HaS Text Model (Q8_0)

HaS (Hide and Seek) is an on-device privacy model providing a complete pipeline from entity recognition to anonymization and restoration.

• 📦 0.6B parameters, Q8_0 quantization, 639 MB
• 🔒 Data never leaves device — local inference, no network required
• 🌍 8 languages natively supported: Chinese, English, Portuguese, French, Spanish, German, Korean, Japanese
• ⚡ Apple M4 benchmark: prefill 1,600–2,800 tok/s, generation 96–120 tok/s

1. Core Capabilities

Traditional anonymization (regex, Presidio, etc.) only does pattern matching. HaS is an on-device Agentic privacy pipeline — a set of composable atomic capabilities that solve multi-turn consistency, reversible restoration, and post-anonymization data usability.

2. Six Atomic Capabilities

3. Structured Semantic Tags & Coreference Resolution

3-Level Semantic Tags

Tags use a <EntityType[ID].Category.Attribute> three-level structure:

<Address[1].City.CityName>             ← identifies this as a city name
<Address[2].StreetAddress.FullAddress>  ← identifies this as a detailed address
<Amount[1].ContractAmount.NumberSymbol> ← identifies this as a contract amount
<Phone[1].Mobile.FullNumber>           ← identifies this as a mobile number

Comparison with traditional approaches:

Coreference Resolution

The same entity often appears in multiple forms. HaS automatically recognizes they refer to the same object and unifies them under one ID:

Original forms               Unified tag
───────────────────           ───────────────────────
CloudGenius Inc.          →   <Organization[1].Company.Name>
CloudGenius               →   <Organization[1].Company.Name>
云创智能                   →   <Organization[1].Company.Name>
CG                        →   <Organization[1].Company.Name>

This ensures anonymized text remains logically coherent — LLMs seeing multiple <Organization[1]> know it's the same company. Critical for multi-turn conversations and long document chunking: entity IDs remain globally consistent across turns and chunks.

4. Quick Start

Recommended deployment with llama.cpp:

llama-server -m has_text_model.gguf -ngl 999 -c 8192 -np 1 -fa on -ctk q8_0 -ctv q8_0

• Listens on http://127.0.0.1:8080/v1 by default
• OpenAI Chat Completions compatible API
• ~1.56 GB total memory with recommended settings

5. Usage Scenarios

The 6 atomic capabilities can be composed into various privacy pipelines:

6. Prompt Templates

⚠️ Templates must match character-for-character — the model was trained on these exact templates. Any deviation may degrade output quality.

NER

Recognize the following entity types in the text.
Specified types:{types_json_array}
<text>{text}</text>

Hide_with (with mapping)

Turn 1: Same as NER template

Turn 2:

Replace the above-mentioned entity types in the text according to the existing mapping pairs:{mapping_json}

Hide_without (without mapping)

Turn 1: Same as NER template

Turn 2 (fixed text, no variables):

Replace the above-mentioned entity types in the text.

Pair

<original>{original_text}</original>
<anonymized>{anonymized_text}</anonymized>
Extract the mapping from anonymized entities to original entities.

Split

Split each composite anonymized key into atomic keys.
Composite mapping:
{composite_mapping_json_array}

Seek

The mapping from anonymized entities to original entities:
{mapping_json}
Restore the original text based on the above mapping:
{text_with_tags}

7. Speed Benchmarks

Test platform: Apple M4, Q8_0 model, llama-server recommended settings

HaS ships with a CLI tool has-text that orchestrates model capabilities with programmatic tools into ready-to-use commands (scan, hide, seek). The following are end-to-end CLI times:

• scan = Model-NER
• hide = Model-NER → Model-Hide → Tool-Pair → Tool-Mapping Merge (with self-check; Model-Split called for composite tags)
• seek = Tool-Language Detection → Tool-Seek (string replacement) or Model-Seek (cross-language) → self-check

• Same-language restoration uses string replacement (constant time); cross-language automatically switches to model inference
• 8K context per chunk. With recursive chunking, can process documents of hundreds of thousands of tokens

8. Quantization Versions

---

中文版

HaS Text Model (Q8_0)

HaS（Hide and Seek） 是一个端侧部署的隐私模型，提供从实体识别到脱敏还原的完整管线。

• 📦 0.6B 参数，Q8_0 量化，639 MB
• 🔒 数据不出设备，本地推理，无需联网
• 🌍 8 语言原生支持：中、英、葡、法、西、德、韩、日
• ⚡ Apple M4 实测：prefill 1,600–2,800 tok/s，生成 96–120 tok/s

