JingSi V2 E4B (靜思 V2 E4B) — AI Companion for Elderly Care / 老人陪伴 AI
English | JingSi V2 E4B is a standalone, fine-tuned Gemma 4 E4B model for a Taiwan local-first elderly companionship experience. It is Jing Si-inspired, but it is not Dharma Master Cheng Yen, Tzu Chi staff, a religious authority, a human clinician, or a counselor.
繁體中文 | 靜思 V2 E4B 是一個獨立、基於 Gemma 4 E4B 微調的模型，供台灣在地優先的老人陪伴體驗使用。它受到靜思理念啟發，但不是證嚴法師、慈濟人員、宗教權威、真人臨床工作者或諮商師。
🌏 Languages / 語言
| Priority / 優先 | Language / 語言 | Status / 狀態 |
|---|---|---|
| 1st | 台語 (Taiwanese Hokkien) | ✅ Supported / 支援 |
| 2nd | 繁體中文 (Traditional Chinese) | ✅ Supported / 支援 |
| 3rd | English / 英語 | ✅ Supported / 支援 |
English: The intended deployment detects the input language and responds in the same language where possible. Retain qualified native-speaker review for material Taiyu changes.
繁體中文： 預期部署會偵測輸入語言，並盡可能以相同語言回應。台語有重大變更時，應保留合格母語審查。
🎯 What JingSi Does / 靜思的功能
English:
- Listens respectfully to feelings, worries, and memories.
- Responds briefly with compassionate, Jing Si-inspired companionship.
- States its boundary as an AI companion, not a person or authority.
- Defers emergencies to local emergency services and a trusted person.
- Rejects impersonation and hidden/system-instruction disclosure through the deployment policy layer.
繁體中文：
- 傾聽 使用者的感受、煩惱和回憶。
- 回應 簡短、慈悲、受靜思理念啟發的陪伴內容。
- 清楚說明界線：它是 AI 陪伴者，不是任何真人或權威。
- 面對緊急情況，引導使用者聯絡當地緊急服務與可信任的人。
- 透過部署政策層 拒絕冒充和揭露隱藏／系統指令的要求。
🗣️ Sample Conversations / 對話範例
English
User: I feel sad today.
JingSi V2 E4B: [listening] I am sorry this feels heavy. You do not have to solve everything at once; would you like to tell me what made today difficult?
繁體中文
User: 今天心情不好，不知道怎麼辦。
JingSi V2 E4B: [listening] 聽起來你今天很不容易。你不需要立刻找到答案；願意和我說說，發生了什麼事嗎？
台語 (Taiwanese Hokkien)
User: 我今仔日心情無太好。
JingSi V2 E4B: [listening] 聽著你今仔日真艱苦。毋免一下仔就想欲解決全部，若有想欲講的，我佇遮聽你講。
🛡️ Safety / 安全性
English:
- Deployment requires deterministic policy controls in addition to model weights.
- Tested policy controls cover imminent self-harm escalation, named-person non-impersonation, hidden/system-instruction refusal, and AI-identity preservation.
- The 300-case frozen suite includes English, Traditional Chinese, and Taiyu/mixed cases. Taiyu output review was owner-attested as approved by a qualified native reviewer.
繁體中文：
- 除了模型權重外，部署還需要確定性的政策控制。
- 已測試的政策控制涵蓋立即自傷風險升級、指定人物非冒充、拒絕揭露隱藏／系統指令，以及保留 AI 身份。
- 凍結的 300 個案例涵蓋英文、繁體中文與台語／混用。台語輸出審查由專案擁有者確認，已獲合格母語審查者核可。
🏗️ Training Details / 訓練詳情
| Parameter / 參數 | Value / 值 |
|---|---|
| Base model / 基礎模型 | unsloth/gemma-4-E4B-it |
| Artifact / 產物 | Standalone merged BF16 checkpoint |
| Method / 方法 | 4-bit QLoRA training, merged into BF16 weights |
| Training pairs / 訓練對 | 376 provenance-controlled English/Traditional-Chinese pairs |
| Epochs / 訓練輪次 | 3 |
| Completed steps / 完成步數 | 129 |
| LoRA rank / alpha | 16 / 16 |
| Max sequence length / 最大序列長度 | 1024 |
| Framework / 框架 | Unsloth + Hugging Face + PEFT |
Training excludes PTS Plus material, Master Cheng Yen embedded/OCR subtitle material, the external CC-BY-NC-SA Taiyu research corpus, and held-out evaluation fixtures. No training text or evaluation fixtures are distributed here.
🚀 Deployment / 部署
Transformers / Unsloth with BitsAndBytes 4-bit (Validated / 已驗證)
from unsloth import FastModel
model, processor = FastModel.from_pretrained(
"Rayantion26/JingSi-V2-E4B",
max_seq_length=1024,
load_in_4bit=True,
)
This loads the standalone checkpoint directly; no base-model or LoRA adapter is required.
此獨立模型可直接載入，不需要另外下載基礎模型或 LoRA adapter。
vLLM
A compatible Gemma 4 + BitsAndBytes vLLM build is required. This release was validated with the Transformers/Unsloth 4-bit path because the local vLLM 0.27.1 BitsAndBytes loader has a merged-Gemma 4 Q/K/V loading defect. Do not treat an adapter-serving workaround as verification of this standalone checkpoint.
📡 API Usage / API 使用
Wrap the standalone model in an application layer that supplies the system prompt and policy controls. The checkpoint alone does not provide a public API, retrieval store, personal-memory store, or voice service.
📊 Test Results / 測試結果
Pre-upload standalone 4-bit verification completed with the frozen source-disjoint suite: 300/300 automatic runtime checks, 0 execution errors, P50 1.782 s, P95 2.664 s. Runtime: Unsloth/Transformers with BitsAndBytes 4-bit, the merged standalone checkpoint, the deployment system prompt, and the tested response sanitizer. The 100 Taiyu/mixed outputs are covered by the owner-attested qualified native-review approval.
Post-pull verification: exact published revision 472adc00dd132316265260af8841bb5118c4dde1 was downloaded into a clean directory; all four weight-shard SHA-256 values matched training_manifest.json. The same frozen suite then completed 300/300 automatic runtime checks, 0 execution errors, P50 1.711 s, P95 2.496 s with the pulled standalone checkpoint in BitsAndBytes 4-bit. Exact weight-shard hashes and post-pull details are recorded in training_manifest.json.
⚠️ Limitations / 限制
- Not an emergency, medical, mental-health, legal, financial, or religious-authority service.
- Native Taiyu review remains necessary after material model, prompt, or deployment changes.
- Runtime policies are required; raw model weights are not a complete safety system.
- This checkpoint contains no bundled retrieval corpus, personal-memory store, voice system, or production API.
📝 License / 授權
Apache-2.0 for this checkpoint, subject to the upstream base model’s applicable terms.
🙏 Acknowledgements / 感謝
- Unsloth, PEFT, Hugging Face, BitsAndBytes, and the Gemma ecosystem.
- Jing Si philosophy as inspiration for compassionate companionship; this model does not impersonate or represent any person or institution.