code/set4_multiblimp.py

"""SET 4, ACCURACY arm: MultiBLiMP 1.0 (jumelet/multiblimp) on the Goldfish merges.

The Δfloor tables for SET 4 are likelihood only. This asks the accuracy question directly, on the
same merges: does anything survive as *grammatical competence*? Minimal pairs are `sen` (grammatical)
vs `wrong_sen`; a model is correct when it assigns the higher total log-probability to `sen`.
Chance = 0.500.

CAVEAT built into the design: the merged models live in the ENGLISH parent's token-id space, so
partner-language items must be tokenized with the English tokenizer. The per-cell UNK rate is
reported alongside every number; where it is high (Greek) the partner-language accuracy is not
interpretable as grammatical competence and is marked as such."""
import os, sys, json, time, csv, argparse, gc
sys.path.insert(0, "/root/compose-audit")
from common import *
import gpt2_align as G2
from mergeschool.core.models import load_hf
from huggingface_hub import hf_hub_download

ap = argparse.ArgumentParser()
ap.add_argument("--pairs", default="nld_Latn:nld_latn:nld,spa_Latn:spa_latn:spa,ell_Grek:ell_grek:ell,pol_Latn:pol_latn:pol")
ap.add_argument("--n_sent", type=int, default=500)
ap.add_argument("--max_items", type=int, default=1200)
ap.add_argument("--bs", type=int, default=48)
A = ap.parse_args()
OUT = "/root/compose-audit/results/set4_multiblimp.jsonl"
DEV = "cuda"
ENG_REPO = "goldfish-models/eng_latn_1000mb"


def log(*a):
    print(f"[{time.strftime('%H:%M:%S')}]", *a, flush=True)


def mb_pairs(lang, n):
    p = hf_hub_download("jumelet/multiblimp", f"{lang}/data.tsv", repo_type="dataset")
    rows = list(csv.DictReader(open(p, encoding="utf-8"), delimiter="\t"))[:n]
    return [(r["sen"], r["wrong_sen"], r.get("phenomenon", "?")) for r in rows
            if r.get("sen") and r.get("wrong_sen")]


def encode(tok, sents, maxlen=64):
    e = tok(sents, return_tensors="pt", padding=True, truncation=True, max_length=maxlen)
    return e["input_ids"], e["attention_mask"]


@torch.no_grad()
def score(model, ids, am, dev, bs):
    out = []
    for i in range(0, ids.shape[0], bs):
        x, m = ids[i:i + bs].to(dev), am[i:i + bs].to(dev)
        lp = torch.log_softmax(model(x, attention_mask=m).logits.float()[:, :-1], -1)
        out.append((lp.gather(-1, x[:, 1:].unsqueeze(-1)).squeeze(-1) * m[:, 1:].float()).sum(1).cpu())
    return torch.cat(out).numpy()


def unk_rate(tok, ids, am):
    u = tok.unk_token_id
    if u is None: return 0.0
    return float(((ids == u) & (am.bool())).sum().item() / max(1, am.sum().item()))


log("loading eng parent")
m_e, tok_e = load_hf(ENG_REPO, dtype=torch.float32, device=DEV); m_e.eval()
cfg = m_e.config
D, NH, NL, V = cfg.n_embd, cfg.n_head, cfg.n_layer, cfg.vocab_size
SD_E = sd_np(m_e)
shell = m_e

sys.path.insert(0, "/root/compose-audit")
from set4_goldfish_lib import sent_acts   # noqa

eng_lines = flores_lines("eng_Latn")[: A.n_sent]
acts_e = sent_acts(m_e, tok_e, eng_lines, DEV)
MB_ENG = mb_pairs("eng", A.max_items)
ENC_ENG_E = (encode(tok_e, [g for g, b, p in MB_ENG]), encode(tok_e, [b for g, b, p in MB_ENG]))
log(f"MultiBLiMP eng items={len(MB_ENG)}  UNK(eng tok)={unk_rate(tok_e, *ENC_ENG_E[0]):.2%}")


def acc(sd, enc_g, enc_b):
    sd_load(shell, sd, DEV)
    sg = score(shell, *enc_g, DEV, A.bs); sb = score(shell, *enc_b, DEV, A.bs)
    return float((sg > sb).mean())


