Skip to page content
SAVRN
Search Plan a project

Open-weight model · Zero-shot classification

chinese-decision-2b

by Quincyqiang quincyqiang/chinese-decision-2b

chinese-decision-2b is an open-weight model for zero-shot classification from Quincyqiang, released under Apache License 2.0. It has 1.9B parameters and a 262,144-token context. At 16-bit it needs about 4.5 GB of GPU memory, which fits on 1x MI300X from $1.85 an hour, at the lowest prices in the SAVRN Index.

Chinese Decision 是一个面向中文场景的轻量级决策模型。它不聊天、不生成文本:给定上下文和一组固定候选项,一次前向直接返回每个候选项的概率,适合意图识别、工单分流、内容审核、数据打标、给模型输出打分、智能体动作选择等"有限选项判断"任务。 - 一次前向,零输出 token:读取最后位置上候选字母 A–Z 的 logits,softmax 后即为概率。 - 三种题型:noul(是/否)、choice(单选)、score(有序等级,返回期望分值)。 - 系列中准确率最高:CJ-Bench 全量整体…

Parameters1.9B
Context262,144
Weights3.8 GB
Licenseapache-2.0
AccessOpen weights
Monthly Downloads—

Runs On

What it takes to serve chinese-decision-2b (1.9B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.

PrecisionWeightsMemory neededCheapest setupPer hourAlso fits
16-bit 3.8 GB 4.5 GB 1x MI300X (192 GB)
Vultr
$1.85 1x H100 $1.99 · 1x MI325X $2.00
8-bit 1.9 GB 2.3 GB 1x MI300X (192 GB)
Vultr
$1.85 1x H100 $1.99 · 1x MI325X $2.00
4-bit 0.9 GB 1.1 GB 1x MI300X (192 GB)
Vultr
$1.85 1x H100 $1.99 · 1x MI325X $2.00

Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Oct 20, 2026.

chinese-decision-2b on every accelerator the SAVRN Index prices, at every precision

Model Card

By Quincyqiang, published under apache-2.0, revision 4edf2b416603.

Chinese Decision 是一个面向中文场景的轻量级决策模型。它不聊天、不生成文本:给定上下文和一组固定候选项,一次前向直接返回每个候选项的概率,适合意图识别、工单分流、内容审核、数据打标、给模型输出打分、智能体动作选择等"有限选项判断"任务。 - 一次前向,零输出 token:读取最后位置上候选字母 A–Z 的 logits,softmax 后即为概率。 - 三种题型:noul(是/否)、choice(单选)、score(有序等级,返回期望分值)。 - 系列中准确率最高:CJ-Bench 全量整体 ACC 84.11%,医疗 / 法律 / 金融平均 78.25%,比 chinese-decision-0.8b 分别高 1.01 / 1.35 个百分点。 - 概率经过校准:附带在独立校准划分上拟合的温度,整体 ECE 0.64%。 更新于 2026-10-17。BF16 合并模型已完成全量 CJ-Bench、保存/重载和实际 HTTP 三题型/双路由验证,并已在 Hugging Face 与 ModelScope 公开发布。由本模型直接转换的 Q80 GGUF 也已完成全量 CJ-Bench 对照,见 2B Q80 模型卡。 训练数据版本 a3963a61f71e00051a80721e587d720d8926e8e7。最终 adapter SHA256 为 9b55e8eda00997544b76fa945d890605c3acae4199041c197ef5a09275142a7c;合并权重 SHA256 为…

Read Quincyqiang's full model card

中文概率决策模型:一次前向,直接返回候选项概率。

A Chinese decision model fine-tuned from Qwen3.5-2B with LoRA and CE + RLCD. It supports yes/no, single-choice and ordinal decisions across general, medical, legal and finance tasks. This package contains merged BF16 weights and the existing validation-fitted temperature.