一、核心能力

传统脱敏方案（正则、Presidio 等）只做模式匹配。HaS 的定位是端侧 Agentic 隐私管线——用一组可组合的原子能力解决多轮一致、可逆还原和脱敏后数据可用性的问题。

二、6 个原子能力

三、结构化语义标签与指代消解

三级语义标签

脱敏后的标签采用 <实体类型[编号].分类.属性> 三级结构：

<地址[1].城市.市名>          ← 知道这是一个城市名
<地址[2].街道门牌.完整地址>   ← 知道这是一个详细地址
<金额[1].合同金额.数字符号>   ← 知道这是一笔合同金额，不只是普通数字
<电话[1].手机号.完整号码>     ← 知道这是手机号，不是座机或传真

对比传统脱敏方案：

同样是地址，三级标签区分了"城市名"和"详细地址"；同样是金额，标签告诉你这是"合同金额"。脱敏后的文本对 LLM 仍然可理解、可推理。

指代消解

同一实体在文本中往往以多种形式出现。HaS 会自动识别它们指向同一对象，统一归为同一编号：

原文中的写法              脱敏后统一为
───────────────────       ───────────────────────
云创智能科技有限公司   →   <组织[1].企业.名称>
云创智能              →   <组织[1].企业.名称>
CloudGenius           →   <组织[1].企业.名称>
云创                  →   <组织[1].企业.名称>

这确保了脱敏后的文本逻辑自洽——LLM 看到多处 <组织[1]> 就知道是同一家公司，而不会误以为是不同实体。在多轮对话和长文档分块中尤为关键：跨轮次、跨分块的实体编号全局一致。

四、快速开始

推荐使用 llama.cpp 推理框架：

llama-server -m has_text_model.gguf -ngl 999 -c 8192 -np 1 -fa on -ctk q8_0 -ctv q8_0

• 默认监听 http://127.0.0.1:8080/v1
• API 兼容 OpenAI Chat Completions 格式
• 推荐配置下总内存约 1.56 GB

五、使用场景

6 个原子能力可以组合成多种隐私管线：

上述场景仅为典型示例。6 个原子能力可自由组合，适配任意隐私需求。

六、提示词模板

⚠️ 模板必须逐字符精确匹配，模型基于这些模板训练。任何偏差（空格、换行、标点）都可能降低输出质量。

NER

Recognize the following entity types in the text.
Specified types:{types_json_array}
<text>{text}</text>

• {types_json_array}：JSON 数组，紧接 Specified types: 无空格。如 ["组织","地址","人名"]
• {text}：用户原始文本

输出：JSON 对象，key 为实体类型，value 为识别到的实体数组。

Hide_with（带映射脱敏）

第 1 轮：与 NER 模板相同

第 2 轮：

Replace the above-mentioned entity types in the text according to the existing mapping pairs:{mapping_json}

• {mapping_json}：已有映射字典，类型 Record<string, string[]>

输出：脱敏后文本。

Hide_without（无映射脱敏）

第 1 轮：与 NER 模板相同

第 2 轮（固定文本，无变量）：

Replace the above-mentioned entity types in the text.

输出：脱敏后文本，标签由模型自主生成。

Pair（提取映射）

<original>{original_text}</original>
<anonymized>{anonymized_text}</anonymized>
Extract the mapping from anonymized entities to original entities.

输出：Record<string, string[]> 映射 JSON。

Split（拆分复合标签）

Split each composite anonymized key into atomic keys.
Composite mapping:
{composite_mapping_json_array}

输出：拆分后的原子映射。

Seek（还原）

The mapping from anonymized entities to original entities:
{mapping_json}
Restore the original text based on the above mapping:
{text_with_tags}

输出：还原后的原文。

七、速度评估

测试平台：Apple M4，Q8_0 模型，llama-server 推荐配置

HaS 配套了 CLI 工具 has-text，将模型的原子能力与程序化工具编排为开箱即用的命令（scan、hide、seek）。以下为 CLI 端到端耗时，每个命令内部组合了模型调用与工具调用：

• scan = Model-NER
• hide = Model-NER → Model-Hide → Tool-Pair → Tool-Mapping Merge（含自检；遇复合标签时额外调用 Model-Split）
• seek = Tool-Language Detection → Tool-Seek（字符串替换）或 Model-Seek（跨语言时）→ 自检

• 扫描+脱敏 包含多步编排（NER → 替换 → 映射提取），比 仅扫描 慢 2–3 倍
• 同语言还原为字符串替换，耗时恒定；经过翻译/改写后自动切换模型推理
• 8K 是单次窗口。配合递归分块，可处理数十万 token 级文档

八、量化版本