new model

README.md

@@ -1,12 +1,11 @@
 ---
-title: 真空ジェネレータ
+title: 真空ジェネレータ v2
 emoji: ⚡
 colorFrom: red
 colorTo: purple
 python_version: 3.11
 models:
-- llm-jp/llm-jp-3-1.8b-instruct
-- vericava/llm-jp-3-1.8b-instruct-lora-vericava17
+- vericava/gpt2-medium-vericava-posts-v3
 sdk: gradio
 sdk_version: 5.23.1
 app_file: app.py

app.py

@@ -21,19 +21,9 @@ MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "32768"))
 
 
 if torch.cuda.is_available():
-    model_id = "vericava/llm-jp-3-1.8b-instruct-lora-vericava17"
-    base_model_id = "llm-jp/llm-jp-3-1.8b-instruct"
-    tokenizer = AutoTokenizer.from_pretrained(base_model_id, trust_remote_code=True)
-    tokenizer.chat_template = "{{bos_token}}{% for message in messages %}{% if message['role'] == 'user' %}{{ '\\n\\n### 前の投稿:\\n' + message['content'] + '' }}{% elif message['role'] == 'system' %}{{ '以下は、SNS上の投稿です。あなたはSNSの投稿生成botとして、次に続く投稿を考えなさい。説明はせず、投稿の内容のみを鉤括弧をつけずに答えよ。' + message['content'] }}{% elif message['role'] == 'assistant' %}{{ '\\n\\n### 次の投稿:\\n' + message['content'] + eos_token }}{% endif %}{% if loop.last and add_generation_prompt %}{{ '\\n\\n### 次の投稿:\\n' }}{% endif %}{% endfor %}"
-    model = AutoModelForCausalLM.from_pretrained(
-        base_model_id,
-        trust_remote_code=True,
-    )
-    model.load_adapter(model_id)
     my_pipeline=pipeline(
         task="text-generation",
-        model=model,
-        tokenizer=tokenizer,
+        model="vericava/gpt2-medium-vericava-posts-v3",
         do_sample=True,
         num_beams=1,
     )
@@ -49,37 +39,29 @@ def generate(
     top_k: int = 50,
     repetition_penalty: float = 1.0,
 ) -> Iterator[str]:
-    from datetime import datetime, timezone, timedelta
-
-    d=datetime.now(timezone(timedelta(hours=9), 'JST'))
-    m=d.month
-    if m < 3 or m > 11:
-        season = '冬'
-    elif m < 6:
-        season = '春'
-    elif m < 9:
-        season = '夏'
-    else:
-        season = '秋'
-
-    h=d.hour
-    go = '午前' if h < 12 else '午後'
-    h = h % 12
-    minute = d.minute
-    time = go + str(h) + '時' + str(minute) + '分'
-
-    messages = [
-        {"role": "system", "content": "なお今は日本の" + season + "で、時刻は" + time + "であるものとする。また、あなたは真空という名前のユーザであるとする。"},
-        {"role": "user", "content": message},
-    ]
+    user_input = " ".join(message.strip().split("\n"))
+    
+    user_input = user_input if (
+        user_input.endswith("。")
+        or user_input.endswith("?")
+        or user_input.endswith("!")
+        or user_input.endswith("？")
+        or user_input.endswith("！")
+    ) else user_input + "。"
 
     output = my_pipeline(
-        messages,
+        user_input,
         temperature=temperature,
         max_new_tokens=max_new_tokens,
+        repetition_penalty=repetition_penalty,
+        top_k=top_k,
+        top_p=top_p,
     )
     print(output)
-    yield output[-1]["generated_text"][-1]["content"]
+    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
+    yield gen_text
 
 demo = gr.ChatInterface(
     fn=generate,