代码 · 模型总览 · Decision Studio · 训练数据 · 测试集

0.8B BF16 · 0.8B Q8_0 · 2B BF16 · 2B Q8_0

相关资源

资源 说明 链接
代码 训练、评测、GGUF 导出与 HTTP 服务(GPU / CPU) GitHub
测试平台 Decision Studio:决策实验台、场景库、鲁棒性实验室、评测中心与性能压测 GitHub
2B BF16(本仓库) 合并权重与校准温度,Transformers 推理 Hugging Face · ModelScope
2B GGUF Q8_0 llama.cpp 推理,CPU 即可部署,不需要 PyTorch Hugging Face · ModelScope
0.8B 系列 更小、更快;部分法律判断与金融评分任务更好 BF16 · GGUF

简介

Chinese Decision 是一个面向中文场景的轻量级决策模型。它不聊天、不生成文本:给定上下文和一组固定候选项,一次前向直接返回每个候选项的概率,适合意图识别、工单分流、内容审核、数据打标、给模型输出打分、智能体动作选择等"有限选项判断"任务。

  • 一次前向,零输出 token:读取最后位置上候选字母 A–Z 的 logits,softmax 后即为概率。
  • 三种题型:noul(是/否)、choice(单选)、score(有序等级,返回期望分值)。
  • 单一模型覆盖四个领域:通用、医疗、法律、金融混合训练。
  • 系列中准确率最高:CJ-Bench 全量整体 ACC 84.11%,医疗 / 法律 / 金融平均 78.25%,比 chinese-decision-0.8b 分别高 1.01 / 1.35 个百分点。
  • 概率经过校准:附带在独立校准划分上拟合的温度,整体 ECE 0.64%。

模型与交付状态

更新于 2026-10-17。BF16 合并模型已完成全量 CJ-Bench、保存/重载和实际 HTTP 三题型/双路由验证,并已在 Hugging Face 与 ModelScope 公开发布。由本模型直接转换的 Q8_0 GGUF 也已完成全量 CJ-Bench 对照,见 2B Q8_0 模型卡。

项目 值
基础模型 Qwen3.5-2B,文本决策主干
训练目标 候选分布交叉熵 + RLCD;LoRA rank 32
训练覆盖 1 epoch,13,048,000 条决策,4,994 步,4 卡
best 验证集 distribution NLL 选择 step 4994,NLL 0.5860035124
RLCD group size 4,sigma 按已训练样本进度从 0.4 降至 0.1
交付权重 model.safetensors,BF16,3,763,692,048 字节
默认温度 1.074872374534607,独立验证划分拟合
输入 / 候选项 最长 16,384 tokens / 单题 2–26 项(训练覆盖 2–16 项),类别更多用 /v1/tournament 分组淘汰
提示版本 hmm-zh-v1,关闭 thinking,末位候选字母 logits
训练结束 2026-10-17 17:25:54,Asia/Shanghai

训练数据版本 a3963a61f71e00051a80721e587d720d8926e8e7。最终 adapter SHA256 为 9b55e8eda00997544b76fa945d890605c3acae4199041c197ef5a09275142a7c;合并权重 SHA256 为 a8fadd326b19a278f8c26216ca481b607369baca33962aac0d8fda603c830068。

文件说明

文件 内容
model.safetensors、config.json 合并模型,无需再次加载 PEFT adapter
tokenizer.json、tokenizer_config.json、chat_template.jinja tokenizer 与聊天模板
decision_config.json 模型身份、温度、最大长度与候选数
training_config.json、training_summary.json 配置、训练覆盖、best 的来源
validation_report.json 选模验证集的校准前后指标,不是测试集结果
LICENSE、LICENSES/、THIRD_PARTY_NOTICES.md 许可与上游方法归属

快速开始

下载

hf download quincyqiang/chinese-decision-2b --local-dir models/chinese-decision-2b
# 或 ModelScope
modelscope download --model quincyqiang/chinese-decision-2b --local_dir models/chinese-decision-2b

GPU HTTP 服务

在 代码仓库 根目录安装推理依赖,模型目录放在 models/chinese-decision-2b:

pip install -e ".[torch]"
CUDA_VISIBLE_DEVICES=0 OMP_NUM_THREADS=6 TOKENIZERS_PARALLELISM=false \
  python scripts/serve_gpu.py --base models/chinese-decision-2b \
  --host 127.0.0.1 --port 8789 --token-budget 16384 --max-batch 32

