code/bgpt_ceiling.py

"""What would SUCCESS look like? The jointly-trained bilingual ceiling.

SET 4 shows that merging two monolingual Goldfish models produces a model that is destroyed by
likelihood and badly degraded by accuracy. That is only interpretable against what a bilingual model
of the same budget actually achieves. B-GPT (Arnett et al.) trains English+X jointly with a single
shared tokenizer — the target the composition literature is trying to reach without joint training.

Reports the same two metrics on the same two corpora: nats per UTF-8 byte on FLORES-200 devtest, and
MultiBLiMP 1.0 accuracy."""
import os, sys, json, time, csv, argparse
sys.path.insert(0, "/root/compose-audit")
from common import *
from mergeschool.core.models import load_hf
from huggingface_hub import hf_hub_download

ap = argparse.ArgumentParser()
ap.add_argument("--pairs", default="nld_Latn:nl:nld,spa_Latn:es:spa,ell_Grek:el:ell,pol_Latn:pl:pol")
ap.add_argument("--variant", default="simultaneous")
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=8)
A = ap.parse_args()
OUT = "/root/compose-audit/results/bgpt_ceiling.jsonl"
DEV = "cuda"


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


def build_blocks(tok, text, block=128, max_blocks=200):
    ids = tok(text)["input_ids"]
    n = max(1, min(max_blocks, len(ids) // block))
    arr = torch.from_numpy(np.asarray(ids[: n * block], dtype=np.int64).reshape(n, block))
    nb = sum(len(tok.decode(list(arr[i, 1:].numpy())).encode("utf-8")) for i in range(n))
    return arr, nb


@torch.no_grad()
def nll_total(model, blocks, dev, bs=8):
    tot, ntok = 0.0, 0
    for i in range(0, blocks.shape[0], bs):
        x = blocks[i:i + bs].to(dev)
        lp = torch.log_softmax(model(x).logits.float()[:, :-1], -1)
        tgt = x[:, 1:]
        tot += (-lp.gather(-1, tgt.unsqueeze(-1)).squeeze(-1)).sum().item(); ntok += tgt.numel()
    return tot, ntok


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"]) for r in rows if r.get("sen") and r.get("wrong_sen")]


@torch.no_grad()
def mb_acc(model, tok, pairs, dev, bs=48, maxlen=64):
    def sc(sents):
        e = tok(sents, return_tensors="pt", padding=True, truncation=True, max_length=maxlen)
        ids, am = e["input_ids"], e["attention_mask"]
        o = []
        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)
            o.append((lp.gather(-1, x[:, 1:].unsqueeze(-1)).squeeze(-1) * m[:, 1:].float()).sum(1).cpu())
        return torch.cat(o).numpy()
    sg, sb = sc([g for g, _ in pairs]), sc([b for _, b in pairs])
    unk = 0.0
    if tok.unk_token_id is not None:
        e = tok([g for g, _ in pairs], return_tensors="pt", padding=True, truncation=True, max_length=maxlen)
        unk = float(((e["input_ids"] == tok.unk_token_id) & e["attention_mask"].bool()).sum().item()
                    / max(1, e["attention_mask"].sum().item()))
    return float((sg > sb).mean()), unk


eng_text = "\n".join(flores_lines("eng_Latn")[: A.n_sent])
MB_ENG = mb_pairs("eng", A.max_items)
BLOCK = 128            # B-GPT's n_positions. Every model in this table is scored at the SAME
                       # context length so the nats/byte numbers are comparable.


def eval_model(m, tok, x_text, mbx):
    be, nbe = build_blocks(tok, eng_text, BLOCK); bx, nbx = build_blocks(tok, x_text, BLOCK)
    te, _ = nll_total(m, be, DEV, bs=A.bs); tx, _ = nll_total(m, bx, DEV, bs=A.bs)
    ae, ue = mb_acc(m, tok, MB_ENG, DEV)
    ax, ux = mb_acc(m, tok, mbx, DEV)
    return {"nats_per_byte_eng": te / nbe, "nats_per_byte_x": tx / nbx,
            "multiblimp_eng": ae, "multiblimp_x": ax, "unk_rate_eng": ue, "unk_rate_x": ux}

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, x2, mb = spec.split(":")
    if fcode in done: continue
    repo = f"catherinearnett/B-GPT_en_{x2}_{A.variant}"
    try:
        m, tok = load_hf(repo, dtype=torch.float32, device=DEV); m.eval()
    except Exception as e:
        log("FAILED to load", repo, type(e).__name__, str(e)[:200]); continue
    x_text = "\n".join(flores_lines(fcode)[: A.n_sent])
    MBX = mb_pairs(mb, A.max_items)
    arms = {"bgpt_joint_bilingual": eval_model(m, tok, x_text, MBX)}
    log(f"  B-GPT joint: {arms['bgpt_joint_bilingual']}")
    del m; torch.cuda.empty_cache()

    # --- the same table for the Goldfish parents and their merges, at the SAME 128-token context
    gcode = {"nld_Latn": "nld_latn", "spa_Latn": "spa_latn", "ell_Grek": "ell_grek", "pol_Latn": "pol_latn"}[fcode]
    m_e, tok_e = load_hf("goldfish-models/eng_latn_1000mb", dtype=torch.float32, device=DEV); m_e.eval()
    SD_E = sd_np(m_e)
    arms["goldfish_eng_parent"] = eval_model(m_e, tok_e, x_text, MBX)
    m_x, tok_x = load_hf(f"goldfish-models/{gcode}_1000mb", dtype=torch.float32, device=DEV); m_x.eval()
    SD_X = sd_np(m_x)
    arms["goldfish_partner_parent"] = eval_model(m_x, tok_x, x_text, MBX)
    del m_x; torch.cuda.empty_cache()
    V = int(m_e.config.vocab_size)
    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
    for nm, sd in (("merge_M0_naive", MG.average([SD_E, SD_X])),
                   ("merge_M1a_vocab", MG.average([SD_E, SD_X_V]))):
        sd_load(m_e, sd, DEV)
        arms[nm] = eval_model(m_e, tok_e, x_text, MBX)
        log(f"  {nm}: {arms[nm]}")
    del m_e; torch.cuda.empty_cache()

    r = {"set": "bgpt_ceiling", "lang": fcode, "repo": repo, "variant": A.variant,
         "context_tokens": BLOCK, "n_items_eng": len(MB_ENG), "n_items_x": len(MBX),
         "arms": arms}
    fh.write(json.dumps(r) + "\n"); fh.flush()
fh.close()
log("DONE bgpt")