hamo-score-0.6b/README.md

---
license: other
license_name: hamo-rail-s-1.0
license_link: LICENSE
language:
  - zh
  - en
base_model: Qwen/Qwen3-0.6B
pipeline_tag: text-generation
tags:
  - mental-wellness
  - affect-scoring
  - psychology
  - knowledge-distillation
  - qwen3
  - gguf
  - mlx
  - hamo
---

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

*给每句话把脉的小模型（中文版说明见下半部分）*

> ### 👉 Start with the toolkit, not the weights
> `pip install hamo-score` → [**hamo-score-toolkit**](https://github.com/HamoAI/hamo-score-toolkit) (Apache-2.0)
> is the other half of this model: the exact prompt format, the crisis gate this model
> *requires upstream of it*, the smoothing math its scores are designed to feed, a
> one-command Docker server, and a 195-question self-check exam. Weights alone invite the
> one deployment shape this model was designed against.
> **Disagree with a score?** That is the single most useful thing you can send us —
> [open a disagreement report](https://github.com/HamoAI/hamo-score-toolkit/issues/new?template=score_disagreement.md);
> they feed the human gold-label program that steers future versions.

**hamo-score-0.6b** reads one message from a mental-wellness conversation and scores five
psychological pulse signals. It never writes replies. It is the first production-distilled
component of [Hamo AI](https://www.hamo.ai)'s closed-loop wellness engine, released so that
practitioner-supervised tools can run state scoring **locally** — no API, no data leaving the room.

> ⚠️ **What this model is NOT.** It is not a chatbot, not a diagnostic instrument, and
> **not a crisis detector**. In Hamo's own production system, crisis and self-harm content is
> short-circuited by an independent deterministic mechanism *upstream* of this model — it never
> reaches the scorer. Any deployment must reproduce that pattern (see LICENSE §3c).

## The five pulses (AWEHB)

Each user message gets five scores on a 0.0–3.0 scale (0.5 grid):

| Dim | Name | Plain reading |
|---|---|---|
| **A** | Agency | Is the person doing something for themselves? (incl. small plans, coping statements) |
| **W** | Withdrawal | Giving up, avoiding, disengaging? |
| **E** | Extremity | Catastrophizing chains, all-or-nothing thinking? (bounded realistic worry stays LOW) |
| **H** | Hostility | Attacking someone? (venting frustration without a target is NOT hostility) |
| **B** | Boundary | Can they speak from an "I" position — needs, limits, clear stance? |

**A note on B.** Its theoretical root is *differentiation of self* (family-systems sense:
a bounded two-person relationship vs. an enmeshed, undifferentiated one). A per-message scorer
cannot see the relationship — it sees language. So B measures the **linguistic footprint** of
boundaries: "I need… / I'm not willing… / this is my limit" scores high; panicked venting
(self dissolved in affect) scores low; insults are H, not B. B is a per-message signal,
not a relationship diagnosis.

The scores are designed to feed **deterministic downstream code** (stress update, state
buckets, action gating) — in Hamo, an exponential blend `0.8 × history + 0.2 × message`
smooths per-message noise 5× before any decision is taken. We recommend the same pattern.

## Quickstart

> 🚀 **What the toolkit gives you**, in detail:
> - **Library** — prompt format, parsing, the smoothing math, and the license-required
>   crisis gate in `pip install` + a few lines of code;
> - **Reference server** — `docker compose up` fetches the GGUF, warms the model, and
>   exposes the full gate → score → smooth → bucket pipeline as `POST /score`;
> - **Self-check exam** — 195 synthetic teacher-labeled questions + 10 handwritten gate
>   cases, with an official reference band (JSON 100% · dimension-level 84.0% · gate 10/10)
>   so you can verify your wiring reproduces the official numbers;
> - **Fine-tuning guide** — [`docs/finetune.md`](https://github.com/HamoAI/hamo-score-toolkit/blob/main/docs/finetune.md),
>   the six-generation playbook (including the two rejected generations and why) for
>   adapting the scorer to your own population with your own consented data.
>
> Release notes: [EN](https://www.hamoai.tech/blog/open-sourcing-hamo-score-toolkit/) ·
> [中文](https://www.hamo.ai/blog/open-sourcing-hamo-score-toolkit/).

