Spaces:
Sleeping
Sleeping
feat: gguf for spaces
Browse files- .gitignore +3 -0
- README.md +20 -1
- requirements.txt +3 -66
- src/app.py +11 -32
.gitignore
CHANGED
|
@@ -1,5 +1,8 @@
|
|
|
|
|
| 1 |
models/
|
|
|
|
| 2 |
src/data/
|
|
|
|
| 3 |
tmp/
|
| 4 |
|
| 5 |
# Byte-compiled / optimized / DLL files
|
|
|
|
| 1 |
+
fig_memo/
|
| 2 |
models/
|
| 3 |
+
practice/
|
| 4 |
src/data/
|
| 5 |
+
src/model/
|
| 6 |
tmp/
|
| 7 |
|
| 8 |
# Byte-compiled / optimized / DLL files
|
README.md
CHANGED
|
@@ -46,4 +46,23 @@ LLMやBERTなどの自然言語処理技術を使ったプロジェクトの練
|
|
| 46 |
├── collect # データセットを作成する
|
| 47 |
├── data
|
| 48 |
└── app.py
|
| 49 |
-
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 46 |
├── collect # データセットを作成する
|
| 47 |
├── data
|
| 48 |
└── app.py
|
| 49 |
+
```
|
| 50 |
+
|
| 51 |
+
## 実行方法
|
| 52 |
+
|
| 53 |
+
- ローカル
|
| 54 |
+
|
| 55 |
+
|
| 56 |
+
## メモ
|
| 57 |
+
### モデルについて
|
| 58 |
+
- (2025/5/10)LLMをCPUで使用するのはかなり厳しい。gguf形式のものを適切に使用すれば可能かもしれないが、まずはt5などを使用する?
|
| 59 |
+
- LLMについて比較を行った結果
|
| 60 |
+
- SakanaAI/TinySwallow-1.5B-Instruct(1.5Bということを考慮に入れるとgemma3以上?)
|
| 61 |
+
- gguf形式ならCPUでも推論可能なはず。だけどcolabで6分かかる、、、
|
| 62 |
+
- google/gemma-3-4b-it(圧倒的。1bは英語のみ対応)
|
| 63 |
+
- Rakuten/RakutenAI-2.0-mini-instruct(かなり良い)
|
| 64 |
+
- rinna/gemma-2-baku-2b-it(そこそこ。実行方法が悪い?)
|
| 65 |
+
- google/gemma-2-2b-jpn-it(同)
|
| 66 |
+
- meta-llama/Llama-3.2-3B-Instruct(日本語対応してない)
|
| 67 |
+
- microsoft/Phi-4-mini-instruct
|
| 68 |
+
- lightblue/DeepSeek-R1-Distill-Qwen-1.5B-Multilingual
|
requirements.txt
CHANGED
|
@@ -1,66 +1,3 @@
|
|
| 1 |
-
|
| 2 |
-
|
| 3 |
-
|
| 4 |
-
attrs==25.1.0
|
| 5 |
-
blinker==1.9.0
|
| 6 |
-
cachetools==5.5.1
|
| 7 |
-
certifi==2025.1.31
|
| 8 |
-
charset-normalizer==3.4.1
|
| 9 |
-
click==8.1.8
|
| 10 |
-
comm==0.2.2
|
| 11 |
-
debugpy==1.8.9
|
| 12 |
-
decorator==5.1.1
|
| 13 |
-
diskcache==5.6.3
|
| 14 |
-
executing==2.1.0
|
| 15 |
-
gitdb==4.0.12
|
| 16 |
-
GitPython==3.1.44
|
| 17 |
-
idna==3.10
|
| 18 |
-
ipykernel==6.29.5
|
| 19 |
-
ipython==8.30.0
|
| 20 |
-
jedi==0.19.2
|
| 21 |
-
Jinja2==3.1.4
|
| 22 |
-
jsonschema==4.23.0
|
| 23 |
-
jsonschema-specifications==2024.10.1
|
| 24 |
-
jupyter_client==8.6.3
|
| 25 |
-
jupyter_core==5.7.2
|
| 26 |
-
llama_cpp_python==0.3.2
|
| 27 |
-
markdown-it-py==3.0.0
|
| 28 |
-
MarkupSafe==3.0.2
|
| 29 |
-
matplotlib-inline==0.1.7
|
| 30 |
-
mdurl==0.1.2
|
| 31 |
-
narwhals==1.27.1
|
| 32 |
-
nest-asyncio==1.6.0
|
| 33 |
-
numpy==2.1.3
|
| 34 |
-
packaging==24.2
|
| 35 |
-
pandas==2.2.3
|
| 36 |
-
parso==0.8.4
|
| 37 |
-
pexpect==4.9.0
|
| 38 |
-
pillow==11.1.0
|
| 39 |
-
platformdirs==4.3.6
|
| 40 |
-
prompt_toolkit==3.0.48
|
| 41 |
-
protobuf==5.29.3
|
| 42 |
-
psutil==6.1.0
|
| 43 |
-
ptyprocess==0.7.0
|
| 44 |
-
pure_eval==0.2.3
|
| 45 |
-
pyarrow==19.0.0
|
| 46 |
-
pydeck==0.9.1
|
| 47 |
-
Pygments==2.18.0
|
| 48 |
-
python-dateutil==2.9.0.post0
|
| 49 |
-
pytz==2024.2
|
| 50 |
-
pyzmq==26.2.0
|
| 51 |
-
referencing==0.36.2
|
| 52 |
-
requests==2.32.3
|
| 53 |
-
rich==13.9.4
|
| 54 |
-
rpds-py==0.22.3
|
| 55 |
-
six==1.17.0
|
| 56 |
-
smmap==5.0.2
|
| 57 |
-
stack-data==0.6.3
|
| 58 |
-
streamlit==1.42.1
|
| 59 |
-
tenacity==9.0.0
|
| 60 |
-
toml==0.10.2
|
| 61 |
-
tornado==6.4.2
