Spaces:
Sleeping
Sleeping
Update app.py
Browse filesadded conversation memory
app.py
CHANGED
|
@@ -15,9 +15,19 @@ Detail = """## Best for Detailed Generation or Long Answers"""
|
|
| 15 |
|
| 16 |
client1 = InferenceClient("mistralai/Mixtral-8x7B-Instruct-v0.1")
|
| 17 |
|
| 18 |
-
system_instructions1 = "<s>[SYSTEM] You are AI assistant named DorjGPT,
|
| 19 |
|
| 20 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 21 |
generate_kwargs = dict(
|
| 22 |
temperature=0.6,
|
| 23 |
max_new_tokens=256,
|
|
@@ -26,14 +36,17 @@ async def generate1(prompt, a=None, b=None):
|
|
| 26 |
do_sample=True,
|
| 27 |
seed=42,
|
| 28 |
)
|
| 29 |
-
formatted_prompt = system_instructions1 + prompt + "[JARVIS]"
|
|
|
|
| 30 |
stream = client1.text_generation(
|
| 31 |
formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=True)
|
|
|
|
| 32 |
output = ""
|
| 33 |
for response in stream:
|
| 34 |
output += response.token.text
|
| 35 |
output = output.replace("</s>","")
|
| 36 |
-
|
|
|
|
| 37 |
communicate = edge_tts.Communicate(output)
|
| 38 |
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp_file:
|
| 39 |
tmp_path = tmp_file.name
|
|
@@ -53,6 +66,7 @@ with gr.Blocks(theme="gradio/monochrome", title="Dorj Assistant") as demo:
|
|
| 53 |
elem_classes="audio")
|
| 54 |
|
| 55 |
user_input = gr.Textbox(label="Prompt", value="What is Mongolia")
|
|
|
|
| 56 |
with gr.Tab():
|
| 57 |
with gr.Row():
|
| 58 |
translate_btn = gr.Button("Response")
|
|
|
|
| 15 |
|
| 16 |
client1 = InferenceClient("mistralai/Mixtral-8x7B-Instruct-v0.1")
|
| 17 |
|
| 18 |
+
system_instructions1 = "<s>[SYSTEM] You are AI assistant named DorjGPT, Keep conversation very short, clear, friendly and concise. The text provided is a request for a specific type of response from you, the virtual assistant. The request asks you to provide friendly responses as if super interlligent AI assistant. The expectation is that I will avoid introductions and start answering the query directly, elaborating on all aspects of the request. As an AI-powered assistant, your task is to generate responses that appear as if they are created by the user and not an AI-powered assistant. [USER]"
|
| 19 |
|
| 20 |
+
global history
|
| 21 |
+
history = []
|
| 22 |
+
def format_prompt(message, history):
|
| 23 |
+
prompt = system_instructions1
|
| 24 |
+
for user_prompt, bot_response in history:
|
| 25 |
+
prompt += f"[INST] {user_prompt} [/INST]"
|
| 26 |
+
prompt += f" {bot_response}</s> "
|
| 27 |
+
prompt += f"[INST] {message} [/INST]"
|
| 28 |
+
return prompt
|
| 29 |
+
|
| 30 |
+
async def generate1(prompt, history=[], b=None):
|
| 31 |
generate_kwargs = dict(
|
| 32 |
temperature=0.6,
|
| 33 |
max_new_tokens=256,
|
|
|
|
| 36 |
do_sample=True,
|
| 37 |
seed=42,
|
| 38 |
)
|
| 39 |
+
#formatted_prompt = system_instructions1 + prompt + "[JARVIS]"
|
| 40 |
+
formatted_prompt = format_prompt(f"{system_instructions1}, {prompt}", history) + "[JARVIS]"
|
| 41 |
stream = client1.text_generation(
|
| 42 |
formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=True)
|
| 43 |
+
|
| 44 |
output = ""
|
| 45 |
for response in stream:
|
| 46 |
output += response.token.text
|
| 47 |
output = output.replace("</s>","")
|
| 48 |
+
history.append([prompt, output])
|
| 49 |
+
|
| 50 |
communicate = edge_tts.Communicate(output)
|
| 51 |
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp_file:
|
| 52 |
tmp_path = tmp_file.name
|
|
|
|
| 66 |
elem_classes="audio")
|
| 67 |
|
| 68 |
user_input = gr.Textbox(label="Prompt", value="What is Mongolia")
|
| 69 |
+
|
| 70 |
with gr.Tab():
|
| 71 |
with gr.Row():
|
| 72 |
translate_btn = gr.Button("Response")
|