启动前选择空闲 GPU;本命令不需要访问训练 run 或加载 adapter。测试服务:

curl http://127.0.0.1:8789/health
curl http://127.0.0.1:8789/v1/systemone -H 'Content-Type: application/json' --data-binary @examples/request.json
python scripts/verify_export.py --url http://127.0.0.1:8789

Python 直接读取候选概率

以下示例仅需 PyTorch 和 Transformers,不需要训练库或文本生成。项目实测环境为 torch 2.8.0、transformers 5.3 系列;安装约束见 pyproject.toml。固定候选顺序并读取 decision_config.json 的温度,不使用通用 text-generation pipeline。

import json

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

path = "models/chinese-decision-2b"
tokenizer = AutoTokenizer.from_pretrained(path)
model = AutoModelForCausalLM.from_pretrained(path, dtype=torch.bfloat16, attn_implementation="sdpa").to("cuda").eval()
config = json.load(open(f"{path}/decision_config.json", encoding="utf-8"))
temperature, max_length = config["temperature"], config["max_length"]

LETTERS = "ABCDEFGHIJKLMNOPQRSTUVWXYZ"
letter_ids = [tokenizer.encode(c, add_special_tokens=False)[0] for c in LETTERS]


def text(value):
    return value if isinstance(value, str) else json.dumps(value, ensure_ascii=False, separators=(",", ":"))


def options_for(question):
    kind, criteria = question["type"], question.get("criteria")
    if kind == "choice":
        return [(str(k), text(v)) for k, v in criteria.items()]
    if kind == "score":
        return [(str(i), text(v)) for i, v in enumerate(criteria)]
    criteria = criteria or {}
    return [("false", text(criteria.get("false", "否"))), ("true", text(criteria.get("true", "是")))]


def build_prompt(state, instructions, options):
    content = ("待判断的上下文:\n" + text(state)
               + "\n\n问题:\n" + text(instructions)
               + "\n\n候选项:\n" + "\n".join(f"{LETTERS[i]}: {k} — {v}" for i, (k, v) in enumerate(options))
               + "\n只返回候选项对应的字母。")
    return tokenizer.apply_chat_template([{"role": "user", "content": content}], tokenize=False,
                                         add_generation_prompt=True, enable_thinking=False)


@torch.inference_mode()
def decide(state, question):
    options = options_for(question)
    ids = tokenizer.encode(build_prompt(state, question["instructions"], options), add_special_tokens=False)
    if len(ids) > max_length:
        raise ValueError(f"input is {len(ids)} tokens; the limit is {max_length}")
    logits = model(torch.tensor([ids], device=model.device), logits_to_keep=1, use_cache=False).logits
    probs = (logits[0, -1, letter_ids[:len(options)]].float() / temperature).softmax(-1).tolist()
    k, best = len(probs), max(range(len(probs)), key=probs.__getitem__)
    answer = {"probabilities": {key: round(p, 4) for (key, _), p in zip(options, probs)},
              "confidence": round((k * probs[best] - 1) / (k - 1), 4)}
    if question["type"] == "noul":
        answer["noul"] = round(probs[1], 4)
    elif question["type"] == "choice":
        answer["choice"] = options[best][0]
    else:
        answer["score"] = round(sum(i * p for i, p in enumerate(probs)), 4)
    return answer


state = "客户留言:付款后一直没有收到发票,我已经催了三次,今天必须拿到。"
questions = {
    "urgent": {"type": "noul", "instructions": "客户是否表达了紧迫性?"},
    "department": {"type": "choice", "instructions": "应该由哪个部门处理?",
                   "criteria": {"billing": "支付、发票及退款", "technical": "故障和系统问题", "sales": "购买咨询与升级"}},
    "frustration": {"type": "score", "instructions": "客户有多不满?", "criteria": ["平静", "不满", "非常愤怒"]},
}
for name, question in questions.items():
    print(name, decide(state, question))
接口 用途
POST /v1/systemone 决策;可追加 ?debug=1 查看候选 logits
POST /api/alpha/decisions 兼容 hmm 的同一决策路由
POST /v1/batch 多请求批处理,单个请求失败不影响其他请求
POST /v1/compile 编译输入并计数 tokens
POST /v1/tournament 类别更多的 choice 问题:分组淘汰后决赛,每组不超过 16 个候选
GET /health、GET /v1/info 模型身份、step、温度及运行信息

