import torch
import sys
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import json
tokenizer = AutoTokenizer.from_pretrained('google/gemma-2-2b-it')
# Configure 4-bit quantization using BitsAndBytesConfig
quantization_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_compute_dtype=torch.bfloat16,
bnb_4bit_quant_type="nf4",
)
# Load the model with the quantization configuration
model = AutoModelForCausalLM.from_pretrained(
'google/gemma-2-2b-it',
device_map="auto",
quantization_config=quantization_config,
)
# Definir el prompt para generar un JSON con eventos anidados
prompt = (
"Genera un JSON que describa una serie de eventos consecutivos en un formato similar al siguiente:\n\n"
"{\n"
" \"events\": {\n"
" \"event\": {\n"
" \"event_number\": 1,\n"
" \"name\": \"conflict_start\",\n"
" \"description\": \"Tensions escalate between Iran and Israel\",\n"
" \"probability\": 70,\n"
" \"duration_days\": 30,\n"
" \"subevents\": {\n"
" \"event\": {\n"
" \"event_number\": 2,\n"
" \"name\": \"diplomatic_failure\",\n"
" \"description\": \"Diplomatic negotiations fail\",\n"
" \"probability\": 60,\n"
" \"duration_days\": 15,\n"
" \"subevents\": {\n"
" \"event\": {\n"
" \"event_number\": 3,\n"
" \"name\": \"military_clash\",\n"
" \"description\": \"Initial military clash at the border\",\n"
" \"probability\": 50,\n"
" \"duration_days\": 10,\n"
" \"subevents\": {\n"
" \"event\": [\n"
" {\n"
" \"event_number\": 4,\n"
" \"name\": \"escalation\",\n"
" \"description\": \"Conflict escalates into full-scale war\",\n"
" \"probability\": 40,\n"
" \"duration_days\": 180,\n"
" \"subevents\": {\n"
" \"event\": [\n"
" {\n"
" \"event_number\": 5,\n"
" \"name\": \"regional_involvement\",\n"
" \"description\": \"Other Middle Eastern countries get involved\",\n"
" \"probability\": 30,\n"
" \"duration_days\": 365,\n"
" \"subevents\": {\n"
" \"event\": [\n"
" {\n"
" \"event_number\": 6,\n"
" \"name\": \"ceasefire\",\n"
" \"description\": \"International powers broker a ceasefire\",\n"
" \"probability\": 20,\n"
" \"duration_days\": 30\n"
" },\n"
" {\n"
" \"event_number\": 7,\n"
" \"name\": \"prolonged_conflict\",\n"
" \"description\": \"Conflict continues for over a year\",\n"
" \"probability\": 50,\n"
" \"duration_days\": 365\n"
" }\n"
" ]\n"
" }\n"
" },\n"
" {\n"
" \"event_number\": 8,\n"
" \"name\": \"international_intervention\",\n"
" \"description\": \"UN or other international organizations intervene\",\n"
" \"probability\": 25,\n"
" \"duration_days\": 60\n"
" }\n"
" ]\n"
" }\n"
" },\n"
" {\n"
" \"event_number\": 9,\n"
" \"name\": \"containment\",\n"
" \"description\": \"Conflict is contained and doesn't escalate\",\n"
" \"probability\": 30,\n"
" \"duration_days\": 90\n"
" }\n"
" ]\n"
" }\n"
" },\n"
" \"event\": {\n"
" \"event_number\": 10,\n"
" \"name\": \"sanctions\",\n"
" \"description\": \"Increased sanctions on Iran\",\n"
" \"probability\": 70,\n"
" \"duration_days\": 180,\n"
" \"subevents\": {\n"
" \"event\": [\n"
" {\n"
" \"event_number\": 11,\n"
" \"name\": \"iran_retaliates\",\n"
" \"description\": \"Iran retaliates with cyberattacks\",\n"
" \"probability\": 40,\n"
" \"duration_days\": 60\n"
" },\n"
" {\n"
" \"event_number\": 12,\n"
" \"name\": \"israel_response\",\n"
" \"description\": \"Israel responds with targeted airstrikes\",\n"
" \"probability\": 50,\n"
" \"duration_days\": 60\n"
" }\n"
" ]\n"
" }\n"
" }\n"
" }\n"
" },\n"
" \"event\": {\n"
" \"event_number\": 13,\n"
" \"name\": \"diplomatic_success\",\n"
" \"description\": \"Successful diplomatic negotiations\",\n"
" \"probability\": 40,\n"
" \"duration_days\": 30,\n"
" \"subevents\": {\n"
" \"event\": [\n"
" {\n"
" \"event_number\": 14,\n"
" \"name\": \"peace_agreement\",\n"
" \"description\": \"Iran and Israel sign a peace agreement\",\n"
" \"probability\": 20,\n"
" \"duration_days\": 60\n"
" },\n"
" {\n"
" \"event_number\": 15,\n"
" \"name\": \"temporary_truce\",\n"
" \"description\": \"A temporary truce is established\",\n"
" \"probability\": 30,\n"
" \"duration_days\": 30\n"
" }\n"
" ]\n"
" }\n"
" }\n"
" }\n"
" }\n"
" }\n"
"}\n\n"
"Ahora, genera un JSON similar con eventos anidados, pero cambia los detalles y nĂºmeros para hacer que sea con el input que viene a continuacion, respondiendo solo el JSON empezando con y terminando con :"
)
def generate(event):
combined_input = f"{prompt} {event}"
prompt_msg = [{'role': 'user', 'content': combined_input}]
inputs = tokenizer.apply_chat_template(
prompt_msg,
add_generation_prompt=True,
return_tensors='pt'
)
tokens = model.generate(
inputs.to(model.device),
max_new_tokens=1024,
temperature=0.5,
do_sample=True
)
# Get the length of the input tokens (adjust based on your tokenizer)
input_length = len(tokenizer.encode(combined_input))
output_text = tokenizer.decode(tokens[0][input_length:], skip_special_tokens=True)
print(output_text)
json_start_index = output_text.find("")
json_end_index = output_text.find("")
if json_start_index != -1 and json_end_index != -1:
json_string = output_text[json_start_index + 6:json_end_index].strip()
# Debugging: Print the extracted JSON string to check its contents
print("Extracted JSON String:", json_string)
# Load and return the JSON data
try:
data = json.loads(json_string)
return data
except json.JSONDecodeError as e:
return f"Error: Invalid JSON - {e}"
else:
return "Error: or not found in generated output"