yam-peleg/Hebrew-Mistral-7B-200K

README.md

@@ -1,5 +1,5 @@
 ---
-title: Yam-Peleg Hebrew-Mistral-7B
+title: Yam-Peleg Hebrew-Mistral-7B-200K
 emoji: 🕎
 colorFrom: blue
 colorTo: gray
@@ -11,10 +11,9 @@ license: apache-2.0
 suggested_hardware: a10g-small
 ---
 
-# Yam-Peleg's Hebrew-Mistral-7B
+# Yam-Peleg's Hebrew-Mistral-7B-200K
 
-Hebrew-Mistral-7B was introduced in [this Facebook post](https://www.facebook.com/groups/MDLI1/posts/2701023256728372/).
+Hebrew-Mistral-7B-200K was introduced in [this Facebook post](https://www.facebook.com/groups/MDLI1/posts/2708679492629415/).
 
-Please, check the [original model card](https://huggingface.co/yam-peleg/Hebrew-Mistral-7B) for more details.
+Please, check the [original model card](https://huggingface.co/yam-peleg/Hebrew-Mistral-7B-200K) for more details.
 You can see the other Hebrew models by Yam [here](https://huggingface.co/collections/yam-peleg/hebrew-models-65e957875324e2b9a4b68f08)
-

app.py

@@ -3,24 +3,27 @@ from threading import Thread
 from typing import Iterator
 
 import gradio as gr
+# from gradio import MultimodalTextbox
+# from gradio.data_classes import FileData
 import spaces
 import torch
 from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
+from typing_extensions import NotRequired, TypedDict
 
 MAX_MAX_NEW_TOKENS = 1024
 DEFAULT_MAX_NEW_TOKENS = 256
-MAX_INPUT_TOKEN_LENGTH = 512
+MAX_INPUT_TOKEN_LENGTH = 50000
 
 DESCRIPTION = """\
-# Yam-Peleg's Hebrew-Mistral-7B
+# Yam-Peleg's Hebrew-Mistral-7B-200K
 
-Hebrew-Mistral-7B was introduced in [this Facebook post](https://www.facebook.com/groups/MDLI1/posts/2701023256728372/).
+Hebrew-Mistral-7B-200K was introduced in [this Facebook post](https://www.facebook.com/groups/MDLI1/posts/2708679492629415/).
 
-Please, check the [original model card](https://huggingface.co/yam-peleg/Hebrew-Mistral-7B) for more details.
+Please, check the [original model card](https://huggingface.co/yam-peleg/Hebrew-Mistral-7B-200K) for more details.
 You can see the other Hebrew models by Yam [here](https://huggingface.co/collections/yam-peleg/hebrew-models-65e957875324e2b9a4b68f08)
 
 # Note: Use this model for only for completing sentences.
-## While the user interface is of a chatbot for convenience, this is a base model and is not fine-tuned for chatbot tasks or instruction following tasks. As such, the model is not provided a chat history and will complete your text based on the last given prompt only.
+## While the user interface is of a chatbot for convenience, this is a base model and is not fine-tuned for chatbot tasks or instruction following tasks.
 """
 
 LICENSE = """
@@ -38,9 +41,9 @@ if not torch.cuda.is_available():
 
 
 if torch.cuda.is_available():
-    model_id = "yam-peleg/Hebrew-Mistral-7B"    
+    model_id = "yam-peleg/Hebrew-Mistral-7B-200K"    
     model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", torch_dtype=torch.bfloat16, low_cpu_mem_usage=True)
-    tokenizer_id = "yam-peleg/Hebrew-Mistral-7B"
+    tokenizer_id = "yam-peleg/Hebrew-Mistral-7B-200K"
     tokenizer = AutoTokenizer.from_pretrained(tokenizer_id)
         
 
@@ -54,7 +57,14 @@ def generate(
     top_k: int = 30,
     repetition_penalty: float = 1.0,
 ) -> Iterator[str]:
-    
+    historical_text = ""
+    #Prepend the entire chat history to the message with new lines between each message
+    for user, assistant in chat_history:
+        historical_text += f"\n{user}\n{assistant}"
+
+    message = historical_text + f"\n{message}"
+    if len(historical_text) > 0:
+        message = historical_text + f"\n{message}"
     input_ids = tokenizer([message], return_tensors="pt").input_ids
     if input_ids.shape[1] > MAX_INPUT_TOKEN_LENGTH:
         input_ids = input_ids[:, -MAX_INPUT_TOKEN_LENGTH:]
@@ -102,28 +112,28 @@ chat_interface = gr.ChatInterface(
             minimum=0.1,
             maximum=4.0,
             step=0.1,
-            value=0.3,
+            value=0.9,
         ),
         gr.Slider(
             label="Top-p (nucleus sampling)",
             minimum=0.05,
             maximum=1.0,
             step=0.05,
-            value=0.3,
+            value=0.7,
         ),
         gr.Slider(
             label="Top-k",
             minimum=1,
             maximum=1000,
             step=1,
-            value=30,
+            value=40,
         ),
     ],
     stop_btn=None,
     examples=[
         ["מתכון לעוגת שוקולד:"],
         ["שפת התכנות פייטון היא"],
-        ["העלילה של סינדרלה"],
+        ["האיש האחרון בעולם ישב לבד בחדרו, כשלפתע"],
         ["שאלה: מהי עיר הבירה של מדינת ישראל?\nתשובה:"],
         ["שאלה: אני ממש עייף, מה כדאי לי לעשות?\nתשובה:"],
     ],
@@ -136,4 +146,4 @@ with gr.Blocks(css="style.css") as demo:
     gr.Markdown(LICENSE)
 
 if __name__ == "__main__":
-    demo.queue(max_size=20).launch()
+    demo.queue(max_size=20).launch()