每个请求最多 64 个问题,每题 2–26 个候选(类别更多时用 /v1/tournament,示例 examples/tournament.json 有 41 个客服意图),最长 16,384 tokens;超长或不合法输入返回 422,不静默截断。score 输出从 0 开始的等级期望值,不是最大概率等级;confidence = (K*p_max-1)/(K-1) 衡量分布尖锐程度,不等于经验准确率。

CJ-Bench 全量结果

CJ-Bench 全量为 307,900 条(304,822 条硬标签、3,078 条软标签),通用/医疗/法律/金融分别为 100,000 / 200,000 / 4,300 / 3,600 条。全部保留原始候选顺序,无过滤、无截断。ACC/ECE 仅计硬标签,ECE 使用 15 个等宽分箱;分布 NLL 与 Brier 包含软标签;Score MAE 覆盖 38,746 条评分题,计算预测期望等级与目标期望等级的绝对差。

整体 micro 按硬标签样本数加权,四领域 macro 是四个领域的等权平均,专业平均仅对医疗、法律、金融等权。不同量不能互换,整体 ECE 也不是领域 ECE 的平均。选模和温度拟合只使用验证数据,未使用 CJ-Bench 标签。

下表均为已发布的正式权重(LoRA 已合并,使用默认温度);本仓库为 chinese-decision-2b,chinese-decision-2b-gguf 由它直接转换。加粗表示 2B 相对同格式 0.8B 提升的指标。

关于温度 T:概率按 p = softmax(z / T) 计算,z 为候选字母的 logits。T = 1 表示直接对 logits 做 softmax、不做温度校准。发布权重附带在独立验证划分上拟合的温度(0.8B 为 1.108,2B 为 1.075),只调整概率的尖锐程度,不改变选出的答案:准确率不变,ECE / NLL 更好。

模型 格式 Micro ACC (%) ↑ Macro ACC (%) ↑ ECE (%) ↓ NLL ↓ Brier ↓ Score MAE ↓
chinese-decision-0.8b BF16 83.1052 75.7376 0.2706 0.431318 0.223833 0.584427
chinese-decision-0.8b-gguf Q8_0 83.1233 75.8092 0.2696 0.431293 0.223815 0.583973
chinese-decision-2b BF16 84.1107 77.1750 0.6361 0.416520 0.213295 0.564216
chinese-decision-2b-gguf Q8_0 84.1094 77.3282 0.6534 0.416288 0.213189 0.563738

相对 chinese-decision-0.8b,本模型 Micro +1.01、Macro +1.44、专业平均 +1.35 个百分点,NLL / Brier / Score MAE 均更低;整体 ECE 略高(0.64% vs 0.27%)。训练阶段表中的 checkpoint 为合并前、T = 1 的结果,与发布权重相差在合并的数值误差范围内(同批次验证集审计为 14/1200 条 argmax 变化)。

分领域与分题型

模型 格式 General ACC (%) Medical ACC (%) Legal ACC (%) Finance ACC (%) Professional avg. ACC (%)
chinese-decision-0.8b BF16 72.26 88.85 67.84 74.00 76.90
chinese-decision-0.8b-gguf Q8_0 72.23 88.89 67.84 74.28 77.00
chinese-decision-2b BF16 73.96 89.50 68.74 76.50 78.25
chinese-decision-2b-gguf Q8_0 73.91 89.50 69.40 76.50 78.47
题型 样本数 chinese-decision-0.8b ACC (%) chinese-decision-2b ACC (%) 2B ECE (%) 2B NLL 2B Score MAE
noul 134,577 94.96 95.11 0.34 0.114612 —
choice 134,577 78.32 80.12 0.90 0.559377 —
score 38,746 56.45 57.67 1.65 0.968959 0.564216

