👉 Start with the toolkit, not the weights

pip install hamo-score → 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; 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 component distilled from the production scoring of 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 measured on the v7 bf16 weights at temperature 0 (JSON 100% · dimension-level 83.5%, expected band 81–86% · gate 10/10; the shipped q8 GGUF with neutral sampling scores 83.8%) so you can verify your wiring reproduces the official numbers;

Fine-tuning guide — docs/finetune.md, the seven-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 · 中文.

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/. 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 (this release's shipped v7 q8 GGUF), sampling as the only variable: fabrication rate 2.9% → 8.7%, boundary sign-flips (a true 0.0 scored ≥2.0) 0 → 1; the miss rate falls from 12.8% to 6.7%, which is the same upward push, not an improvement. The same switch on the earlier v6.1 q8 GGUF: fabrication 13.5% → 25.0%, sign-flips 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 is designed to replace; 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 the 453-turn final split of the real-conversation exam (never trained on; gold labels include 8 human corrections). Routine checkpoint selection used the separate 305-turn selection split, but in the last step of choosing between v7's two finalist checkpoints we also looked at their final-split scores — they tied at 97.1% decision-level (85.0% vs 85.1% dimension-level), and the tie was broken on boundary-exam error type, not on this split.

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.

Each column measured on that release's 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 and again in v7 (final-exam misses 11 → 5 → 4 → 3) — but see the crisis disclaimer above: recall here is defense-in-depth, not the defense.

Community quantizations — and what we measured on v7

mradermacher/hamo-score-0.6b-GGUF provides static GGUF quants (Q2_K to f16) of the v4 weights this repository carried on 2026-08-05 — two releases before v7 — twelve build targets we never shipped ourselves. Thanks to mradermacher for the work, and for carrying the RAIL-S license terms through redistribution. Those builds are now outdated: anyone who wants a v7 Q6_K or Q4_K_M currently has to quantize this repository's v7 weights themselves (llama.cpp: convert_hf_to_gguf.py to f16, then llama-quantize).

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 measured it on v7 itself. Every column below comes from the v7 weights, run through the same 453-turn final exam with the same prompt and parser: bf16 via MLX (this column reproduces the headline figures above), the GGUF builds via llama.cpp with neutral sampling. Q8_0 is the GGUF shipped in this repository; Q6_K and Q4_K_M are our own quantizations of v7 (a v7 f16 GGUF from llama.cpp's convert_hf_to_gguf.py, then llama.cpp's llama_model_quantize — the library function the llama-quantize command above wraps, so the same path) and are not published anywhere.

v7 weights, 453-turn final split	bf16 (MLX)	Q8_0 (shipped)	Q6_K (ours, unpublished)	Q4_K_M (ours, unpublished)
Dimension-level ±0.5	85.1%	85.3%	85.3%	84.5%
Decision-level (state bucket)	97.1%	97.1%	96.9%	96.7%
Crisis W-misses (gold ≥2.5 → pred <0.5, n=37)	3	3	3	3
Mean W on those 37 crisis-adjacent turns (gold 2.84)	2.58	2.58	2.58	2.50
On those 37 turns, W lower / higher than Q8_0	0 / 0	—	1 / 1	6 / 1
Mean score shift vs Q8_0, all 453 turns	≈0	—	every dimension within ±0.011	A −0.05, W −0.03, E −0.01, H +0.00, B +0.00
File size	—	0.64 GB	0.50 GB	0.40 GB
P50 latency (Apple M1 Pro; llama.cpp on Metal unless noted)	0.83 s (MLX)	0.70 s	0.56 s	0.54 s

What the table says. On v7, Q6_K is indistinguishable from Q8_0 on this exam. Against Q8_0, Q4_K_M loses 0.8 pt at dimension level and 0.4 pt at decision level, and did not add a crisis miss here — but it still attenuates one-sidedly where it matters least forgivingly: on the 37 crisis-adjacent turns (gold W ≥ 2.5) it scores W lower than Q8_0 on 6 and higher on 1, pulling that subset's mean withdrawal signal from 2.58 to 2.50, and across all 453 turns its mean shift is small (A −0.05, W −0.03, the rest within ±0.01). The methodological point still stands on v7: bucket agreement moves only 0.4 points across the three GGUF builds — the state buckets are coarse enough to absorb a damped signal, so bucket agreement alone would never have surfaced this. That is why the toolkit's compare_quants.py prints a directional table, not just agreement.

An earlier version of this section compared mradermacher's v4 community builds against our v6.1 Q8_0 and so overstated quantization loss — its "Q4_K_M lower than Q8_0 on 20 of 37", "Q6_K ~1 point below" and "Agency drift of −0.19" findings were mostly the v4 → v6.1 version gap, and the all-v7 table above supersedes them.

What we recommend.

Q8_0 — the reference build, shipped in this repository. Use it when the read-outs gate behaviour and you have the 0.64 GB.
Q6_K, quantized from v7 — validated for gating use when memory is tight: indistinguishable from Q8_0 on the final exam, including the crisis-adjacent subset, and it saves 22% of the size. (The community Q6_K is built from v4 and is not covered by this validation.)
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, keep deterministic crisis detection upstream where it belongs (LICENSE §3c requires that pattern at any quantization), and consider lowering your withdrawal thresholds to compensate for the one-sided W attenuation above.

