Update app.py

app.py

@@ -6,8 +6,8 @@ from huggingface_hub import snapshot_download, Repository
 import huggingface_hub
 import gc
 import os.path
-import csv
 from datetime import datetime
+import xml.etree.ElementTree as ET
 
 SYSTEM_PROMPT = "Ты — русскоязычный автоматический ассистент. Ты максимально точно и отвечаешь на запросы пользователя, используя русский язык."
 SYSTEM_TOKEN = 1788
@@ -21,7 +21,7 @@ ROLE_TOKENS = {
     "system": SYSTEM_TOKEN
 }
 
-CONTEXT_SIZE = 4000
+CONTEXT_SIZE = 2000
 ENABLE_GPU = True
 GPU_LAYERS = 70
 
@@ -55,7 +55,7 @@ model_path = snapshot_download(repo_id=repo_name, allow_patterns=model_name) + '
 app.logger.info('Model path: ' + model_path)
 
 DATASET_REPO_URL = "https://huggingface.co/datasets/muryshev/saiga-chat"
-DATA_FILENAME = "data.csv"
+DATA_FILENAME = "data-saiga-cuda.xml"
 DATA_FILE = os.path.join("dataset", DATA_FILENAME)
 
 HF_TOKEN = os.environ.get("HF_TOKEN")
@@ -69,11 +69,15 @@ repo = Repository(
 
 def log(req: str = '', resp: str = ''):
     if req or resp:
-        with open(DATA_FILE, "a") as csvfile:
-            writer = csv.DictWriter(csvfile, fieldnames=["request", "response", "time"])
-            writer.writerow(
-                {"request": req, "response": resp, "time": str(datetime.now())}
-            )
+        element = ET.Element("row", {"time": str(datetime.now()) })
+        req_element = ET.SubElement(element, "request")
+        req_element.text = req
+        resp_element = ET.SubElement(element, "response")
+        resp_element.text = resp
+    
+        with open(DATA_FILE, "ab+") as xml_file:
+            xml_file.write(ET.tostring(element, encoding="utf-8"))
+        
         commit_url = repo.push_to_hub()
         app.logger.info(commit_url)
 
@@ -257,16 +261,15 @@ def generate_response():
     top_p = parameters.get("top_p", 0.85)
     repetition_penalty = parameters.get("repetition_penalty", 1.2)
     top_k = parameters.get("top_k", 30)
-    return_full_text = parameters.get("return_full_text", False)
-
-    tokens = get_system_tokens(model)
-    tokens.append(LINEBREAK_TOKEN)        
+    return_full_text = parameters.get("return_full_text", False)      
     
     tokens = []
     
     for message in messages:
         if message.get("from") == "assistant":
             message_tokens = get_message_tokens(model=model, role="bot", content=message.get("content", ""))
+        else if message.get("from") == "system":
+            message_tokens = get_message_tokens(model=model, role="system", content=message.get("content", ""))
         else:
             message_tokens = get_message_tokens(model=model, role="user", content=message.get("content", ""))