训练阶段

每 1000 步按验证集分布 NLL 选出的 checkpoint,均在全量 307,900 条上评测(合并前,T = 1,即不做温度校准)。进度按实际已训练样本计算;最后一行为发布权重。

Checkpoint 训练进度 General ACC (%) Medical ACC (%) Legal ACC (%) Finance ACC (%) Professional avg. ACC (%) ECE (%)
step 1000 20.08% 70.55 87.69 55.67 66.94 70.10 2.03
step 2000 40.13% 72.56 88.54 62.67 71.44 74.22 1.27
step 3000 60.11% 71.87 88.99 65.56 74.31 76.28 1.63
step 4000 80.18% 73.77 89.42 68.40 76.03 77.95 0.85
step 4994(最终) 100.00% 73.95 89.50 69.05 76.03 78.19 1.49
chinese-decision-2b(发布权重) — 73.96 89.50 68.74 76.50 78.25 0.64

法律、金融提升最多(+13.37 / +9.08 个百分点)。step3000 通用领域的回落来自商品分类来源的暂时崩落,step4000 恢复;后段准确率趋于平台。完整诊断见 训练交付记录。

与其他决策模型对比

下表的外部模型数字为 Chinese-Jev 原文 Table 1 / Table 2 的作者报告值,2026-10-17 核对,本项目没有重跑这些模型;本项目行为发布权重。加粗为每列最优。

通用领域(Table 1)

模型 ACC (%) ↑ ECE (%) ↓
Qwen3.5-2B 32.95 43.68
Open-Jev 52.33 10.54
SemIF 44.21 27.27
Laya Multilingual 40.22 23.04
Jev (API) 68.35 11.45
Chinese-Jev General / Specialists 69.20 3.78
chinese-decision-0.8b 72.26 1.68
chinese-decision-2b 73.96 1.93

专业领域(Table 2)

模型 Medical ACC Medical ECE Legal ACC Legal ECE Finance ACC Finance ECE Avg. ACC Avg. ECE
Qwen3.5-2B 53.88 28.71 31.53 44.82 38.11 49.12 41.17 40.88
Open-Jev 73.08 2.88 44.07 23.15 48.11 10.19 55.09 12.07
SemIF 62.17 10.39 43.07 19.99 47.89 18.45 51.04 16.28
Laya Multilingual 24.55 43.05 34.72 31.49 38.31 26.18 32.53 33.57
Jev (API) 84.29 5.70 68.35 8.03 78.39 2.98 77.01 5.57
Chinese-Jev General 39.77 17.33 38.76 15.37 35.39 14.43 37.97 15.71
Chinese-Jev General / Specialists 87.65 7.25 60.10 17.45 64.72 11.55 70.82 12.08
chinese-decision-0.8b 88.85 1.12 67.84 4.05 74.00 2.45 76.90 2.54
chinese-decision-2b 89.50 0.24 68.74 3.37 76.50 2.13 78.25 1.92

本模型相对 Chinese-Jev(General + 三个 Specialist)在通用 / 医疗 / 法律 / 金融分别高 4.76 / 1.85 / 8.64 / 11.78 个百分点,专业平均高 7.43;相对闭源 Jev,专业平均高 1.24,金融仍低 1.89。四个领域的 ECE 都低于论文中的所有模型。Chinese-Jev General 加上医疗、法律和金融 Specialist 是 四个分别模型;本项目每个尺寸各使用一个模型覆盖四领域。Jev 为闭源 API。论文使用 H200,本地使用 A800;延迟口径和硬件不同,不做等价排名。

2B 与 0.8B 的模型规模、batch 和训练目标同时变化,因此结果不能当作 RLCD 的独立收益。两者 best 已见样本分别为 13,048,000 与 12,879,232,应按样本量比较进度。