Builds below Q4_K_M remain unvalidated by us and should be assumed worse until measured; mradermacher's v4 builds are outdated and not covered by these recommendations. The toolkit's eval/compare_quants.py runs the same comparison, including the directional table, on the public synthetic exam. Its absolute numbers will differ from the 453-turn table above, but the shape transfers. 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:

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.
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.
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).
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): 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（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，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从生产评分中蒸馏出来的组件：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 条终评切分，金标含 8 处人工修正；常规选点用的是另一份 305 条选型集，但 v7 最后两个候选检查点二选一时也看过它们在终评切分上的成绩—— 二者决策级同为 97.1%、维度级 85.0% 对 85.1%，最终按边界判别卷的错误性质定夺，而非按终评分数；评分真值来自它要替代的大模型评分器；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。

两列各测于该版本随包发布的 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（MLX bf16）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。v7 分开判卷的结果——这 42 题维度级 88.1%、决策级 100%；其余 411 题无任何正文重叠，维度级 84.8%、决策级 96.8%。也就是说，上面两个成绩各含约 +0.3 个百分点的乐观。幅度不大，但我们选择自己写出来。（v6.1 用同样方法测得：重叠部分 89.0% / 100%，无重叠部分 85.2% / 95.9%。）（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，以及我们在 v7 上实测的量化：社区志愿者 mradermacher 制作了 Q2_K→f16 共 12 个静态 GGUF 量化档 （量化自 2026-08-05 时本仓库的 v4 权重，比 v7 早两个版本）——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。 这些档位现已过时：目前想要 v7 的 Q6_K / Q4_K_M，需要自己从本仓库的 v7 权重量化（llama.cpp：先用 convert_hf_to_gguf.py 转成 f16，再用 llama-quantize）。

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，所以我们直接在 v7 上量了：下表每一列都来自 v7 权重，用同一套 453 题终评、同一段提示词、同一个解析器——bf16 走 MLX（这一列复现了上文成绩单），各 GGUF 档走 llama.cpp + 中性采样。Q8_0 是本仓库随包发布的 GGUF； Q6_K 与 Q4_K_M 是我们自己从 v7 量化的（先用 llama.cpp 的 convert_hf_to_gguf.py 得到 v7 f16 GGUF，再经 llama.cpp 的 llama_model_quantize 量化——即上文 llama-quantize 命令所封装的库函数，路径相同），未在任何地方发布：

v7 权重，453 题终评	bf16（MLX）	Q8_0（随包发布）	Q6_K（自量化，未发布）	Q4_K_M（自量化，未发布）
维度级 ±0.5 一致率	85.1%	85.3%	85.3%	84.5%
决策级（状态桶）	97.1%	97.1%	96.9%	96.7%
危机 W 漏检（金标 ≥2.5 → 预测 <0.5，n=37）	3	3	3	3
那 37 条危机相邻样本的 W 均值（金标 2.84）	2.58	2.58	2.58	2.50
这 37 条中 W 低于 / 高于 Q8_0 的条数	0 / 0	—	1 / 1	6 / 1
全部 453 条相对 Q8_0 的平均分偏移	≈0	—	各维度均在 ±0.011 以内	A −0.05、W −0.03、E −0.01、H +0.00、B +0.00
体积	—	0.64 GB	0.50 GB	0.40 GB
P50 延迟（Apple M1 Pro；除注明外为 llama.cpp + Metal）	0.83 秒（MLX）	0.70 秒	0.56 秒	0.54 秒

表里说了什么：在 v7 上，Q6_K 与 Q8_0 在这份考卷上分不出差别。Q4_K_M 相对 Q8_0 维度级少 0.8 个百分点、决策级少 0.4 个百分点，这次也没有多漏检危机——但它仍在最不容有失的地方单向衰减：在 37 条危机相邻样本（金标 W≥2.5）上，有 6 条 W 打得比 Q8_0 更低、仅 1 条更高，把该子集的退缩信号均值从 2.58 拉到 2.50；全卷 453 条上的平均偏移很小（A −0.05、W −0.03，其余在 ±0.01 以内）。方法论上的那一点在 v7 上依然成立：三个 GGUF 档位的状态桶一致率只相差 0.4 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。 这正是工具包的 compare_quants.py 除了一致率还要打印方向性表格的原因。

本节的早先版本拿 mradermacher 的 v4 社区档位去对比我们 v6.1 的 Q8_0，因而夸大了量化损失——其中「Q4_K_M 在 37 条里有 20 条低于 Q8_0」「Q6_K 低约 1 个百分点」「行动力下移 −0.19」等结论主要是 v4 → v6.1 的版本差，现已由上面这张全 v7 的表取代。

我们的建议：

Q8_0 —— 参考档，随本仓库发布。读数用于门控行为、且你付得起 0.64 GB 时，用它。
Q6_K（从 v7 量化） —— 内存紧张时可用于门控，已验证：在终评上（含危机相邻子集）与 Q8_0 分不出差别，体积省 22%。（社区的 Q6_K 量化自 v4，不在这项验证范围内。）
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，务必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），并考虑相应下调退缩维度的阈值，以补偿上面那种单向衰减。

比 Q4_K_M 更低的档位我们未做验证，应默认更差，直到有人量过；mradermacher 的 v4 档位已过时，不在上述建议范围内。工具包里的 eval/compare_quants.py 在公开的合成考卷上做同样的对比（含方向性表格）——绝对数值与上面 453 题的表不可直接比，但对比的形状可迁移；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 （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（v7 bf16 权重、温度 0 实测：JSON 100%、维度级 83.5%（合格带 81–86%）、闸门 10/10；随包 q8 GGUF 在中性采样下为 83.8%）验证你的部署接线；还有一份微调指南（docs/finetune.md，七代打法，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》。

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》。