The model was trained on **exactly one prompt format** (its rubric is baked into the weights —
do not add scoring instructions):

```
给来访者最新消息打分（AWEHB，0.0-3.0）。
此前对话:
user: <turn>
assistant: <turn>
最新消息: <message to score>
```

The context block (`此前对话:`) is optional; up to 5 turns are accepted, and the official
toolkit trims to the production-validated guard — last 3 turns × 200 chars, message capped
at 500 chars. Apply the Qwen3
chat template with thinking disabled, temperature 0. Output is a single JSON object.

**transformers**

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch, json, re

m = AutoModelForCausalLM.from_pretrained("HamoAI/hamo-score-0.6b", torch_dtype=torch.bfloat16)
tok = AutoTokenizer.from_pretrained("HamoAI/hamo-score-0.6b")

prompt = "给来访者最新消息打分（AWEHB，0.0-3.0）。\n此前对话:\nassistant: 这周过得怎么样？\n最新消息: 今天试着出门散了个步"
text = tok.apply_chat_template([{"role": "user", "content": prompt}],
                               add_generation_prompt=True, tokenize=False, enable_thinking=False)
out = m.generate(**tok(text, return_tensors="pt"), max_new_tokens=80, do_sample=False)
print(re.search(r"\{[^{}]*\}", tok.decode(out[0])).group())
# {"A": 1.5, "W": 0.0, "E": 0.0, "H": 0.0, "B": 1.0}
```

**ollama / llama.cpp** — a ready `q8_0` GGUF is in [`gguf/`](https://huggingface.co/HamoAI/hamo-score-0.6b/tree/main/gguf). Modelfile:

```
FROM ./hamo-score-0.6b-v7.q8.gguf
TEMPLATE """<|im_start|>user
{{ .Prompt }}<|im_end|>
<|im_start|>assistant
<think>