Q8_0 版本

由本模型直接转换的 2B Q8_0 GGUF(2.01 GB,llama.cpp,CPU 可运行,不需要 PyTorch)沿用同一温度,在 CJ-Bench 全部 307,900 条上与本模型逐条配对:整体 ACC 84.1094%(本模型 84.1107%,净少对 4 条),首选项一致率 99.00%,ECE 0.6534%(本模型 0.6361%),NLL、Brier、Score MAE 略降。四领域 macro 77.3282%、专业平均 78.4663%,高出的部分主要来自法律领域(4,300 条中净多对 28 条),不代表量化提升了能力。

全量在 llama.cpp CUDA 上测得,CPU 只验证了有限输入。分领域、分题型、来源变化、BF16 GGUF 后端对照与 CPU / GPU 延迟见 2B Q8_0 模型卡。

校准验证与测试的区别

温度只在母记录分组后的验证划分 0 上拟合,选模使用划分 1–4。后者固定 1200 条验证数据上:

合并模型验证指标 T=1 T=1.0748723745
硬标签 ACC (%) 76.0906 76.0906
ECE(10 bins,%) 2.3324 1.0608
分布 NLL 0.585104 0.582978
Brier 0.316658 0.315944

上述验证 ECE 为 10 bins,不能混入使用 15 bins 的 CJ-Bench 测试表。全局温度不能保证所有领域、来源或题型都改善。

性能与已知局限

同服务器、25 条输入、每条预热后测 3 次,Transformers BF16 GPU 的单请求 p50 为 44.77 ms;短输入(92–255 tokens)p50 为 44.68 ms。这是预先分词后的前向及概率计算,不含分词、HTTP 或网络,且为共享 A800 服务器。完整长度范围为 92–6521 tokens,不能从这些输入推断生产吞吐。

  • 2B 的整体准确率更高,但法律判断与金融评分仍落后于 0.8B:法律 noul 为 82.07%(chinese-decision-0.8b 为 85.60%),金融 score 为 80.33%(0.8B 为 82.83%);Q8_0 为 82.73% 和 80.58%。
  • 商品分类来源 3,151 条在训练中波动:step3000 严重退化,step4000 恢复,发布权重 ACC 为 70.96%、NLL 1.5125(合并前 checkpoint 在 T = 1 时为 70.99%、1.6029)。整体指标不能代替来源诊断。
  • 温度校准降低了整体 ECE,但 Score MAE 比 T = 1 时略高(0.5642 vs 0.5607);金融领域仍低于闭源 Jev 的作者报告准确率。
  • 不用于聊天或开放式生成;仅针对中文固定候选任务训练。标签含义来自原任务,不能代替专业正确性判断。阈值与人工复核区间需用实际业务数据验证。

原始报告:reports/cj-bench-2b-step4994-20261017 与 reports/cj-bench-2b-step4994-calibrated-20261017。数据 manifest SHA256 为 461bb692676e8c01bbc89a5478a6cda9e5d0096afd2c82919d73f45a018ea351。更多训练进度和来源诊断见 训练交付记录。

归属、许可与引用

模型权重保留基础 Qwen3.5-2B 的 Apache-2.0 LICENSE。项目代码以 MIT 为主,RLCD 改编部分保留 Apache-2.0;原文与声明见 LICENSES/ 和 THIRD_PARTY_NOTICES.md。数据许可按来源分别适用,模型交付不授予数据再分发权利。

方法参考 hmm,RLCD 与数据参考 Chinese-Jev。本项目没有复现 Jev 或 Chinese-Jev 的训练过程。使用数据或基线时请同时引用 Chinese-Jev 论文。

@misc{chinese-decision-2026,
  title = {Chinese Decision: A Lightweight Chinese Probabilistic Decision Model},
  author = {yanqiangmiffy},
  year = {2026},
  howpublished = {\url{https://github.com/yanqiangmiffy/chinese-decision}}
}

Configuration

Architecture
Qwen3_5ForCausalLM
Context length (tokens)
262,144
Layers
24
Hidden size
2,048
Feed-forward size
6,144
Attention heads
8
Key/value heads
2
Head dimension
256
Vocabulary size
248,320
Model type
qwen3_5_text

Identity and Version

Repository
quincyqiang/chinese-decision-2b
Publisher
Quincyqiang
Task
Zero-shot classification
Modality
Text
Library
transformers
Parameters
1.9B parameters
Languages
zh
Revision
4edf2b416603d1c2774236d8f29b9e24f69d482c
First published
2026-10-17
Last updated
2026-10-18

Files and Weights

16 files, 3.8 GB in total. The weights are 1 file totalling 3.8 GB in safetensors.

