| import json,glob,os,sys |
| sys.path.insert(0,"..") |
| from tokenizers import Tokenizer |
| from items import ITEMS |
| en={(d["link"],d["question_number"]):d for d in map(json.loads,open("eng_Latn.jsonl"))} |
| ta={(d["link"],d["question_number"]):d for d in map(json.loads,open("tam_Taml.jsonl"))} |
| common=[k for k in ta if k in en]; print("aligned q",len(common)) |
| pe={};pt={} |
| for k in common: pe[k[0]]=en[k]["flores_passage"]; pt[k[0]]=ta[k]["flores_passage"] |
| keys=list(pt); print("passages",len(keys)) |
| E=[pe[k] for k in keys]; T=[pt[k] for k in keys] |
| tw=sum(len(t.split()) for t in T); ew=sum(len(t.split()) for t in E) |
| tl=[i["tl"] for i in ITEMS]; en50=[i["en"] for i in ITEMS]; ta50=[i["ta"] for i in ITEMS] |
| rows=[] |
| for f in sorted(glob.glob("toks/*.json")): |
| k=os.path.basename(f)[:-5]; t=Tokenizer.from_file(f) |
| n=lambda L: sum(len(t.encode(x,add_special_tokens=False).ids) for x in L) |
| ne,nt=n(E),n(T); n50e,n50t,n50l=n(en50),n(ta50),n(tl) |
| rows.append(dict(model=k,vocab=t.get_vocab_size(),flores_en=ne,flores_ta=nt,ratio=round(nt/ne,2),tok_per_ta_word=round(nt/tw,2),tok_per_en_word=round(ne/ew,2),chars_per_ta_tok=round(sum(map(len,T))/nt,2),tb_en=n50e,tb_ta=n50t,tb_tl=n50l,tl_ratio=round(n50l/n50e,2))) |
| rows.sort(key=lambda r:r["ratio"]) |
| for r in rows: print(r) |
| json.dump(dict(passages=len(keys),ta_words=tw,en_words=ew,rows=rows),open("tokenizer_league.json","w"),indent=1) |
|
|