</think>

"""
PARAMETER temperature 0
PARAMETER num_predict 80
PARAMETER stop <|im_end|>
PARAMETER repeat_penalty 1.0
PARAMETER top_k 0
PARAMETER top_p 1.0
```

> ⚠️ **The last three parameters are not optional.** ollama defaults to
> `repeat_penalty 1.1`. This model's output — `{"A": 0.0, "W": 0.0, "E": 0.0,
> "H": 0.0, "B": 0.0}` — is deliberately repetitive, so penalising repeated
> tokens pushes every score *away from zero*, fabricating signal that isn't
> there. Measured on our 300-question boundary-discrimination exam, same
> weights, sampling as the only variable: fabrication rate **13.5% → 25.0%**,
> boundary sign-flips (a true `0.0` scored `≥2.0`) **6 → 10**. Earlier
> revisions of this card omitted them; if you deployed from those
> instructions, add them and re-create the model.

Parse the first `{...}` in the response (the model may emit an empty `<think>` block first).

## Evaluation

Held-out exam: **758 real, pseudonymised production turns** (labels = the production-scale
LLM scorer this model replaces; the exam turns are **never trained on** — the only real data
in training is the separately disclosed 440 consented staff turns, see "How it was trained").

| Metric | Better is | **v7 (this release)** | v6.1 (previous) | Teacher (DeepSeek, qualification paper) | Reference scorer self-consistency* |
|---|---|---|---|---|---|
| Dimension-level, within ±0.5 | ↑ higher | **85.1%** (A84 / W87 / E87 / H94 / B74) | 85.6% | 88.7% | 94–98% |
| Decision-level (state bucket after deterministic stress calc) | ↑ higher | **97.1%** | 96.2% | 97.5% | — |
| Crisis-phrase W-recall misses (count, not %) | ↓ **lower** | **3** | 4 | — | — |
| JSON validity | ↑ higher | ~100% | ~100% | — | — |

**Read the two headline rows together.** v7 trades ~0.5 pt of raw agreement with the
reference scorer's labels for +0.9 pt on the decision the product actually consumes, and one
fewer crisis miss. The dimension-level dip is expected rather than a regression: v7 was
trained on a re-labelled corpus (the teacher re-scored every row at temperature 0, removing
label noise), so the student is now more faithful to the *denoised* labels and correspondingly
slightly less aligned with the noisier reference it was originally distilled from. Measured
directly: agreement with the temperature-0 teacher on boundary scores rose **57.3% → 69.3%**
while agreement with the original reference scorer on the same dimension stayed flat.

Evaluated on an untouched 453-turn final split of the real-conversation exam (never trained
on, never used for checkpoint selection; gold labels include 8 human corrections).

**Residual textual overlap, measured rather than assumed.** The exam and training splits are
disjoint by sample and by conversation, but the consented staff contributors repeat themselves
across sessions: 42 of the 453 final-exam turns (9.3%) carry a message text that also occurs
somewhere in the 440 consented training turns, 11 of them (2.4%) with the same recent context.
We scored those rows separately. For v7: on the 42 overlapping turns the model reaches 88.1%
dimension-level and 100% decision-level; on the 411 with no textual overlap, **84.8% and
