Hugging Face's logo

shahidul034
/
readctrl

Safetensors
Model card Files
xet
Community
readctrl
File size: 4,974 Bytes
1db7196
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
 
import os
# Environment Setup
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = "2"
import argparse
parser=argparse.ArgumentParser()
parser.add_argument("--lang",type=str,default="pt",help="language code")
args=parser.parse_args()
lang_code=args.lang

import json
import tqdm
import numpy as np
import pandas as pd
import textstat
import spacy
import torch
from sentence_transformers import SentenceTransformer, util
from datasets import load_dataset

# lang_code="pt"
# device = "cuda" if torch.cuda.is_available() else "cpu"

# 1. Load Models Efficiently
model = SentenceTransformer('all-MiniLM-L6-v2')
# model = SentenceTransformer('Qwen/Qwen3-Embedding-0.6B')
# Disable unnecessary components in Spacy to save time/memory
nlp = spacy.load(f"{lang_code}_core_news_sm", disable=["ner", "lemmatizer", "attribute_ruler"])

def get_parse_tree_stats(text):
    doc = nlp(text)
    depths = []
    for sent in doc.sents:
        def walk_tree(node, depth):
            if not list(node.children): return depth
            return max(walk_tree(child, depth + 1) for child in node.children)
        depths.append(walk_tree(sent.root, 1))
    return np.mean(depths) if depths else 0

# 2. Data Loading
ds = load_dataset("wikimedia/wikipedia", f"20231101.{lang_code}", split='train', streaming=True)
# Taking a subset for the anchor pool to keep memory manageable
wiki_list = [item['text'] for item in ds.take(1000000)] 

# 3. PRE-PROCESS WIKI ANCHORS (Do this ONCE)
print("Chunking and Encoding Wikipedia...")
wiki_chunks = []
for text in wiki_list:
    paragraphs = [p.strip() for p in text.split('\n\n') if len(p.split()) > 20]
    wiki_chunks.extend(paragraphs)

# Encode all chunks at once and keep on GPU
chunk_embs = model.encode(wiki_chunks, convert_to_tensor=True, show_progress_bar=True)

# 4. Load Target Docs
with open(f"/home/mshahidul/readctrl/data/synthetic_dataset_diff_labels/syn_data_diff_labels_{lang_code}_v1.json", "r") as f:
    res = json.load(f)

my_target_documents = []
for item in res:
    for key, value in item['diff_label_texts'].items():
        my_target_documents.append({"index": item['index'], "label": key, "text": value})

# Load Progress
save_path = f"/home/mshahidul/readctrl/data/data_annotator_data/crowdsourcing_input_{lang_code}_v1.json"
processed_data = []
if os.path.exists(save_path):
    with open(save_path, "r") as f:
        processed_data = json.load(f)
processed_keys = {(d['index'], d['label']) for d in processed_data}

# 5. Process with Batching logic where possible
print("Starting Matching Loop...")
for doc in tqdm.tqdm(my_target_documents):
    if (doc['index'], doc['label']) in processed_keys:
        continue

    # A. Robust Anchor Finding (Optimized)
    doc_emb = model.encode(doc['text'], convert_to_tensor=True)
    doc_len = len(doc['text'].split())
    
    hits = util.semantic_search(doc_emb, chunk_embs, top_k=25)[0]
    
    wiki_anchor = None
    best_fallback = None
    min_delta = float('inf')

    for hit in hits:
        cand_text = wiki_chunks[hit['corpus_id']]
        cand_len = len(cand_text.split())
        len_diff = abs(cand_len - doc_len)
        
        # Track fallback while looking for strict match
        if len_diff < min_delta:
            min_delta = len_diff
            best_fallback = cand_text
            
        if 0.8 <= (cand_len / doc_len) <= 1.2:
            wiki_anchor = cand_text
            break
    
    if not wiki_anchor:
        wiki_anchor = best_fallback

    # B. Calculate Metrics
    doc_metrics = {
        "fkgl": textstat.flesch_kincaid_grade(doc['text']),
        "word_count": doc_len
    }
    wiki_metrics = {
        "fkgl": textstat.flesch_kincaid_grade(wiki_anchor),
        "word_count": len(wiki_anchor.split())
    }
    
    # C. Store results
    processed_data.append({
        "index": doc['index'],
        "label": doc['label'],
        "original_doc": doc['text'],
        "wiki_anchor": wiki_anchor,
        "doc_fkgl": doc_metrics['fkgl'],
        "wiki_fkgl": wiki_metrics['fkgl'],
        "doc_tree_depth": get_parse_tree_stats(doc['text']),
        "wiki_tree_depth": get_parse_tree_stats(wiki_anchor),
        "fkgl_delta": doc_metrics['fkgl'] - wiki_metrics['fkgl']
    })

    # Save every 20 to reduce disk I/O overhead
    if len(processed_data) % 20 == 0:
        with open(save_path, "w") as f:
            json.dump(processed_data, f, indent=2)

# Final Save
with open(save_path, "w") as f:
    json.dump(processed_data, f, indent=2)


# python /home/mshahidul/readctrl/code/vectordb_build/data_annotate_data_prep_test_v2.py --lang pt
python /home/mshahidul/readctrl/code/vectordb_build/data_annotate_data_prep_test_v2.py --lang en
# python /home/mshahidul/readctrl/code/vectordb_build/data_annotate_data_prep_test_v2.py --lang es
# python /home/mshahidul/readctrl/code/vectordb_build/data_annotate_data_prep_test_v2.py --lang fr
python /home/mshahidul/readctrl/code/readability_control.py