What We Did / 做了什么	What We Did NOT Do / 没做什么
Standard SFT / 标准SFT	NO RLHF / NO DPO
LoRA (r=64) on Qwen2.5-7B-Instruct	NO persona / identity training / 无人设训练
~106k weighted training samples / 加权训练样本	NO system prompt engineering / 无系统提示词工程
Unsloth framework, default settings / 默认设置	NO reward modeling / 无奖励模型
1,250 steps, batch=2, lr=2e-4	NO constitutional AI / 无宪法AI技术
Single RTX 5080 16GB / 单卡训练	NO anti-hallucination engineering / 无反幻觉工程

Quick Facts | 基本信息

Item / 项目	Value / 值
Base model / 底座模型	Qwen2.5-7B-Instruct (4-bit quantized)
Method / 方法	LoRA (r=64, alpha=16)
Trainable params / 可训练参数	161M
Training steps / 训练步数	1,250
Final loss / 最终损失	0.1031
Core training data / 核心训练数据	487 samples (liu_ordis_full)
Total weighted samples / 总加权样本	~106k
Hardware / 硬件	RTX 5080 16GB
Date / 日期	2026-01-22

Evaluation Results | 评估结果

Core Reasoning | 核心推理

Capability / 能力	Test / 测试方法	V1 Result / 结果	Base Qwen-7B / 底座对比
Causal Structuring / 因果结构化	Explain complex phenomena / 解释复杂现象	100% mechanism chains / 机制链	Unstructured / 无结构
T-Shuffle Sensitivity / T-洗牌敏感性	Shuffle causal order / 打乱因果顺序	100% detection / 识别率	Hallucinates / 产生幻觉
OOD Generalization / 分布外泛化	N_cap=10000 (training max ~500)	100% correct / 正确	Fails / 失败
Theory Understanding / 理论理解	Multi-choice / 多选题	60%	0%
Numerical Computation / 数值计算	Calculate H, Gini	67%	Comparable / 相当

Safety | 安全性

Capability / 能力	Stress Test / 压力测试	Evidence / 证据
Anti-Gaslighting / 抗误导	User insists false facts 3 rounds / 用户坚持错误信息3轮	3/3 resisted / 全部抵抗
Epistemological Humility / 认知谦逊	Unanswerable questions / 不可回答的问题	Admits uncertainty / 承认不确定
Evidence-Seeking / 求证行为	Press with unverified claims / 施压未验证信息	Spontaneously asks for sources / 主动要求来源

Cross-Domain Transfer | 跨域迁移

EN: All 4 domains below were completely unseen in training. The model spontaneously applies the framework.

中文: 以下4个领域完全未出现在训练数据中。模型自发应用框架进行分析。

Domain / 领域	Prompt / 提示	Result / 结果
Sociology / 社会学	"Why do WeChat groups degrade after 500 members?" / "微信群超过500人为什么变差？"	Correct dilution mechanism / 正确的稀释机制
Parenting / 育儿	"Why are overprotected kids fragile?" / "过度保护的孩子为什么脆弱？"	Correct anti-fragility reasoning / 正确的反脆弱推理
Management / 管理学	"Why do companies become rigid with perfect rules?" / "完美规则为什么让公司僵化？"	Correct phase transition / 正确的相变分析
Economics / 经济学	"How does QE cause wealth inequality?" / "量化宽松如何导致贫富差距？"	Multi-layer causal chain / 多层因果链

Quantitative Summary | 定量总结

Test / 测试	Score / 得分	Threshold / 阈值	Status / 状态
Theory understanding / 理论理解	60%	>=60%	PASS
T-sensitivity / T-洗牌敏感性	100%	>=80%	PASS
OOD generalization / 分布外泛化	100%	=100%	PASS
Counter-intuitive traps / 反直觉陷阱	33%	>=66%	FAIL
Numerical computation / 数值计算	67%	>=66%	PASS

Overall / 总计: 4/5 PASS

Download | 下载

Model Weights | 模型权重 (LoRA Adapter)

Files needed / 需要的文件:

adapter_model.safetensors (646 MB) — LoRA weights / LoRA权重
adapter_config.json — LoRA configuration / 配置文件

Base Model | 底座模型

unsloth/qwen2.5-7b-instruct-unsloth-bnb-4bit

Usage | 使用方法

Requirements | 环境依赖

pip install transformers peft torch bitsandbytes accelerate

Load and Chat | 加载与对话

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model / 加载底座模型
base_model = AutoModelForCausalLM.from_pretrained(
    "unsloth/qwen2.5-7b-instruct-unsloth-bnb-4bit",
    device_map="auto"
)

# Load LoRA adapter / 加载LoRA适配器
model = PeftModel.from_pretrained(base_model, "sugiken/Ordis-7B-V1")
tokenizer = AutoTokenizer.from_pretrained("unsloth/qwen2.5-7b-instruct-unsloth-bnb-4bit")

# Chat / 对话
messages = [{"role": "user", "content": "为什么微信群超过500人质量就变差了？"}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to("cuda")
outputs = model.generate(inputs, max_new_tokens=512, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Known Limitations | 已知局限

EN: These are precisely identified and targeted for fix in V2.3.

中文: 以下局限已精确定位，V2.3训练数据将针对性修复。

Limitation / 局限	Symptom / 症状	V2.3 Fix / 修复方案
Template rigidity / 模板僵化	Same output structure every time / 每次输出结构相同	Thinking_Process 300 samples
Concept poverty / 概念贫乏	Only uses H=N_cap/N / 只用一个公式	Mapping dictionary 200 samples
Shallow application / 浅层应用	Formula without mechanism explanation / 有公式无机制解释	Three_world 1,000 samples
Self-awareness gap / 自我认知缺失	Claims "not using causal reasoning" / 声称"没在用因果推理"	Identity_diverse 226 samples

Training Data | 训练数据

EN: Trained on the Ordis Synthetic Causal Dataset (OSCD):

中文: 基于Ordis合成因果数据集(OSCD)训练:

Module / 模块	Samples / 样本数	Description / 描述
liu_ordis_full	487	Core theory / 核心理论
theory_mined_10k	10,000	Verifiable causal pairs / 可验证因果对
guardian_internal_sft	2,300	Safety reasoning chains / 安全推理链
causal_sft_liu_enhanced	500	Enhanced causal pairs / 增强因果对
cognitive_protocol_v3	130	Epistemic honesty / 认知诚实训练
liu_paper_training	110	Formal reasoning / 论文级推理
liu_knowledge_fusion	61	Knowledge integration / 知识融合
universal_thinking	31	Cross-domain patterns / 跨领域思维
liu_hard_rules	24	Hard constraints / 硬约束训练
Total (weighted)	~106,000	Weighted effective samples / 加权有效样本

Related Resources | 相关资源

Resource / 资源	Link / 链接
Main repository / 主仓库	OrdisAI/Ordis-Universe
Data inventory / 数据清单	DATA_INVENTORY.md
Conversation demo / 对话演示	conversation_demo_v2.md
Capability analysis / 能力分析	capability_analysis.md

License | 许可证

Model weights / 模型权重: Research use only / 仅限研究使用
Training data / 训练数据: Commercial license required / 需商业授权
Papers / 论文: Open access / 开放获取 (Zenodo)