96.8%**. So the headline figures carry roughly **+0.3 points** of optimism from this
effect — small, and stated here rather than left for someone else to find. (v6.1 measured the
same way: 89.0%/100% on the overlap, 85.2%/95.9% clean.) (Only 5 of the 42
also share the training row's label vector: the same sentence usually earns different scores in
a different turn, so these are not free points, merely easier ones.)

### Boundary discrimination — what v7 was actually built to fix

Agreement metrics hide the failure that mattered most. On a 300-question exam built
specifically to probe the B (Boundary) decision surface — 60 matched pairs plus 180 singletons
across nine trap cells — v6.1 was reading **self-erasure as strong boundary**: "行，我全听你的，
你说哪天去就哪天去" ("fine, I'll do whatever you say, you decide") scored B=2.5 where the truth
is 0.0. That is a sign error on B's negative pole, not a calibration wobble, and B carries the
largest single weight in the downstream stress formula.

| 300-question boundary exam | Better is | v6.1 | **v7** |
|---|---|---|---|
| Fabrication — a true 0 scored ≥1.0 (sees boundary that isn't there) | ↓ **lower** | 13.5% | **2.9%** |
| Miss — a true high scored ≤0.5 (misses a real boundary) | ↓ **lower** | 16.1% | **12.8%** |
| **Sign flips** — a true 0.0 scored ≥2.0 (count; reads self-erasure as strong boundary) | ↓ **lower** | 6 | **0** |
| Paired direction accuracy — ranks the higher-boundary arm above the lower | ↑ higher | 81.7% | **98.3%** |

Three of the four rows are error rates, so **lower is better on the first three and higher on
the last**; v7 improves on all four.

Both columns measured on the shipped q8 GGUF with neutral sampling, so they describe the
artifact you download rather than an internal checkpoint. Direction accuracy is the cleanest
of the four — it asks only whether the model ranks the higher-boundary arm of a matched pair
above the lower one, so it is immune to absolute calibration.

\* Self-consistency = the same messages scored twice by the reference scorer in two live
environments; its own agreement is only 94–98% at dimension level — the practical ceiling.

Latency (single message, warm): ~0.8 s on Apple M1 Pro (MLX bf16); 1.5–2.9 s on a
2-vCPU ARM server (q8 GGUF, CPU-only). Crisis-phrase W-recall improved 2× in the v4
generation and edged further down in v6.1 (final-exam misses 11 → 5 → 4) — but see the
crisis disclaimer above: recall here is defense-in-depth, not the defense.

## Community quantizations — and what we measured on them

[**mradermacher/hamo-score-0.6b-GGUF**](https://huggingface.co/mradermacher/hamo-score-0.6b-GGUF)
provides static GGUF quants of this model from Q2_K to f16 — twelve build targets we never
shipped ourselves. Thanks to mradermacher for the work, and for carrying the RAIL-S license
terms through redistribution.

Our published metrics were measured on **Q8_0**. Because this model's read-outs gate how deep
a conversation may go, quantization damage here is a clinical question rather than a
perplexity number — so we ran two community builds through the same 453-turn final exam, same
