File size: 1,919 Bytes
7debf49
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
"""Robustness pass (no inference): for every atlas/math/rewrite row add
  lid_noquote        GlotLID profile after dropping segments that quote the prompt (passage/options/problem)
  lingua_eng         English share by an independent detector (Lingua)
  lingua_eng_noquote the same, quote-stripped
Rows are rewritten in place. Usage: python relid.py"""
import json, os
from concurrent.futures import ThreadPoolExecutor
from datasets import load_dataset
from lid import profile, lingua_eng_share
from atlas import LANGS as ATLAS_LANGS, mc_prompt

tps = json.load(open("data/atlas/templates.json", encoding="utf-8"))
par = json.load(open("data/math/parallel.json", encoding="utf-8"))
bel = {}


def belebele(lang):
    if lang not in bel:
        bel[lang] = {(r["link"], r["question_number"]): r for r in load_dataset("facebook/belebele", lang, split="test")}
    return bel[lang]


def prompt_of(r):
    if "link" in r:
        return mc_prompt(belebele(r["lang"])[(r["link"], r["qnum"])], tps[r["lang"]])
    return par[str(r["id"])][r["lang"]]["text"]


def enrich(r):
    p = prompt_of(r)
    r["lid_noquote"] = profile(r["reasoning"], r["lang"], exclude=p)
    r["lingua_eng"] = lingua_eng_share(r["reasoning"])
    r["lingua_eng_noquote"] = lingua_eng_share(r["reasoning"], exclude=p)
    return r


for path in ("data/atlas/runs_low.jsonl", "data/atlas/runs_high.jsonl", "data/math/runs_low.jsonl",
             "data/math/runs_high.jsonl"):
    if not os.path.exists(path):
        continue
    rows = [json.loads(l) for l in open(path, encoding="utf-8")]
    for lang in {r["lang"] for r in rows if "link" in r}:
        belebele(lang)  # load once, single-threaded
    with ThreadPoolExecutor(4) as ex:
        rows = list(ex.map(enrich, rows))
    with open(path, "w", encoding="utf-8") as f:
        f.writelines(json.dumps(r, ensure_ascii=False) + "\n" for r in rows)
    print("enriched", path, len(rows))