apak commited on
Commit
eedcecc
·
verified ·
1 Parent(s): c7530d6

Update ner_logic.py

Browse files
Files changed (1) hide show
  1. ner_logic.py +19 -50
ner_logic.py CHANGED
@@ -1,42 +1,22 @@
1
  import torch
2
- import wikipedia # Hata almamak için pip install wikipedia yapılmalı
3
- import re
4
  import json
5
- import spaces
6
- from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
7
  from labels_config import CUSTOM_LABELS
8
  from prompts import SYSTEM_COT_PROMPT
 
9
 
10
- # --- MODELLER ---
11
- ner_pipe = pipeline("ner", model="xlm-roberta-large-finetuned-conll03-english", aggregation_strategy="simple", device=-1)
12
- model_id = "Qwen/Qwen2.5-1.5B-Instruct"
13
- tokenizer = AutoTokenizer.from_pretrained(model_id)
14
- llm_model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True)
15
-
16
- def get_wiki_summary(term):
17
- """Wikipedia'dan bilgi çeken kritik fonksiyon."""
18
- try:
19
- wikipedia.set_lang("tr")
20
- search_results = wikipedia.search(term)
21
- if not search_results:
22
- return "Wikipedia'da bu terim için sonuç bulunamadı."
23
- # İlk sonucun özetini al
24
- summary = wikipedia.summary(search_results[0], sentences=2, auto_suggest=False)
25
- return summary
26
- except Exception as e:
27
- return f"Wikipedia sorgu hatası: {str(e)}"
28
 
29
  @spaces.GPU
30
  def refine_labels_batch(misc_items, full_sentence):
31
- if not misc_items:
32
- return []
33
-
34
- device = "cuda" if torch.cuda.is_available() else "cpu"
35
 
36
- # Label kurallarını oluşturma
37
  label_rules = ""
38
  for k, v in CUSTOM_LABELS.items():
39
- label_rules += f"### {k}\nTANIM: {v['tanim']}\nANAHTARLAR: {', '.join(v['anahtar_kelimeler'])}\nORNEKLER: {', '.join(v.get('ornekler', []))}\n\n"
40
 
41
  targets_text = "".join([f"- VARLIK: {item['word']} | WIKI: {item['wiki']}\n" for item in misc_items])
42
 
@@ -48,37 +28,26 @@ def refine_labels_batch(misc_items, full_sentence):
48
 
49
  messages = [{"role": "user", "content": final_prompt}]
50
  text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
51
- model_inputs = tokenizer([text], return_tensors="pt").to(device)
52
 
53
  try:
54
- # LLM Üretimi
55
- generated_ids = llm_model.generate(
56
- model_inputs.input_ids,
57
- max_new_tokens=1500,
58
- do_sample=False
59
- )
60
  output = tokenizer.batch_decode(generated_ids[:, model_inputs.input_ids.shape[1]:], skip_special_tokens=True)[0]
61
 
62
- # JSON bloğunu ayıklamak için regex kullan (try bloğu içinde kalmalı)
63
- json_match = re.search(r'\[.*\]', output, re.DOTALL)
64
-
65
  if json_match:
66
- raw_json = json_match.group(0)
 
67
 
68
- # Eksik parantez kontrolü ve tamamlama
69
- if raw_json.count('{') > raw_json.count('}'):
70
- raw_json += "}]"
71
-
72
- # JSON Parse işlemi
73
  results = json.loads(raw_json)
 
 
 
 
74
  return results
75
  else:
76
- print(f"LLM Yanlış Format Döndürdü: {output}")
77
  return []
78
-
79
- except json.JSONDecodeError as e:
80
- print(f"JSON Çözümleme Hatası: {str(e)}")
81
- return []
82
  except Exception as e:
83
- print(f"Süreç Hatası: {str(e)}")
84
  return []
 
1
  import torch
 
 
2
  import json
3
+ import re
4
+ from transformers import AutoTokenizer, AutoModelForCausalLM
5
  from labels_config import CUSTOM_LABELS
6
  from prompts import SYSTEM_COT_PROMPT
7
+ import spaces
8
 
9
+ # Global modeller (Örnek olarak Qwen kullanılmıştır)
10
+ tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")
11
+ llm_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct", torch_dtype="auto", device_map="auto")
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
12
 
13
  @spaces.GPU
14
  def refine_labels_batch(misc_items, full_sentence):
15
+ if not misc_items: return []
 
 
 
16
 
 
17
  label_rules = ""
18
  for k, v in CUSTOM_LABELS.items():
19
+ label_rules += f"### {k}\nTANIM: {v['tanim']}\nANAHTARLAR: {', '.join(v['anahtar_kelimeler'])}\n\n"
20
 
21
  targets_text = "".join([f"- VARLIK: {item['word']} | WIKI: {item['wiki']}\n" for item in misc_items])
22
 
 
28
 
29
  messages = [{"role": "user", "content": final_prompt}]
30
  text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
31
+ model_inputs = tokenizer([text], return_tensors="pt").to(llm_model.device)
32
 
33
  try:
34
+ generated_ids = llm_model.generate(model_inputs.input_ids, max_new_tokens=1500, do_sample=False)
 
 
 
 
 
35
  output = tokenizer.batch_decode(generated_ids[:, model_inputs.input_ids.shape[1]:], skip_special_tokens=True)[0]
36
 
37
+ json_match = re.search(r'\[\s*\{.*\}\s*\]', output, re.DOTALL)
 
 
38
  if json_match:
39
+ raw_json = json_match.group(0).strip()
40
+ if raw_json.count('{') > raw_json.count('}'): raw_json += "}]"
41
 
 
 
 
 
 
42
  results = json.loads(raw_json)
43
+ # app.py uyumu için anahtar kontrolü
44
+ for r in results:
45
+ if 'karar' not in r: r['karar'] = "MISC"
46
+ if 'reasoning' not in r: r['reasoning'] = "Analiz adımları oluşturulamadı."
47
  return results
48
  else:
49
+ print(f"Format Hatası: {output}")
50
  return []
 
 
 
 
51
  except Exception as e:
52
+ print(f"Hata: {str(e)}")
53
  return []