hamo-score-0.6b

hamo-score-0.6b — the little model that takes your pulse

A 0.6B model that reads one message from a wellness conversation, in Chinese or English, and returns five scores as one JSON object: Agency, Withdrawal, Extremity, Hostility, Boundary (AWEHB), each 0.0–3.0. It does not write replies; its scores feed deterministic code (stress update, state buckets, action gating). 中文说明见下方。

• Scores locally, on CPU. No external API, no conversation data leaving the machine. Distilled to do the job of the LLM scorer behind Hamo AI's production rubric (the reference scorer).
• Small, and fast enough for live use. 0.64 GB as q8 GGUF (base: Qwen3-0.6B). On an ARM CPU server through ollama, with the toolkit's prompt trimming: median 2.13 s per message, slowest 2.91 s of 138 synthetic items.
• Ready to wire in. The toolkit (pip install hamo-score, Apache-2.0): prompt format, upstream keyword gate, one-command Docker server.
• Free to use, commercially too. HAMO-RAIL-S 1.0, with use restrictions.

Message: 今天试着出门散了个步 (tried going out for a walk today), after the assistant asked 这周过得怎么样？ (how was your week?)

Scores (v10): {"A": 1.0, "W": 0.0, "E": 0.0, "H": 0.0, "B": 1.5}

A one-off walk is mid-band Agency; Boundary also credits a report of positive action.

Quickstart

pip install hamo-score

• Library: score_message() runs the keyword gate first and skips scoring if it fires.
• Server: docker compose up in server/ of the toolkit repository serves v10 as POST /score on port 8080.
• Self-check: python eval/run_exam.py in a clone of the repository, against your own ollama endpoint (--base-url); the Docker server exposes only port 8080.
• Weights: gguf/hamo-score-0.6b-v10.q8.gguf. Sampling must be neutral (temperature 0, repeat_penalty 1.0, top_k 0, top_p 1.0): ollama's default repeat_penalty 1.1 pushes scores away from zero. The toolkit's OllamaClient sends the last three with each request.

Prompt template, transformers snippet and Modelfile: technical record.

The five scores

The scores are not meant to be read raw.

Numbers at a glance

The real final exam is 453 pseudonymised turns from three consenting internal staff members, so these are not accuracy figures for the general population. The synthetic exams are in-distribution: passing shows no regression inside the known range, not generalisation.

What it is not

Not a chatbot, not a diagnostic instrument, not a crisis detector. It does not detect or handle crises, and none of its scores, W included, is a crisis signal. Crisis handling is the job of deterministic code that runs before the model: the toolkit's keyword gate is a starting point, not a complete screen, and LICENSE §3c makes an independent upstream mechanism a condition of consumer-facing mental-wellness deployments.

Evaluation

v10 met 17 of the 18 checks of its signed pre-registration. The one it did not meet, G5a, was a stand-in for crisis handling, which the founder ruled outside this model after seeing the result; under the registration as signed the verdict was "rejected", and v10 is released by the founder's decision. The reported checks, exams and method are in the technical record.

Limitations & known residuals

• Scores are not comparable across v7, v9 and v10. Store the version with every score.
• A follows a newer rubric (v8) with no measurement on real conversations yet.
• B can flip sign. With no context, v10 scores B 2.5 on a known self-erasure sentence (giving up one's own position) whose correct B is 0; v7 and v9 score it 0.
• The exams are narrow and reused. The real ones cover three staff members and have judged earlier versions.
• Check the toolkit's stress cut-offs on your own data before they gate anything.
• Full list: technical record.

Training data

Synthetic conversations labelled by an open-weight teacher model, plus 440 pseudonymised turns from consenting internal staff. External user conversations never enter training. Training messages: 76.0% Chinese, 13.1% English, 10.9% mixed; results were not broken down by language.

Versions and files

• v10 is the default on main since 2026-10-06 (UTC). Pin it with revision="f3869312f0222992c3eb2e938e78090de38eb80a".
• Files: gguf/hamo-score-0.6b-v10.q8.gguf; the v9, v7 and v6.1 GGUFs stay in gguf/. Digests and rollback commands: technical record.
• Coming from v9: its narrower B rubric makes computed stress drift upward wherever B has a negative weight, and re-tuning thresholds does not repair it. Move to v10 (or back to v7); do not carry over stress accumulated from v9 scores (full correction).

Links and license

• Toolkit: GitHub · PyPI · integration guide · FAQ · fine-tuning guide.
• Everything else: technical record.
• Disagree with a score? Open a disagreement report.
• License: HAMO-RAIL-S 1.0 (LICENSE): free commercial and non-commercial use, modification and redistribution, with four use restrictions that pass on to recipients (§3a–d). It is deliberately not an OSI-approved open-source license.