Weights1 file · 3.8 GB
Configuration6 files · 40.5 KB
Tokenizer2 files · 20.0 MB
Documentation3 files · 35.9 KB
Other2 files · 17.9 KB
Repository2 files · 3.9 KB
Every file
FileTypeSizeSHA-256
model.safetensorsWeights3.8 GB a8fadd326b19
config.jsonConfiguration1.8 KB —
decision_config.jsonConfiguration234 B —
generation_config.jsonConfiguration116 B —
training_config.jsonConfiguration815 B —
training_summary.jsonConfiguration1.6 KB —
validation_report.jsonConfiguration36.0 KB —
LICENSEDocumentation11.3 KB —
README.mdDocumentation21.8 KB —
THIRD_PARTY_NOTICES.mdDocumentation2.8 KB —
LICENSES/Apache-2.0.txtOther10.2 KB —
chat_template.jinjaOther7.8 KB —
.gitattributesRepository1.6 KB —
.ms_upload_cacheRepository2.4 KB —
tokenizer.jsonTokenizer20.0 MB a5cd9732badc
tokenizer_config.jsonTokenizer1.5 KB —

License and Download

License
apache-2.0
Access
Open weights, no gate
Download size
3.8 GB
Download from Quincyqiang

Released by Quincyqiang through its official repository on Hugging Face. Read the license.

Built From

  • Derived from Qwen/Qwen3.5-2B
  • Trained on (disclosed) bbldCVer-hf/CJ-Dataset

Memory Requirements

PrecisionWeights in memory
As published3.8 GB
16-bit3.8 GB
8-bit1.9 GB
4-bit0.9 GB

Weights only, from the published parameter count; the key-value cache and runtime add to this.

Built on This Model

Questions About chinese-decision-2b

How much GPU memory does chinese-decision-2b need?

About 4.5 GB at 16-bit and 1.1 GB at 4-bit: the weights (1.9B parameters) plus a working margin. A long context needs more.

What is the cheapest GPU to run chinese-decision-2b on?

At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.

Can I use chinese-decision-2b commercially?

Yes. chinese-decision-2b is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.

What is chinese-decision-2b's context length?

262,144 tokens, from the maximum position embeddings in its published configuration.

Similar Models

Model · Zero-shot classification

d3-nano

vLLM Semantic Router

d3-nano is the 2B multimodal foundation decision model of Decision 3.0, the decision models of vLLM Semantic Router. Give it an input (text or JSON, optionally with images) and the questions you need answered: pick one of several options, say yes or no, or rate on a scale. It answers them all in one call and returns a probability for every answer, without generating text. - Jev Decision Index 0.3, public suite: 42.14, measured with the official 0.3 kit on the released weights: all 140,178 public requests answered, none unsupported. - +12.1 on the public suite over Decision 2.0 (its 2B model: 30.05 on the board), ahead in all five areas. Images go before the text of the request, each read at…

Open weights apache-2.0 2.2B parameters 262,144 tokens transformers

Model · Zero-shot classification

finecat-nli-v2-xl

Lee Miller

This model is a fine-tune of jhu-clsp/ettin-encoder-1b, trained on the dleemiller/FineCat-NLI-Reason dataset, a compilation of several high-quality NLI data sources with quality screening and reduction of easy samples in the training split. Also includes additional NLI, fact verification and long document sources (2.05M pairs in total). The training uses logit distillation from an NLI finetune of Qwen3.8-27B. \begin{equation} \mathcal{L} = \alpha \cdot \mathcal{L}{\text{CE}}(z^{(s)}, y) + \beta \cdot T^2 \cdot \mathrm{KL}\left(\sigma(z^{(t)}/T) \,\|\, \sigma(z^{(s)}/T)\right) \end{equation} where \\(z^{(s)}\\) and \\(z^{(t)}\\) are the student and teacher logits, \\(y\\) are the ground…

