TownPass on-device assistant models

Model files downloaded by the 台北一點就通!微服務入口整合助理 (an on-device assistant for the TownPass / 台北通 app fork, https://github.com/goodmorning1226/TownPass) on first use, after the user agrees. Everything runs on the phone; no recording or text is sent anywhere. Only files we converted or built are here; the other models (Whisper tiny/small, Ministral 3B, Gemma 4 E2B) are downloaded from their authors' own Hugging Face repositories.

The App checks every file's size and SHA-256 against its built-in manifest (assets/assistant/models.json) before using it.

本 repo 放的是「台北一點就通!微服務入口整合助理」在手機上執行時,使用者同意後才下載的模型檔。 只放我們自己轉檔或建立的檔案;其餘模型從原作者的 Hugging Face 下載。每個檔案的授權如下表。

Path Size SHA-256 Original License What we did
e5/multilingual-e5-small-q8_0.gguf 132,441,760 1d76aaf0627f8385090b4027fcd3827861e7d51f1e27c9f0f479551b6da810d3 intfloat/multilingual-e5-small (Microsoft) MIT, see LICENSES/MIT-multilingual-e5.txt Converted to GGUF and quantized to q8_0 for llama.cpp
decision/townpass-decision-e5-ce-v1-q8_0.gguf 132,441,120 ae1745a7a24c1217271f08e53f140df1ea9e3c50f098362045625290d6b6e22f Fine-tuned by this project from intfloat/multilingual-e5-small (Microsoft) MIT, see LICENSES/MIT-multilingual-e5.txt Trained as a cross-encoder that scores a request against one App page (decision classifier), on synthetic requests written for this project plus the App's own page names, keywords and examples; evaluation sentences were kept out. Converted to GGUF, quantized to q8_0
faq_rag/embeddings.bin 12,473,856 483043bfeddb364bcd43a84866341f99b552e861d10f1c74df04a670028c9a19 Embeddings of the Taipei City FAQ (below) by the e5 model above Open Government Data License, version 1.0 (data) 8,121 entries × 384 float32, normalized, row-major
faq_rag/metadata.json 6,572,186 b771de1a423dcb13ae504e49d0f522499ed463085810a86aeb9ab31410e0bbd4 臺北市市政網站整合平台之常見問答, Taipei City Government Open Government Data License, version 1.0 Questions, answers and departments, as published
faq_rag/model_info.json 409 999d8469d836348696b62882e13fc2c3097937b7ce4c9c3cb24cb8c9bcd049c1 — — Index description (model, prefixes, dimensions)
vietnamese/ggml-phowhisper-base-q8_0.bin 81,768,602 9e78100d845978524462f17c2a2f22770e288393719973056033526335621d2a vinai/PhoWhisper-base (VinAI) BSD-3-Clause, see LICENSES/BSD-3-Clause-PhoWhisper.txt Converted to whisper.cpp ggml and quantized to q8_0
taiwanese/ggml-breeze-asr-26-q8_0.bin 1,656,129,708 e878f611bf3ef407cff3c1e2ecf5f60dc92bc6dd09a85b22166731c0d92764a1 MediaTek-Research/Breeze-ASR-26 (MediaTek Research) Apache-2.0, see LICENSES/Apache-2.0.txt Converted with whisper.cpp convert-h5-to-ggml.py, quantized to q8_0 with whisper-quantize
taiwanese/ggml-breeze-asr-26-q4_k.bin 888,932,908 747896198b388e1ab77397cb82136dd3bd5fc1823f7e0c60757166b20ff9abbe same as above Apache-2.0 Same conversion, quantized to q4_K (used by App builds before 2026-11-13)
taiwanese/ggml-breeze-asr-26-q4_0.bin 888,932,908 0e08c5a2d3a5ba4bae7f85d02212e887537f78d7e623d8694313e2dd65de22ec same as above Apache-2.0 Same conversion, quantized to q4_0 (iPhone, and Android phones with 8 GB of memory, from 2026-11-13: as accurate as q8_0 in our evaluation, less memory)

Conversions change only the file format and numeric precision, not what the models were trained on (the decision classifier above is the one file we trained ourselves). They are not endorsed by the original authors.

Attribution

  • Taipei City FAQ data: 本資料依「政府資料開放授權條款-第1版」提供,來源:臺北市政府(臺北市市政網站整合平台之常見問答)。 The FAQ data is provided under the Open Government Data License, version 1.0 (Taiwan); source: Taipei City Government.
  • Breeze-ASR-26: MediaTek Research. Please cite the original model card and paper (Lan et al., 2026, arXiv:2603.19259).
  • PhoWhisper: VinAI Research.
  • multilingual-e5-small: Wang et al., Microsoft (also the base of the decision classifier).

Use

These files are loaded by whisper.cpp whisper-server (speech) and llama.cpp llama-server --embedding (e5) and llama-server --reranking --pooling rank (decision classifier) inside the App. The App downloads a path here as https://huggingface.co/allensung1226/townpass-assistant-models/resolve/main/<path> and checks its SHA-256 before use.

Downloads last month
153
GGUF
Model size
0.1B params
Architecture
bert
Hardware compatibility
Log In to add your hardware

8-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Paper for allensung1226/townpass-assistant-models