中文

给每句话把脉的小模型：0.6B，读心理支持对话里的一条消息（中文、英文都行），输出一个 JSON 对象，里面是五个分数：行动力、退缩、极端化、敌意、边界感（AWEHB），各 0.0–3.0。它只打分，不写回复；分数是给确定性代码用的（压力更新、状态桶、动作门控）。

• 本地评分，CPU 就能跑。 不调外部 API，对话数据不出机器。它学的是 Hamo AI 按生产口径打分的大模型评分器（参照评分器）。
• 小，而且快到能实时用。 q8 GGUF 0.64 GB（基座 Qwen3-0.6B）。在跑 ollama 的 ARM CPU 服务器上，提示词按工具包规则截短，每条消息中位 2.13 秒，138 道合成题里最慢 2.91 秒。
• 拿来就能接。 工具包（pip install hamo-score，Apache-2.0）：提示词格式、放在模型上游的关键词闸门、一条命令启动的 Docker 服务。
• 可自由使用，商用也行。 许可证是 HAMO-RAIL-S 1.0，附使用限制。

消息： 今天试着出门散了个步（上一句是助手问「这周过得怎么样？」）

打分（v10）： {"A": 1.0, "W": 0.0, "E": 0.0, "H": 0.0, "B": 1.5}

单次散步算中档的行动力；边界感的口径也给「报告积极行动」记分。

快速上手

pip install hamo-score

• 库： score_message() 先跑关键词闸门，命中就不评分。
• 服务： 在工具包仓库的 server/ 下执行 docker compose up，8080 端口 POST /score，跑的是 v10。
• 自检： 克隆仓库后运行 python eval/run_exam.py，对着自己的 ollama 服务（--base-url）；Docker 服务只开放 8080 端口。
• 权重： gguf/hamo-score-0.6b-v10.q8.gguf。采样必须取中性值（温度 0、repeat_penalty 1.0、top_k 0、top_p 1.0）：ollama 默认的 repeat_penalty 1.1 会把分数从 0 往上推。工具包的 OllamaClient 每次请求都带后三项。

提示词模板、transformers 代码和 Modelfile 见技术档案。

五个分数

分数不是直接读的。

关键数字

真实终评是三位已同意的内部员工的 453 轮对话，已假名化，所以这些不是一般人群上的准确率。合成考卷与训练同分布：考过只说明已知范围内没有退步，不说明能推广。

它不是什么

不是聊天机器人，不是诊断工具，也不是危机检测器。它不识别也不处理危机；它的分数，包括 W 在内，没有一个是危机信号。危机处理由先于模型运行的确定性代码负责：工具包的关键词闸门只是起点，不是完整筛查；面向消费者的心理健康部署，上游必须有独立机制，这是许可条件（LICENSE §3c）。

评测

签字版预注册的 18 项检查，v10 过了 17 项。没过的那一项 G5a 当初是作为危机处理的替代指标设的；创始人看到结果后裁定，危机处理不在本模型里判定。按签字版预注册，判定是「拒收」；v10 由创始人决定发布。已报告的各项检查、考卷和方法，都在技术档案里。

已知局限

• v7、v9、v10 的分数互不可比。 分数要连同版本一起存。
• A 用较新的 v8 口径，还没有真实对话上的测量。
• B 可能判反（符号翻转）。 一句已知的自我消融短句（放弃自己的立场），不带上下文时 v10 打 B 2.5，正确的 B 是 0；v7、v9 打 0。
• 考卷窄，而且反复用过。 真实考卷只覆盖三位员工，也给此前的版本判过卷。
• 工具包里压力分桶的阈值，请先用自己的数据核对，再让它把关。
• 完整清单见技术档案。

训练数据

合成对话由开放权重的教师模型打标，另有 440 轮假名化的内部员工对话（经本人同意）。外部用户的对话从不进入训练。训练消息里中文占 76.0%，英文 13.1%，中英混杂 10.9%；成绩没有按语言拆开测。

版本与文件

• main 自 2026-10-06（UTC）起默认是 v10。固定版本请用 revision="f3869312f0222992c3eb2e938e78090de38eb80a"。
• 文件： gguf/hamo-score-0.6b-v10.q8.gguf；v9、v7、v6.1 的 GGUF 仍在 gguf/ 下。校验和与回退命令见技术档案。
• 从 v9 过来： v9 的 B 用的是更窄的口径，喂进 B 为负权重的压力公式，算出的压力会向上漂，重调阈值修不了。请换 v10（或退回 v7），按 v9 分数累积的压力值不要沿用（完整更正）。

链接与许可证

• 工具包： GitHub · PyPI · 集成指南 · FAQ · 微调指南。
• 其余内容： 技术档案（中文在下半部分）。
• 对评分有不同意见？ 提分歧报告。
• 许可证： HAMO-RAIL-S 1.0（LICENSE），商用、非商用均可自由使用，也可修改和再分发；另有四条须随权重传递的使用限制（§3a–d）。它不是 OSI 认可的开源许可证，这是有意为之。