lw = "loss_weights" if lw not in ctrl[0]: pytest.skip("collator does not expose per-token loss weights") for i in range(len(ctrl)): cw = set(ctrl[i][lw].tolist()) tw = set(treat[i][lw].tolist()) assert 0.3 not in cw or 0.3 in tw, "unexpected reasoning weight"