ACC_E_ENG = acc(SD_E, *ENC_ENG_E)
log(f"eng parent, MultiBLiMP-eng = {ACC_E_ENG:.4f}")

done = set()
if os.path.exists(OUT):
    for l in open(OUT):
        try: done.add(json.loads(l)["lang"])
        except Exception: pass
fh = open(OUT, "a")

for spec in A.pairs.split(","):
    fcode, gcode, mb = spec.split(":")
    if fcode in done: continue
    t0 = time.time()
    repo = f"goldfish-models/{gcode}_1000mb"
    log(f"=== {fcode} <- {repo}")
    m_x, tok_x = load_hf(repo, dtype=torch.float32, device=DEV); m_x.eval()
    SD_X = sd_np(m_x)
    x_lines = flores_lines(fcode)[: A.n_sent]
    acts_x = sent_acts(m_x, tok_x, x_lines, DEV)
    MB_X = mb_pairs(mb, A.max_items)
    ENC_X_X = (encode(tok_x, [g for g, b, p in MB_X]), encode(tok_x, [b for g, b, p in MB_X]))
    ENC_X_E = (encode(tok_e, [g for g, b, p in MB_X]), encode(tok_e, [b for g, b, p in MB_X]))
    unk_x_e = unk_rate(tok_e, *ENC_X_E[0]); unk_x_x = unk_rate(tok_x, *ENC_X_X[0])
    sg = score(m_x, *ENC_X_X[0], DEV, A.bs); sb = score(m_x, *ENC_X_X[1], DEV, A.bs)
    ACC_X_X = float((sg > sb).mean())
    del m_x; torch.cuda.empty_cache()
    ACC_E_X = acc(SD_E, *ENC_X_E)                     # English parent on partner-language items
    log(f"  items={len(MB_X)} UNK(eng tok on {mb})={unk_x_e:.2%}  X parent MB-{mb}={ACC_X_X:.4f}  "
        f"eng parent MB-{mb}={ACC_E_X:.4f}  (chance 0.5)")

    vkeys = [k for k in SD_X if k.endswith("wte.weight") or k.endswith("lm_head.weight")]
    SD_X_V, cov = AL.remap_vocab_rows(SD_X, tok_e, tok_x, V, keys=vkeys)
    for k in vkeys:
        W = np.asarray(SD_X_V[k], float)
        if W.shape[0] == V:
            bad = ~np.isfinite(W).all(axis=1); W[bad] = np.asarray(SD_E[k], float)[bad]
        SD_X_V[k] = W
    BODY = [k for k in SD_E if not (k.endswith("wte.weight") or k.endswith("lm_head.weight"))]
    R_emb, n_anch = G2.emb_procrustes(SD_E, SD_X, tok_e, tok_x)
    sd_emb = G2.apply_resid(SD_X_V, D, R=R_emb)
    sd_emb2, _ = G2.align_full(SD_E, sd_emb, D, NH, None, None, "permutation", body_keys=BODY)
    sdp, ip = G2.align_full(SD_E, SD_X_V, D, NH, acts_e, acts_x, "permutation", body_keys=BODY)
    sdo, io = G2.align_full(SD_E, SD_X_V, D, NH, acts_e, acts_x, "orthogonal", body_keys=BODY)
    sdof, _ = G2.align_full(SD_E, SD_X_V, D, NH, acts_e, acts_x, "orthogonal", body_keys=BODY, accept_each=False)

    rungs = {"M0_naive_avg": MG.average([SD_E, SD_X]),
             "M1a_vocab_avg": MG.average([SD_E, SD_X_V]),
             "M1b_vocab_perm_avg": MG.average([SD_E, sdp]),
             "M1c_vocab_orth_avg": MG.average([SD_E, sdo]),
             "M1e_vocab_orth_forced": MG.average([SD_E, sdof]),
             "M1g_emb_procrustes": MG.average([SD_E, sd_emb]),
             "M1h_emb_proc_units": MG.average([SD_E, sd_emb2])}
    res = {}
    for k, sd in rungs.items():
        res[k] = {"mb_eng": acc(sd, *ENC_ENG_E), "mb_x": acc(sd, *ENC_X_E)}
        res[k]["delta_eng_vs_eng_parent"] = res[k]["mb_eng"] - ACC_E_ENG
        res[k]["delta_x_vs_x_parent"] = res[k]["mb_x"] - ACC_X_X

    r = {"set": "set4_multiblimp", "lang": fcode, "mb_lang": mb, "repo_b": repo,
         "metric": "MultiBLiMP 1.0 accuracy (chance=0.5) -- ACCURACY, not likelihood",
         "n_items_eng": len(MB_ENG), "n_items_x": len(MB_X),
         "unk_rate_eng_tok_on_x_items": unk_x_e, "unk_rate_own_tok_on_x_items": unk_x_x,
         "parents": {"eng_on_mb_eng": ACC_E_ENG, "x_on_mb_x": ACC_X_X, "eng_on_mb_x": ACC_E_X},
         "rungs": res, "align_info": {"perm": ip, "orth": io}, "secs": time.time() - t0}
    fh.write(json.dumps(r) + "\n"); fh.flush()
    log("  " + "  ".join(f"{k}: eng={v['mb_eng']:.3f} x={v['mb_x']:.3f}" for k, v in res.items()))
    del rungs, sdp, sdo, sdof, sd_emb, sd_emb2, SD_X, SD_X_V; gc.collect()
fh.close()
log("DONE set4_multiblimp")