prompt, same parser:

| | Q8_0 (ours, published) | Q6_K (community) | Q4_K_M (community) |
|---|---|---|---|
| Dimension-level ±0.5 | 85.5% | 84.4% | 84.2% |
| Decision-level (state bucket) | 96.2% | 95.8% | 95.6% |
| Crisis W-misses (gold ≥2.5 → pred <0.5, n=37) | 4 | 3 | **5** |
| Mean W on those 37 crisis-adjacent turns (gold 2.84) | 2.39 | 2.26 | **2.01** |
| JSON validity | 100% | 100% | 100% |
| File size | 0.64 GB | 0.50 GB | 0.40 GB |
| P50 latency (M-series, Metal) | 0.50 s | 0.44 s | 0.42 s |

Head to head against Q8_0, both builds land in the same state bucket ~98% of the time
(mean |Δstress| 0.09). The headline numbers are nearly indistinguishable — which is exactly
why we looked underneath them.

**What the headline numbers hide: low-bit builds attenuate, and Q4_K_M attenuates most where
it matters least forgivingly.** Every dimension drifts downward relative to Q8_0, and the
drift concentrates on Agency (mean −0.19 for Q4_K_M, −0.11 for Q6_K). On the 37
crisis-adjacent turns of the exam — gold W ≥ 2.5 — **Q4_K_M scores lower than Q8_0 on 20 of
them and higher on exactly 1**, pulling that subset's mean withdrawal signal from 2.39 down to
2.01 and costing one additional missed crisis signal. Q6_K's withdrawal signal survives
essentially intact across the full exam (mean shift +0.01 vs Q4_K_M's −0.04). Note that
bucket agreement moves only 0.4–0.6 points across all three builds: **the state buckets are
coarse enough to absorb a damped signal, so bucket agreement alone would never have surfaced
this.** (On crisis misses specifically, 3 vs 4 out of 37 is within noise — we read Q6_K as
matching Q8_0 there, not beating it.)

**What we recommend.**

- **Q8_0** — the reference build. Use it when the read-outs gate behaviour and you have the 0.64 GB.
- **Q6_K** — the lowest build we would validate for gating use. It costs ~1 point of
  dimension-level agreement and preserves the withdrawal signal; it saves 22% of the size.
- **Q4_K_M** — fine for research, offline analysis, and any use where a human reads the scores
  rather than a system acting on them. If memory forces it into a gating deployment, lower
  your withdrawal thresholds to compensate for the documented damping, and keep deterministic
  crisis detection upstream where it belongs (LICENSE §3c requires that pattern at any
  quantization).

The other nine builds remain unvalidated by us; bit-widths below Q4_K_M should be assumed
worse until measured. The evaluation harness used here is in the
[toolkit](https://github.com/HamoAI/hamo-score-toolkit) — if you validate a build we haven't,
we would be glad to link your numbers.

## How it was trained

A three-stage distillation chain — the full story is in the companion write-up
[*Distilling hamo-score-0.6b: A Plateau, Three Bugs, and Why Data Beat Model Size*](https://www.hamoai.tech/blog/distilling-hamo-score-0-6b/):

1. **Exam by the incumbent**: 1,198 pseudonymised production turns with reference scores,
   split by session hash — 440 calibration turns and a 758-turn evaluation set, the latter
   splitting again into 305 selection and the **453-turn final** that grades this release.
