new model

app.py

@@ -10,20 +10,20 @@ import torch
 from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer, pipeline
 from peft import PeftModel
 
-DESCRIPTION = "# 真空ジェネレータ\n<p>Imitate 真空 (@vericava)'s posts interactively</p>"
+DESCRIPTION = "# 真空ジェネレータ (v3)\n<p>Imitate 真空 (@vericava)'s posts interactively</p>"
 
 if not torch.cuda.is_available():
     DESCRIPTION += "\n<p>Running on CPU 🥶 This demo does not work on CPU.</p>"
 
-MAX_MAX_NEW_TOKENS = 768
-DEFAULT_MAX_NEW_TOKENS = 512
+MAX_MAX_NEW_TOKENS = 128
+DEFAULT_MAX_NEW_TOKENS = 64
 MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "32768"))
 
 
 if torch.cuda.is_available():
     my_pipeline=pipeline(
         task="text-generation",
-        model="vericava/gpt2-medium-vericava-posts-v3",
+        model="vericava/llm-jp-3-vericava-posts-v1",
         do_sample=True,
         num_beams=1,
     )
@@ -59,8 +59,8 @@ def generate(
     )[-1]["generated_text"]
     print(output)
     gen_text = output[len(user_input):]
-    gen_text = gen_text[:gen_text.find("\n")] if "\n" in gen_text else gen_text
-    gen_text = gen_text[:(gen_text.rfind("。") + 1)] if "。" in gen_text else gen_text
+    #gen_text = gen_text[:gen_text.find("\n")] if "\n" in gen_text else gen_text
+    #gen_text = gen_text[:(gen_text.rfind("。") + 1)] if "。" in gen_text else gen_text
     yield gen_text
 
 demo = gr.ChatInterface(
@@ -87,29 +87,29 @@ demo = gr.ChatInterface(
             minimum=0.05,
             maximum=1.0,
             step=0.05,
-            value=0.95,
+            value=0.90,
         ),
         gr.Slider(
             label="Top-k",
             minimum=1,
-            maximum=1000,
+            maximum=100,
             step=1,
-            value=50,
+            value=20,
         ),
         gr.Slider(
             label="Repetition penalty",
             minimum=1.0,
-            maximum=2.0,
+            maximum=4.0,
             step=0.05,
-            value=1.5,
+            value=2.0,
         ),
     ],
     stop_btn=None,
     examples=[
-        ["サマリーを作る男の人，サマリーマン。"],
-        ["やばい場所にクリティカルな配線ができてしまったので掲示した。"],
+        ["おはよ"],
+        ["えらいね"],
         ["にゃん"],
-        ["Wikipedia の情報は入っているのかもしれない"],
+        ["よしよし"],
     ],
     description=DESCRIPTION,
     css_paths="style.css",