QAway-to
commited on
Commit
·
1c31761
1
Parent(s):
87cd86c
Change tokenizer v1.4
Browse files- core/interviewer.py +15 -10
core/interviewer.py
CHANGED
|
@@ -1,25 +1,33 @@
|
|
| 1 |
# core/interviewer.py
|
| 2 |
"""
|
| 3 |
🇬🇧 Interviewer logic module (no instructions)
|
| 4 |
-
Generates random MBTI-style questions using
|
| 5 |
|
| 6 |
🇷🇺 Модуль интервьюера.
|
| 7 |
-
Использует fine-tuned модель для генерации вопросов без
|
| 8 |
"""
|
|
|
|
|
|
|
|
|
|
|
|
|
| 9 |
from transformers import AutoModelForSeq2SeqLM, T5Tokenizer
|
| 10 |
|
|
|
|
|
|
|
|
|
|
| 11 |
QG_MODEL = "f3nsmart/ft-flan-t5-base-qgen"
|
| 12 |
|
|
|
|
| 13 |
tokenizer = T5Tokenizer.from_pretrained(QG_MODEL, use_fast=False)
|
| 14 |
model = AutoModelForSeq2SeqLM.from_pretrained(QG_MODEL)
|
| 15 |
|
| 16 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 17 |
model.to(device).eval()
|
| 18 |
-
|
| 19 |
print(f"✅ Loaded interviewer model (slow tokenizer): {QG_MODEL}")
|
|
|
|
| 20 |
|
| 21 |
# --------------------------------------------------------------
|
| 22 |
-
# 2️⃣
|
| 23 |
# --------------------------------------------------------------
|
| 24 |
PROMPTS = [
|
| 25 |
"Personality and emotions.",
|
|
@@ -33,7 +41,7 @@ PROMPTS = [
|
|
| 33 |
# 3️⃣ Очистка текста
|
| 34 |
# --------------------------------------------------------------
|
| 35 |
def _clean_question(text: str) -> str:
|
| 36 |
-
"""Берёт первую фразу с '?'
|
| 37 |
text = text.strip()
|
| 38 |
m = re.search(r"(.+?\?)", text)
|
| 39 |
if m:
|
|
@@ -50,9 +58,7 @@ def _clean_question(text: str) -> str:
|
|
| 50 |
# 4️⃣ Генерация вопроса
|
| 51 |
# --------------------------------------------------------------
|
| 52 |
def generate_question(user_id: str = "default_user", **kwargs) -> str:
|
| 53 |
-
"""
|
| 54 |
-
Генерирует один MBTI-вопрос без инструкций.
|
| 55 |
-
"""
|
| 56 |
prompt = random.choice(PROMPTS)
|
| 57 |
inputs = tokenizer(prompt, return_tensors="pt", truncation=True).to(device)
|
| 58 |
with torch.no_grad():
|
|
@@ -65,5 +71,4 @@ def generate_question(user_id: str = "default_user", **kwargs) -> str:
|
|
| 65 |
max_new_tokens=60,
|
| 66 |
)
|
| 67 |
text = tokenizer.decode(out[0], skip_special_tokens=True)
|
| 68 |
-
|
| 69 |
-
return question
|
|
|
|
| 1 |
# core/interviewer.py
|
| 2 |
"""
|
| 3 |
🇬🇧 Interviewer logic module (no instructions)
|
| 4 |
+
Generates random MBTI-style questions using the fine-tuned model.
|
| 5 |
|
| 6 |
🇷🇺 Модуль интервьюера.
|
| 7 |
+
Использует fine-tuned модель для генерации вопросов без инструкций.
|
| 8 |
"""
|
| 9 |
+
|
| 10 |
+
import random
|
| 11 |
+
import re
|
| 12 |
+
import torch
|
| 13 |
from transformers import AutoModelForSeq2SeqLM, T5Tokenizer
|
| 14 |
|
| 15 |
+
# --------------------------------------------------------------
|
| 16 |
+
# 1️⃣ Настройки модели
|
| 17 |
+
# --------------------------------------------------------------
|
| 18 |
QG_MODEL = "f3nsmart/ft-flan-t5-base-qgen"
|
| 19 |
|
| 20 |
+
# ✅ Принудительно используем оригинальный SentencePiece-токенайзер
|
| 21 |
tokenizer = T5Tokenizer.from_pretrained(QG_MODEL, use_fast=False)
|
| 22 |
model = AutoModelForSeq2SeqLM.from_pretrained(QG_MODEL)
|
| 23 |
|
| 24 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 25 |
model.to(device).eval()
|
|
|
|
| 26 |
print(f"✅ Loaded interviewer model (slow tokenizer): {QG_MODEL}")
|
| 27 |
+
print(f"Device set to use {device}")
|
| 28 |
|
| 29 |
# --------------------------------------------------------------
|
| 30 |
+
# 2️⃣ Seed-промпты (без инструкций)
|
| 31 |
# --------------------------------------------------------------
|
| 32 |
PROMPTS = [
|
| 33 |
"Personality and emotions.",
|
|
|
|
| 41 |
# 3️⃣ Очистка текста
|
| 42 |
# --------------------------------------------------------------
|
| 43 |
def _clean_question(text: str) -> str:
|
| 44 |
+
"""Берёт первую фразу с '?'"""
|
| 45 |
text = text.strip()
|
| 46 |
m = re.search(r"(.+?\?)", text)
|
| 47 |
if m:
|
|
|
|
| 58 |
# 4️⃣ Генерация вопроса
|
| 59 |
# --------------------------------------------------------------
|
| 60 |
def generate_question(user_id: str = "default_user", **kwargs) -> str:
|
| 61 |
+
"""Генерирует один MBTI-вопрос без инструкций"""
|
|
|
|
|
|
|
| 62 |
prompt = random.choice(PROMPTS)
|
| 63 |
inputs = tokenizer(prompt, return_tensors="pt", truncation=True).to(device)
|
| 64 |
with torch.no_grad():
|
|
|
|
| 71 |
max_new_tokens=60,
|
| 72 |
)
|
| 73 |
text = tokenizer.decode(out[0], skip_special_tokens=True)
|
| 74 |
+
return _clean_question(text)
|
|
|