| """Cross-model 10-pair diagnostics: last pre-logit layer, SGPT pooling, FLORES-200 devtest.""" |
| import itertools, json, os, sys |
| import numpy as np, pandas as pd |
| sys.path.insert(0, "/root/crossarch-diagnostics/scripts") |
| import metrics as M |
| from mergeschool.geometry import gf_core as GC |
|
|
| ROOT = "/root/crossarch-diagnostics" |
| MODELS = [ |
| ("EN_pythia", "EleutherAI/pythia-1.4b", "EN", "GPT-NeoX", 2048, 24, 16, 50304), |
| ("ZH_pythia", "SJTU-CL/Zh-Pythia-1.4B", "ZH", "GPT-NeoX", 2048, 24, 16, 49953), |
| ("PT_tucano", "TucanoBR/Tucano-1b1", "PT", "LLaMA", 2048, 22, 32, 32000), |
| ("PL_bielik", "speakleash/Bielik-1.5B-v3", "PL", "LLaMA-Qwen2.5", 1536, 32, 12, 32000), |
| ("IT_minerva", "sapienzanlp/Minerva-1B-base-v1.0", "IT", "Mistral", 2048, 16, 16, 32768), |
| ] |
|
|
|
|
| def build_uriel(): |
| """URIEL distances from lang2vec feature vectors. |
| |
| This lang2vec build has no `distance()`, so we compute it the way lang2vec defines it: |
| cosine distance between the languages' URIEL feature vectors, on `syntax_knn` (syntactic), |
| `fam` (genetic), and the concatenation syntax+phonology+inventory (featural). Missing |
| entries ("--") are dropped pairwise. |
| """ |
| import itertools as it |
| import lang2vec.lang2vec as l2v |
| iso = {"EN": "eng", "ZH": "cmn", "PT": "por", "PL": "pol", "IT": "ita"} |
| sets = {"uriel_syntactic": "syntax_knn", "uriel_genetic": "fam", |
| "uriel_featural": "syntax_knn+phonology_knn+inventory_knn"} |
| vecs = {k: l2v.get_features(list(iso.values()), v) for k, v in sets.items()} |
|
|
| def cosd(a, b): |
| a = np.asarray(a, float); b = np.asarray(b, float) |
| m = np.isfinite(a) & np.isfinite(b) |
| a, b = a[m], b[m] |
| na, nb = np.linalg.norm(a), np.linalg.norm(b) |
| if na == 0 or nb == 0: |
| return float("nan") |
| return float(1.0 - a @ b / (na * nb)) |
|
|
| out = {} |
| for a, b in it.combinations(list(iso), 2): |
| out[(a, b)] = {k: cosd(vecs[k][iso[a]], vecs[k][iso[b]]) for k in sets} |
| return out, "lang2vec/URIEL (cosine distance on syntax_knn / fam / syntax+phon+inv)" |
|
|
|
|
| def main(): |
| reps, qual = {}, {} |
| for m in MODELS: |
| z = np.load(f"{ROOT}/cache/xmodel_{m[0]}.npz") |
| reps[m[0]] = z["reps"].astype(np.float64) |
| qual[m[0]] = dict(tok_nll=float(z["nll_sum"].sum() / z["ntok"].sum()), |
| sent_nll=float(z["nll_sum"].mean())) |
| uriel, uriel_src = build_uriel() |
| meta = {m[0]: m for m in MODELS} |
| rows = [] |
| for ka, kb in itertools.combinations([m[0] for m in MODELS], 2): |
| A, B = meta[ka], meta[kb] |
| X, Y = reps[ka], reps[kb] |
| pair_seed = int.from_bytes(f"{ka}|{kb}".encode(), "little") % (2 ** 31) |
| r = M.all_rep_metrics(X, Y, pair_seed) |
| u = uriel.get((A[2], B[2])) or uriel[(B[2], A[2])] |
| rows.append(dict( |
| pair=f"{ka}|{kb}", model_a=ka, model_b=kb, lang_a=A[2], lang_b=B[2], |
| arch_a=A[3], arch_b=B[3], hid_a=A[4], hid_b=B[4], L_a=A[5], L_b=B[5], |
| heads_a=A[6], heads_b=B[6], vocab_a=A[7], vocab_b=B[7], |
| same_param_space=bool(A[3] == B[3] and A[4] == B[4] and A[5] == B[5] and A[6] == B[6]), |
| vocab_match=bool(A[7] == B[7]), hidden_match=bool(A[4] == B[4]), |
| depth_match=bool(A[5] == B[5]), **r, **u, |
| mean_tok_nll=0.5 * (qual[ka]["tok_nll"] + qual[kb]["tok_nll"]), |
| max_tok_nll=max(qual[ka]["tok_nll"], qual[kb]["tok_nll"]), |
| n_sentences=X.shape[0], pair_seed=pair_seed)) |
| df = pd.DataFrame(rows) |
| df["cka_delta"] = df.cka - df.cka_shuf |
| df["p_at_1_delta"] = df.p_at_1 - df.p_at_1_shuf |
| df.to_csv(f"{ROOT}/results/crossmodel_pairs.csv", index=False) |
| json.dump(qual, open(f"{ROOT}/results/model_quality.json", "w"), indent=2) |
| json.dump({"uriel_source": uriel_src}, open(f"{ROOT}/results/provenance_uriel.json", "w")) |
| print(df[["pair", "cka", "cka_shuf", "procrustes", "svcca", "p_at_1", "p_at_1_shuf", |
| "same_param_space", "uriel_featural", "mean_tok_nll"]].to_string(index=False)) |
| print("\nURIEL source:", uriel_src, "| estimator max|mine-gf_core| =", |
| "n/a") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