Open weights apache-2.0 1B parameters 7,999 tokens sentence-transformers

Model · Zero-shot classification

d3-lite

vLLM Semantic Router

d3-lite is the 0.8B multimodal foundation decision model of Decision 3.0, the decision models of vLLM Semantic Router. Give it an input (text or JSON, optionally with images) and the questions you need answered: pick one of several options, say yes or no, or rate on a scale. It answers them all in one call and returns a probability for every answer, without generating text. - Jev Decision Index 0.3, public suite: 36.94, measured with the official 0.3 kit on the released weights: all 140,178 public requests answered, none unsupported. - +16.4 on the public suite over Decision 2.0 (its 0.8B model: 20.57 on the board), ahead in all five areas. Images go before the text of the request, each…

Open weights apache-2.0 853M parameters 262,144 tokens transformers

Model · Zero-shot classification

Vela-2.0-0.8B

vLLM Semantic Router

Open Foundation Routing Models Routing decisions. Safety checks. Precise text spans. The compact hybrid member of Vela 2.0: a 756M-parameter model for multilingual routing, safety checks and span-level decisions through one interface. Define options, labels and rubrics at request time. Ask multiple named questions about a request, context and answer, and receive structured decisions with the text spans that support your workflow. 1. Compact deployment. About 3 GB of GPU memory for FP32 parameters, with a 16,384-token input limit. 2. Routing and safety together. Use one request for routing, prompt-attack checks, PII and unsupported-claim detection. 3. Decisions at span resolution. Return…

Open weights apache-2.0 755M parameters

Model · Zero-shot classification

OpenThai-SystemOne

iApp Technology

An open Thai + English "System One" decision model. It does not generate text. Given a state (any text or JSON) and typed questions, it returns calibrated probabilities over the options in one forward pass: The request/response contract mirrors TypeSafe's POST /v1/systemone so code written for the TypeSafe SDK can be pointed at this model unchanged. Typical uses: ticket routing, moderation, intent detection, RAG relevance judging, LLM-output verification, and computer-use / browser-agent action selection (which element to click, which tool to call). Gated-DeltaNet / attention, 262k context), vision encoder removed, then continued-pretrained on ~5B tokens of Thai (web, Wikipedia, parallel…

Open weights apache-2.0 753M parameters 262,144 tokens

Model · Zero-shot classification

chinese-decision-0.8b

Quincyqiang

各尺寸、部署格式与外部模型对比。2B BF16 与 2B Q80 均已公开发布并完成 CJ-Bench 全量评测:2B 整体准确率更高,0.8B 更小更快,部分法律判断与金融评分任务仍是 0.8B 更好。 Chinese Decision 是一个面向中文场景的轻量级决策模型。它不聊天、不生成文本,只做一件事:给定上下文和一组固定候选项,返回每个候选项的概率。 - 一次前向,零输出 token:读取最后位置上候选字母 A–Z 的 logits,softmax 后即为概率。 - 三种题型:noul(是/否)、choice(单选)、score(有序等级,返回期望分值)。 以下截图来自我们的测试平台 Decision Studio,后端运行的就是本模型(step 10000)。这些场景采用同一种设计:业务规则写在代码里,决策模型只负责给出概率。 每张工单在一次请求里同时问 6 个问题:类别(choice),严重程度和情绪(score),是否可复现、是否要求退款、是否有流失风险(noul)。不管后续用不用得上都一并提问,省去多次往返。路由规则写在代码里,根据这些概率决定工单派给哪个队列;类别置信度低于 0.5 时转人工复核。截图中 8 张工单 × 6 个问题共 48 次决策,用时 285 ms。 推荐同时安装 flash-linear-attention,它为 Qwen3.5 的线性注意力层提供高效 kernel: Hugging Face: ModelScope: decisionconfig.json 中的温度只在 softmax 前缩放…

Open weights apache-2.0 752M parameters 262,144 tokens transformers