|
| 62 |
-
traitlets==5.14.3
|
| 63 |
-
typing_extensions==4.12.2
|
| 64 |
-
tzdata==2024.2
|
| 65 |
-
urllib3==2.3.0
|
| 66 |
-
wcwidth==0.2.13
|
|
|
|
| 1 |
+
huggingface-hub
|
| 2 |
+
llama-cpp-python
|
| 3 |
+
streamlit
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
src/app.py
CHANGED
|
@@ -1,32 +1,19 @@
|
|
| 1 |
-
from
|
|
|
|
| 2 |
import streamlit as st
|
| 3 |
-
import torch
|
| 4 |
-
from transformers import AutoTokenizer, pipeline
|
| 5 |
|
| 6 |
-
|
| 7 |
|
| 8 |
def summarize_article(input_text):
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH, local_files_only=True)
|
| 13 |
-
model = ORTModelForSeq2SeqLM.from_pretrained(MODEL_PATH, local_files_only=True)
|
| 14 |
-
|
| 15 |
-
summarizer = pipeline(
|
| 16 |
-
"summarization",
|
| 17 |
-
model=model,
|
| 18 |
-
tokenizer=tokenizer
|
| 19 |
-
)
|
| 20 |
-
return summarizer(
|
| 21 |
-
input_text,
|
| 22 |
-
max_length=600,
|
| 23 |
-
min_length=200,
|
| 24 |
-
do_sample=True,
|
| 25 |
-
temperature=0.5,
|
| 26 |
-
num_beams=4,
|
| 27 |
-
early_stopping=True
|
| 28 |
-
)
|
| 29 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 30 |
|
| 31 |
# ページ設定
|
| 32 |
st.set_page_config(
|
|
@@ -41,19 +28,11 @@ st.subheader("入力を元に要約を生成します")
|
|
| 41 |
|
| 42 |
# 入力フォーム
|
| 43 |
with st.form("input_form"):
|
| 44 |
-
# novel_title = st.text_input("小説のタイトル", placeholder="例:人間失格")
|
| 45 |
input_text = st.text_area("記事内容", height=200, placeholder="例:主人公の葉蔵は自分を「人間失格」だと考えている...")
|
| 46 |
submit_button = st.form_submit_button("生成")
|
| 47 |
|
| 48 |
# 送信ボタンが押されたら結果を表示
|
| 49 |
if submit_button:
|
| 50 |
-
# st.markdown("## 入力内容")
|
| 51 |
-
# # st.write(f"**タイトル:** {novel_title}")
|
| 52 |
-
# st.write("**あらすじや感想メモ:**")
|
| 53 |
-
# st.write(summary)
|
| 54 |
-
|
| 55 |
-
# st.markdown("---")
|
| 56 |
-
|
| 57 |
summary = summarize_article(input_text)
|
| 58 |
|
| 59 |
st.markdown("## 生成された感想記事(デモ)")
|
|
|
|
| 1 |
+
from huggingface_hub import hf_hub_download
|
| 2 |
+
from llama_cpp import Llama
|
| 3 |
import streamlit as st
|
|
|
|
|
|
|
| 4 |
|
| 5 |
+
MAX_OUTPUT_TOKENS = 512
|
| 6 |
|
| 7 |
def summarize_article(input_text):
|
| 8 |
+
repo_id = "SakanaAI/TinySwallow-1.5B-Instruct-GGUF"
|
| 9 |
+
filename = "tinyswallow-1.5b-instruct-q5_k_m.gguf"
|
| 10 |
+
model_path = hf_hub_download(repo_id=repo_id, filename=filename)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 11 |
|
| 12 |
+
# モデルの読み込み
|
| 13 |
+
llm = Llama(model_path=model_path, n_ctx=4096, n_gpu_layers=-1)
|
| 14 |
+
prompt = f"以下のテキストを日本語で約400字程度に要約してください。特に固有名詞や専門用語は正確に含めてください。テキスト: {input_text} 要約: "
|
| 15 |
+
response = llm(prompt, max_tokens=MAX_OUTPUT_TOKENS)
|
| 16 |
+
return response["choices"][0]["text"]
|
| 17 |
|
| 18 |
# ページ設定
|
| 19 |
st.set_page_config(
|
|
|
|
| 28 |
|
| 29 |
# 入力フォーム
|
| 30 |
with st.form("input_form"):
|
|
|
|
| 31 |
input_text = st.text_area("記事内容", height=200, placeholder="例:主人公の葉蔵は自分を「人間失格」だと考えている...")
|
| 32 |
submit_button = st.form_submit_button("生成")
|
| 33 |
|
| 34 |
# 送信ボタンが押されたら結果を表示
|
| 35 |
if submit_button:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 36 |
summary = summarize_article(input_text)
|
| 37 |
|
| 38 |
st.markdown("## 生成された感想記事(デモ)")
|