   The teacher was qualified on the calibration turns; from v6.1 those same 440 turns, all
   consented staff data, also enter training. Elsewhere in this card **440 always means that
   consented set** — the qualification paper is described by its role, not its size, so the
   two uses cannot be mistaken for unrelated numbers that happen to coincide.

   *On "pseudonymised" rather than "anonymised" or "de-identified" — the weaker word is the
   honest one.* The salt is a fixed hard-coded string, so anyone holding the script can
   recompute the mapping; full timestamps are kept; message text is preserved verbatim; and
   the redaction patterns cover mainland-China formats only — Hong Kong 8-digit numbers,
   North American +1 numbers, personal names, WeChat IDs and street addresses were all
   measured passing through. **This data therefore remains personal data, and a deletion
   request still reaches it.** A separate and non-substituting fact: the payload that reaches
   the weights is only the prompt plus five scores, carrying no identifier and no timestamp.
   Both statements are true; neither one covers for the other.
2. **Affordable teacher**: `deepseek-chat` running the exact production rubric, qualified at
   **88.7% dimension-level / 97.5% decision-level** agreement before being allowed to label
   anything.
3. **Synthetic textbook**: ~20,000 admitted dialogue windows across 6 data generations
   (40+ scenario cells with per-cell label-band admission gates, style quotas for short/
   fragmented/code-switched messages, crisis and boundary contrast pairs).
   **No external-client message has ever entered training — by construction.** Starting
   with v6.1, the corpus additionally includes 440 real conversation turns contributed by
   three company-internal staff members (the founder and two staff counselors), with their
   explicit consent, upsampled ×3 (~8% of the corpus).
4. **Student**: Qwen3-0.6B, LoRA on a single MacBook (MLX; prompt-masked loss, cosine decay,
   grad-checkpointing). Total API cost of the whole project: ~US$7.

Key lessons the hard way (kept as disciplines): gradient-mask the prompt (72% of gradient was
being wasted); halve batch size when doubling sequence length (a silent fp16 explosion taught us);
verify every deploy down to a landed row.

## Limitations & known residuals

- **Chinese-primary** (zh 60% / mixed 22% / en 18% in training); English works but is less tested.
- Message-level footprint, not a person-level or relationship-level assessment.
- Mid-band calibration is coarse (0.5 grid; mid-band usage 7.4% vs reference 21–32%).
- Known residuals: a small set of highly implicit severe-distress phrasings remains hard
  (shared across all versions and the reference scorer); occasional over-scoring of bounded
  multi-step worries on E. The conversational-action gap on A was substantially closed in v6.1
  by real-conversation training data (A 81% → 85%).
- Trained against one specific rubric; scores are **relative to that rubric**, not universal
  psychological ground truth.

## Versions

| Version | Change | Dim-level | Decision-level |
|---|---|---|---|
