Upload TRAINING_DOCUMENTATION.md with huggingface_hub
Browse files- TRAINING_DOCUMENTATION.md +15 -11
TRAINING_DOCUMENTATION.md
CHANGED
|
@@ -332,20 +332,24 @@ Step 9: Deploy with vLLM / 部署
|
|
| 332 |
|
| 333 |
### Automated Test Suite / 自動化測試套件
|
| 334 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 335 |
| Category / 類別 | Tests / 測試數 | Pass Rate / 通過率 |
|
| 336 |
|---|---|---|
|
| 337 |
-
| Identity / 身份 |
|
| 338 |
-
| Emotion (EN) / 情緒（英文） |
|
| 339 |
-
| Emotion (ZH) / 情緒（中文） |
|
| 340 |
-
|
|
| 341 |
-
|
|
| 342 |
-
| Wisdom / 智慧 | 26 | 100% |
|
| 343 |
-
| Forbidden phrases / 禁止詞彙 | 16 | 100% |
|
| 344 |
-
| Language consistency / 語言一致性 | 18 | 100% |
|
| 345 |
| Prompt injection / 提示注入 | 20 | 100% |
|
| 346 |
-
|
|
| 347 |
-
|
|
| 348 |
-
|
|
|
|
|
|
|
|
|
|
| 349 |
|
| 350 |
**English:** The automated test suite sends 200 unique prompts to the live API and validates each response against rules: (1) starts with a valid reaction tag, (2) 3-5 sentences, (3) no forbidden phrases, (4) correct language match, (5) refusal detected for non-companion tasks, (6) injection resistance (stays in character, doesn't comply with attacker commands).
|
| 351 |
|
|
|
|
| 332 |
|
| 333 |
### Automated Test Suite / 自動化測試套件
|
| 334 |
|
| 335 |
+
**English:** 220/220 (100%) tests passed via vLLM with 4-bit bitsandbytes quantization in a Podman container — the production deployment path. Tests were run through the Jingsi API (port 8002) which proxies to vLLM (port 8000), applying the system prompt and 6 post-processing guardrails.
|
| 336 |
+
|
| 337 |
+
**繁體中文：** 220/220（100%）測試通過，使用 vLLM 4-bit bitsandbytes 量化於 Podman 容器中 — 即生產部署路徑。測試通過 Jingsi API（埠 8002）代理至 vLLM（埠 8000），套用系統提示和 6 道後處理守護欄。
|
| 338 |
+
|
| 339 |
| Category / 類別 | Tests / 測試數 | Pass Rate / 通過率 |
|
| 340 |
|---|---|---|
|
| 341 |
+
| Identity / 身份 | 40 | 100% |
|
| 342 |
+
| Emotion (EN) / 情緒（英文） | 40 | 100% |
|
| 343 |
+
| Emotion (ZH) / 情緒（中文） | 40 | 100% |
|
| 344 |
+
| Refusal / 拒絕 | 40 | 100% |
|
| 345 |
+
| 台語 (Taiwanese) | 20 | 100% |
|
|
|
|
|
|
|
|
|
|
| 346 |
| Prompt injection / 提示注入 | 20 | 100% |
|
| 347 |
+
| Wisdom / 智慧 | 20 | 100% |
|
| 348 |
+
| **Total / 總計** | **220** | **100%** |
|
| 349 |
+
|
| 350 |
+
**English:** All tests passed through the 4-bit bitsandbytes quantized model served by vLLM in a Podman container, confirming that 4-bit quantization preserves 100% of the model's behavioral quality compared to 16-bit inference.
|
| 351 |
+
|
| 352 |
+
**繁體中文：** 所有測試均通過 vLLM Podman 容器中的 4-bit bitsandbytes 量化模型，確認 4-bit 量化相較於 16-bit 推理保持了 100% 的模型行為品質。
|
| 353 |
|
| 354 |
**English:** The automated test suite sends 200 unique prompts to the live API and validates each response against rules: (1) starts with a valid reaction tag, (2) 3-5 sentences, (3) no forbidden phrases, (4) correct language match, (5) refusal detected for non-companion tasks, (6) injection resistance (stays in character, doesn't comply with attacker commands).
|
| 355 |
|