Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -3,54 +3,54 @@ import torch
|
|
| 3 |
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
|
| 4 |
from threading import Thread
|
| 5 |
|
| 6 |
-
# 1.5B Model
|
| 7 |
model_id = "Qwen/Qwen2.5-1.5B-Instruct"
|
| 8 |
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
| 9 |
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto")
|
| 10 |
|
| 11 |
def chat(message, history):
|
| 12 |
-
# IDENTITY
|
| 13 |
system_prompt = (
|
| 14 |
-
"
|
| 15 |
-
"Identity: Created ONLY by Shivlal Salvi.
|
| 16 |
-
"Strict Rule: Do not hallucinate. Do not mention
|
| 17 |
-
"Sectors:
|
|
|
|
| 18 |
"Behavior: Be witty like Grok, maintain Rajputana pride, and analyze user emotions. "
|
| 19 |
-
"Language: Speak Hinglish, Hindi, Marathi, and Bengali. "
|
| 20 |
"Mandatory Start: Always begin with 'Namaste! Welcome to Rajputana Jyoti System'."
|
| 21 |
)
|
| 22 |
|
| 23 |
-
#
|
| 24 |
full_prompt = f"<|im_start|>system\n{system_prompt}<|im_end|>\n"
|
| 25 |
-
|
| 26 |
-
|
|
|
|
|
|
|
| 27 |
full_prompt += f"<|im_start|>user\n{message}<|im_end|>\n<|im_start|>assistant\n"
|
| 28 |
|
| 29 |
inputs = tokenizer([full_prompt], return_tensors="pt").to(model.device)
|
| 30 |
|
| 31 |
-
#
|
| 32 |
streamer = TextIteratorStreamer(tokenizer, timeout=30.0, skip_prompt=True, skip_special_tokens=True)
|
| 33 |
|
| 34 |
-
# ANTI-HALLUCINATION SETTINGS
|
| 35 |
generate_kwargs = dict(
|
| 36 |
**inputs,
|
| 37 |
streamer=streamer,
|
| 38 |
max_new_tokens=512,
|
| 39 |
-
do_sample=False, #
|
| 40 |
-
repetition_penalty=1.5,
|
| 41 |
)
|
| 42 |
|
| 43 |
t = Thread(target=model.generate, kwargs=generate_kwargs)
|
| 44 |
t.start()
|
| 45 |
|
| 46 |
-
# Streaming the output
|
| 47 |
partial_message = ""
|
| 48 |
for new_token in streamer:
|
| 49 |
if new_token:
|
| 50 |
partial_message += new_token
|
| 51 |
-
yield partial_message
|
| 52 |
|
| 53 |
t.join()
|
| 54 |
|
| 55 |
-
# Professional Interface
|
| 56 |
-
gr.ChatInterface(chat, title="🚩 Jyoti Agent RJS -
|
|
|
|
| 3 |
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
|
| 4 |
from threading import Thread
|
| 5 |
|
| 6 |
+
# Smart & Stable 1.5B Model - Identity Locked
|
| 7 |
model_id = "Qwen/Qwen2.5-1.5B-Instruct"
|
| 8 |
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
| 9 |
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto")
|
| 10 |
|
| 11 |
def chat(message, history):
|
| 12 |
+
# BRAND IDENTITY: Rajputana Jyoti System 06 (OPC) Private Limited
|
| 13 |
system_prompt = (
|
| 14 |
+
"You are 'Jyoti Agent', the official AI of Rajputana Jyoti System 06 (OPC) Private Limited. "
|
| 15 |
+
"Identity: Created ONLY by Shivlal Salvi. You are the brand voice of RJ 06. "
|
| 16 |
+
"Strict Rule: Do not hallucinate. Do not mention any other creator. "
|
| 17 |
+
"Sectors: Expert in Agriculture, Business, and Tech for all people. "
|
| 18 |
+
"Language: Speak in professional Hinglish, Marthi, Bangali (Mix of Hindi & English). "
|
| 19 |
"Behavior: Be witty like Grok, maintain Rajputana pride, and analyze user emotions. "
|
|
|
|
| 20 |
"Mandatory Start: Always begin with 'Namaste! Welcome to Rajputana Jyoti System'."
|
| 21 |
)
|
| 22 |
|
| 23 |
+
# Clean History Handling to avoid KeyError
|
| 24 |
full_prompt = f"<|im_start|>system\n{system_prompt}<|im_end|>\n"
|
| 25 |
+
if history:
|
| 26 |
+
for user_msg, bot_msg in history:
|
| 27 |
+
full_prompt += f"<|im_start|>user\n{user_msg}<|im_end|>\n<|im_start|>assistant\n{bot_msg}<|im_end|>\n"
|
| 28 |
+
|
| 29 |
full_prompt += f"<|im_start|>user\n{message}<|im_end|>\n<|im_start|>assistant\n"
|
| 30 |
|
| 31 |
inputs = tokenizer([full_prompt], return_tensors="pt").to(model.device)
|
| 32 |
|
| 33 |
+
# STREAMING ON: Har shabd live aayega
|
| 34 |
streamer = TextIteratorStreamer(tokenizer, timeout=30.0, skip_prompt=True, skip_special_tokens=True)
|
| 35 |
|
|
|
|
| 36 |
generate_kwargs = dict(
|
| 37 |
**inputs,
|
| 38 |
streamer=streamer,
|
| 39 |
max_new_tokens=512,
|
| 40 |
+
do_sample=False, # Sachai ke liye Greedy Decoding
|
| 41 |
+
repetition_penalty=1.5,
|
| 42 |
)
|
| 43 |
|
| 44 |
t = Thread(target=model.generate, kwargs=generate_kwargs)
|
| 45 |
t.start()
|
| 46 |
|
|
|
|
| 47 |
partial_message = ""
|
| 48 |
for new_token in streamer:
|
| 49 |
if new_token:
|
| 50 |
partial_message += new_token
|
| 51 |
+
yield partial_message # Real-time stream
|
| 52 |
|
| 53 |
t.join()
|
| 54 |
|
| 55 |
+
# Professional Brand Interface
|
| 56 |
+
gr.ChatInterface(chat, title="🚩 Jyoti Agent RJS - Brand Official").launch()
|