| v2 | first distillation (7.5k synthetic) | 81% | 95.4% |
| v3.x | rebalance + defect repair | 81% | 96.8% |
| v4 | 8-agent data audit → 15k corpus, masked loss | 84% | 96.3% |
| v5 | synthetic patch cells — **rejected** (crisis-recall regression; kept as a negative result) | — | — |
| v6 | + real turns with incumbent labels — **rejected** (3 crisis-artifact rows rode into training, crisis misses 5 → 9; kept as a negative result) | — | — |
| v6.1 | + 440 consented internal-staff turns (teacher labels) | 85.6% | 96.2% |
| **v7 (this release)** | corpus re-labelled at temperature 0 (label denoising) + 2,713-row boundary-discrimination patch | 85.1% | **97.1%** |

## License

**HAMO-RAIL-S 1.0** (see [LICENSE](https://huggingface.co/HamoAI/hamo-score-0.6b/blob/main/LICENSE)): free commercial and non-commercial use,
modification and redistribution, with four use restrictions — no standalone clinical
determinations, no consequential decisions about individuals (employment / insurance /
surveillance screening), consumer mental-wellness deployments must keep independent upstream
crisis handling + AI disclosure, no re-identification. Base model Qwen3-0.6B remains Apache-2.0.

---

# 中文说明

> ### 👉 请从工具包开始，而不是从权重开始
> `pip install hamo-score` → [**hamo-score-toolkit**](https://github.com/HamoAI/hamo-score-toolkit)（Apache-2.0）
> 是这个模型的另一半：唯一正确的提示词格式、**必须置于模型上游的危机闸门**、分数该喂进去的
> 平滑折算、一条命令起的 Docker 服务器，以及 195 题自检考卷。只拿权重，恰恰会走成这个模型
> 设计上要防住的那种部署。
> **对某个评分不服？** 那是你能给我们的最有价值的东西——
> [提一条分歧报告](https://github.com/HamoAI/hamo-score-toolkit/issues/new?template=score_disagreement.md)，
> 它会直接进入引导后续版本的人类金标计划。

**hamo-score-0.6b** 是 Hamo AI 闭环疗愈引擎里第一个蒸馏进生产的组件：给心理支持对话中
来访者的每一句话「把脉」，输出五路 0–3 分的脉象（A 行动力 / W 退缩 / E 极端化 / H 敌意 /
B 边界感）。**它从不写回复，也不是危机检测器**——在 Hamo 生产系统里，危机内容在更上游被
独立的确定性机制短路，永远到不了把脉师面前；任何部署都必须复刻这个模式（见 LICENSE §3c）。

**关于 B（边界感）**：它的理论本源是家庭治疗中的「自我分化」——是「我是我、你是你」的二元
关系，还是彼此淹没的混沌一元。逐句评分器看不见关系，只看得见语言，所以 B 测的是边界感的
**语言足迹**：「我需要…」「这是我的底线」得高分；惊慌的倾泻（自我淹没在情绪里）得低分；
骂人算 H 不算 B。B 是逐句信号，不是关系诊断。

**成绩单（v7，本次发布）**：真实假名化生产对话终评（453 条未动用终评切分，金标含 8 处人工
修正；评分真值来自被替换的大模型评分器；考卷数据从未参与训练）：维度级 ±0.5 一致率 **85.1%**、
决策级（经确定性压力折算后的状态桶判定）**97.1%**、危机语漏检 **3 条**（v6.1 对应为 85.6% / 96.2% / 4 条）。
其中前两项是一致率、**越高越好**；「危机语漏检」是条数、**越低越好**——v7 由 4 条降到 3 条。
维度级那 0.5 个点的回落不是退步：v7 的语料由教师在温度 0 下全量重标（去标签噪声），学生因此
更忠于**去噪后**的标签，对当初那个含噪参照的一致率自然略降——实测其与温度 0 教师在 B 维的一致率
由 57.3% 升到 69.3%，而对原参照的 B 一致率纹丝不动。

**v7 真正修好的是边界判别**：在一份专为探测 B 决策面而造的 300 题考卷上（60 组配对 + 180 条单题，
覆盖九类陷阱格子），v6.1 会把**自我消融**读成强边界——「行，我全听你的，你说哪天去就哪天去」
真值 B=0.0，它给 2.5。那是 B 负极上的符号错误，而 B 在下游压力公式里权重最大。

| 300 题边界判别卷 | 越好方向 | v6.1 | **v7** |
|---|---|---|---|
| 造分——真值 0 却给 ≥1.0（看见并不存在的边界） | ↓ **越低越好** | 13.5% | **2.9%** |
| 漏判——真值高却给 ≤0.5（漏掉真实的边界） | ↓ **越低越好** | 16.1% | **12.8%** |
| **符号翻转**——真值 0.0 却给 ≥2.0（条数；把自我消融读成强边界） | ↓ **越低越好** | 6 条 | **0 条** |
| 配对方向正确率——高边界那一臂是否排在低的之上 | ↑ 越高越好 | 81.7% | **98.3%** |

前三行是错误率、第四行是正确率，所以**前三行越低越好、最后一行越高越好**；v7 四项全部改善。

两列均测于随包发布的 q8 GGUF + 中性采样，描述的是你下载到的产物本身。

参照系——同一批消息让原评分器自己打两遍，维度级自洽也只有 94–98%。单条延迟：M1 Pro 约
0.8 秒；2 vCPU ARM 服务器（纯 CPU，q8 GGUF）1.5–2.9 秒。

**残余重叠：我们量了，没有假设掉。** 考卷与训练集按样本、按会话完全不相交，但授权供数的内部
员工会在不同会话里重复说同样的话：终评 453 题中有 42 题（9.3%）的正文在那 440 条授权训练数据
里出现过，其中 11 题（2.4%）连最近上下文也相同。分开判卷的结果——这 42 题维度级 89.0%、
决策级 100%；其余 411 题无任何正文重叠，**维度级 85.2%、决策级 95.9%**。也就是说，上面两个
成绩含约 **+0.4 / +0.3 个百分点**的乐观。幅度不大，但我们选择自己写出来。（42 题里只有 5 题
连标签也相同——同一句话换个轮次通常拿到不同分数，所以它们不是白送的分，只是更容易的分。）

**训练方式**：三级师徒链——生产历史评分出考卷（1,198 条假名化真题——用「假名化」而非「匿名化」
是因为弱的那个词才是诚实的：盐是硬编码固定字符串、完整时间戳保留、正文逐字保留，且脱敏正则只覆盖
大陆格式，香港 8 位号码、北美 +1 号码、人名、微信号与住址实测全部穿过，**故这批数据仍属个人信息，
删除权仍及于它**；另一件必须单独陈述、不可用来顶替上一条的事实是：进入权重的载荷只有提示词与五个
分数，不含任何标识符与时间戳。按会话哈希切分为 440 条
校准集与 758 条评测集，后者再切成 305 条选型集与判定本次成绩的 453 条终评集）→ DeepSeek 在
校准集上通过资格考（维度级 88.7% / 决策级 97.5%）后当教师；本卡中「440」始终指那批经授权的
内部员工轮次，资格考卷按用途称呼、不按题量称呼，以免两处用法被误读成两个撞车的数字 → 约 2 万段合成对话当教材（40+ 场景格子、逐格标签准入闸门、短句/
碎片/中英混杂风格配额）→ Qwen3-0.6B 学生在一台 MacBook 上 LoRA 学成。全项目 API 成本
约 7 美元。**训练语料从不包含任何外部来访者消息（构造上保证）；自 v6.1 起额外加入 440 条公司内部员工（创始人与两位咨询师）明示授权的真实对话轮次（×3 上采样，约占语料 8%）。**

**社区量化档位（我们实测过其中一档）**：社区志愿者 mradermacher 制作了
[**Q2_K→f16 共 12 个静态 GGUF 量化档**](https://huggingface.co/mradermacher/hamo-score-0.6b-GGUF)
——感谢他的工作，也感谢他在再分发中完整保留了 RAIL-S 许可条款。我们公布的指标测于 **Q8_0**；
由于这个模型的读数要门控对话能走多深，低比特量化掉了多少不是困惑度数字而是临床问题，所以我们
用同一套 453 题终评、同一段提示词、同一个解析器，实测了社区的 **Q6_K** 与 **Q4_K_M**：

| | Q8_0（我们的基线） | Q6_K（社区） | Q4_K_M（社区） |
|---|---|---|---|
| 维度级 ±0.5 一致率 | 85.5% | 84.4% | 84.2% |
| 决策级（状态桶） | 96.2% | 95.8% | 95.6% |
| 危机 W 漏检（金标 ≥2.5 → 预测 <0.5，n=37） | 4 | 3 | **5** |
| 那 37 条危机相邻样本的 W 均值（金标 2.84） | 2.39 | 2.26 | **2.01** |
| JSON 合法率 | 100% | 100% | 100% |
| 体积 | 0.64 GB | 0.50 GB | 0.40 GB |
| P50 延迟（M 系列，Metal） | 0.50 秒 | 0.44 秒 | 0.42 秒 |

与 Q8_0 直接对比，两个社区档位落在同一状态桶的比例都约 98%（压力更新偏差均值 0.09）。总分
几乎分不出高下——正因如此，我们才往下面看了一层。

**总分掩盖了什么：低比特档位会「衰减」信号，而 Q4_K_M 恰恰在最不容有失的地方衰减最厉害。**
所有维度相对 Q8_0 系统性下移，且集中在行动力上（Q4_K_M 均值 −0.19，Q6_K −0.11）。在终评集
的 37 条危机相邻样本（金标 W≥2.5）上，**Q4_K_M 有 20 条打得比 Q8_0 更低、仅 1 条更高**，把
该子集的退缩信号均值从 2.39 拉到 2.01，并因此多漏检一条危机信号；而 Q6_K 的退缩信号在全卷上
基本无损（均值偏移 +0.01，Q4_K_M 为 −0.04）。请注意三个档位的状态桶一致率只相差 0.4–0.6 个
百分点：**桶的边界粗到足以吸收一个被压扁的信号，所以只看桶一致率永远发现不了这件事。**
（就危机漏检本身而言，37 条里的 3 与 4 属于噪声范围——我们把 Q6_K 读作「与 Q8_0 持平」，
而不是「优于」。）

**我们的建议**：

- **Q8_0** —— 参考档。读数用于门控行为、且你付得起 0.64 GB 时，用它。
- **Q6_K** —— 我们愿意为门控用途背书的最低档。代价是约 1 个百分点的维度级一致率，退缩信号
  保持完好，体积省 22%。
- **Q4_K_M** —— 适合研究、离线分析，以及分数由人来读而不是由系统据以行动的场景。若内存迫使
  它进入门控部署，请相应下调退缩维度的阈值以补偿上述衰减，并把确定性危机检测保持在上游
  （无论用哪个量化档，LICENSE §3c 都要求这个模式）。

其余九个档位我们未做验证，比 Q4_K_M 更低的比特应默认更差，直到有人量过。评测脚本在
[工具包](https://github.com/HamoAI/hamo-score-toolkit)里——如果你验证了我们没验证过的档位，
我们很乐意把你的数据链上来。

**官方工具包（已开源到 GitHub）**：[hamo-score-toolkit](https://github.com/HamoAI/hamo-score-toolkit)
（Apache-2.0）是这个模型的「另一半」——一条 `pip install` 装上唯一正确的提示词格式、容错
解析、参考版压力折算与许可证要求的危机闸门；一条 `docker compose up` 跑起参考服务器
（`POST /score` 走完整的 闸门→评分→平滑→状态桶 管线）；一份 195 题合成自检考卷 + 10 条
手写闸门用例，对照官方参考带（JSON 100%、维度级 84.0%、闸门 10/10）验证你的部署接线；
还有一份微调指南（[docs/finetune.md](https://github.com/HamoAI/hamo-score-toolkit/blob/main/docs/finetune.md)，
六代打法，含两代拒收的完整原因）。发布文：《[开源 hamo-score-toolkit：把模型的另一半也交出去](https://www.hamo.ai/blog/open-sourcing-hamo-score-toolkit/)》。

**许可证**：HAMO-RAIL-S 1.0——自由商用与修改，但有四条使用限制：不得独立做临床判定、
不得用于对个人的重大决定（雇佣/保险/监控筛查）、面向消费者的心理健康部署必须保留独立的
上游危机处理与 AI 身份披露、不得试图重识别个人。

配套阅读（背景与方法论）：《[hamo-score-0.6b 是怎么蒸出来的：一段平台期、三个坑，以及数据为什么赢了参数量](https://www.hamo.ai/blog/distilling-hamo-score-0-6b/)》。