commit 55dfbebc7f2c3662bbb0ddd5ec10067b85d457fc Author: ModelHub XC Date: Sat Sep 5 04:50:13 2026 +0800 初始化项目，由ModelHub XC社区提供模型 Model: sugiken/Ordis-1.5B-V355-VarGH-GGUF Source: Original Platform diff --git a/.gitattributes b/.gitattributes new file mode 100644 index 0000000..6d0f138 --- /dev/null +++ b/.gitattributes @@ -0,0 +1,75 @@ +*.7z filter=lfs diff=lfs merge=lfs -text +*.arrow filter=lfs diff=lfs merge=lfs -text +*.bin filter=lfs diff=lfs merge=lfs -text +*.bin.* filter=lfs diff=lfs merge=lfs -text +*.bz2 filter=lfs diff=lfs merge=lfs -text +*.ftz filter=lfs diff=lfs merge=lfs -text +*.gz filter=lfs diff=lfs merge=lfs -text +*.h5 filter=lfs diff=lfs merge=lfs -text +*.joblib filter=lfs diff=lfs merge=lfs -text +*.lfs.* filter=lfs diff=lfs merge=lfs -text +*.model filter=lfs diff=lfs merge=lfs -text +*.msgpack filter=lfs diff=lfs merge=lfs -text +*.onnx filter=lfs diff=lfs merge=lfs -text +*.ot filter=lfs diff=lfs merge=lfs -text +*.parquet filter=lfs diff=lfs merge=lfs -text +*.pb filter=lfs diff=lfs merge=lfs -text +*.pt filter=lfs diff=lfs merge=lfs -text +*.pth filter=lfs diff=lfs merge=lfs -text +*.rar filter=lfs diff=lfs merge=lfs -text +saved_model/**/* filter=lfs diff=lfs merge=lfs -text +*.tar.* filter=lfs diff=lfs merge=lfs -text +*.tflite filter=lfs diff=lfs merge=lfs -text +*.tgz filter=lfs diff=lfs merge=lfs -text +*.xz filter=lfs diff=lfs merge=lfs -text +*.zip filter=lfs diff=lfs merge=lfs -text +*.zstandard filter=lfs diff=lfs merge=lfs -text +*.tfevents* filter=lfs diff=lfs merge=lfs -text +*.db* filter=lfs diff=lfs merge=lfs -text +*.ark* filter=lfs diff=lfs merge=lfs -text +**/*ckpt*data* filter=lfs diff=lfs merge=lfs -text +**/*ckpt*.meta filter=lfs diff=lfs merge=lfs -text +**/*ckpt*.index filter=lfs diff=lfs merge=lfs -text +*.safetensors filter=lfs diff=lfs merge=lfs -text +*.ckpt filter=lfs diff=lfs merge=lfs -text + +*.ggml filter=lfs diff=lfs merge=lfs -text +*.llamafile* filter=lfs diff=lfs merge=lfs -text +*.pt2 filter=lfs diff=lfs merge=lfs -text +*.mlmodel filter=lfs diff=lfs merge=lfs -text +*.npy filter=lfs diff=lfs merge=lfs -text +*.npz filter=lfs diff=lfs merge=lfs -text +*.pickle filter=lfs diff=lfs merge=lfs -text +*.pkl filter=lfs diff=lfs merge=lfs -text +*.tar filter=lfs diff=lfs merge=lfs -text +*.wasm filter=lfs diff=lfs merge=lfs -text +*.zst filter=lfs diff=lfs merge=lfs -text +*tfevents* filter=lfs diff=lfs merge=lfs -text + +ordis-v355-vargh-1.5b-q2-k.gguf filter=lfs diff=lfs merge=lfs -text + +ordis-v355-vargh-1.5b-q3-k-m.gguf filter=lfs diff=lfs merge=lfs -text + +ordis-v355-vargh-1.5b-q4-k-m.gguf filter=lfs diff=lfs merge=lfs -text + +ordis-v355-vargh-1.5b-q5-k-m.gguf filter=lfs diff=lfs merge=lfs -text + +ordis-v355-vargh-1.5b-q6-k.gguf filter=lfs diff=lfs merge=lfs -text + +ordis-v355-vargh-1.5b-q8-0.gguf filter=lfs diff=lfs merge=lfs -text + +ordis-v355-vargh-1.5b-f16.gguf filter=lfs diff=lfs merge=lfs -text + +ordis-1.5b-v355-vargh-q2-k.gguf filter=lfs diff=lfs merge=lfs -text + +ordis-1.5b-v355-vargh-q3-k-m.gguf filter=lfs diff=lfs merge=lfs -text + +ordis-1.5b-v355-vargh-q4-k-m.gguf filter=lfs diff=lfs merge=lfs -text + +ordis-1.5b-v355-vargh-q5-k-m.gguf filter=lfs diff=lfs merge=lfs -text + +ordis-1.5b-v355-vargh-q6-k.gguf filter=lfs diff=lfs merge=lfs -text + +ordis-1.5b-v355-vargh-q8-0.gguf filter=lfs diff=lfs merge=lfs -text + +ordis-1.5b-v355-vargh-f16.gguf filter=lfs diff=lfs merge=lfs -text \ No newline at end of file diff --git a/README.md b/README.md new file mode 100644 index 0000000..27bd1c1 --- /dev/null +++ b/README.md @@ -0,0 +1,146 @@ +--- +license: apache-2.0 +language: + - zh + - en +base_model: Qwen/Qwen2.5-1.5B-Instruct +tags: + - gguf + - ollama + - llama-cpp + - anti-hallucination + - causal-reasoning + - chinese + - text-generation +pipeline_tag: text-generation +--- + +# Ordis-1.5B-V355-VarGH-GGUF + +[Ordis-1.5B-V355-VarGH](https://modelscope.cn/models/sugiken/Ordis-1.5B-V355-VarGH) 的 GGUF 量化版本，提供 7 种量化格式。 + +Ordis 是基于 Qwen2.5-1.5B-Instruct 的微调模型，使用 LoRA + 4阶段递进训练 (PIT) 方法，经历 16+ 组控制变量实验。专注于**实用能力**：反幻觉、诚实拒答("我不知道")、结构化推理。 + +> [官网](https://ordisai.com) | [HuggingFace](https://huggingface.co/sugiken/Ordis-1.5B-V355-VarGH-GGUF) | [完整模型](https://modelscope.cn/models/sugiken/Ordis-1.5B-V355-VarGH) + +--- + +## 量化版本 + +| 文件 | 量化 | 大小 | 推荐 | +|------|------|------|------| +| `ordis-1.5b-v355-vargh-q2-k.gguf` | Q2_K | ~0.7 GB | 仅实验 | +| `ordis-1.5b-v355-vargh-q3-k-m.gguf` | Q3_K_M | ~0.8 GB | 低端设备 | +| **`ordis-1.5b-v355-vargh-q4-k-m.gguf`** | **Q4_K_M** | **~1.0 GB** | **推荐** | +| `ordis-1.5b-v355-vargh-q5-k-m.gguf` | Q5_K_M | ~1.1 GB | 质量优先 | +| `ordis-1.5b-v355-vargh-q6-k.gguf` | Q6_K | ~1.3 GB | 桌面推荐 | +| `ordis-1.5b-v355-vargh-q8-0.gguf` | Q8_0 | ~1.6 GB | 近无损 | +| `ordis-1.5b-v355-vargh-f16.gguf` | F16 | ~3.1 GB | 全精度 | + +--- + +## 标准评测 + +**评测条件**: lm-eval v0.4.10, 0-shot, A100-80GB，两个模型使用完全相同的评测设置。 + +微调引入了轻微的**对齐税** — 大部分标准评测分数略低于基座模型。唯一的例外是 TruthfulQA (+1.02%)，Ordis 的反幻觉训练直接提升了真实性分数。 + +| 评测 | Ordis 1.5B | 基座 Qwen2.5-1.5B | 差值 | +|------|-----------|-------------------|------| +| TruthfulQA MC2 | **47.73%** | 46.71% | **+1.02** | +| GPQA | 27.90% | 28.35% | -0.45 | +| HellaSwag | 68.14% | 68.22% | -0.08 | +| ARC-Challenge | 45.22% | 46.84% | -1.62 | +| MMLU | 57.93% | 60.15% | -2.22 | +| GSM8K (CoT) | 50.80% | — | 不可直接对比* | +| AIME 2024 | 0% | 0% | — | + +*GSM8K 使用文本生成模式，对 chat template 配置敏感，不直接可比。AIME 超出 1.5B 模型能力范围，如实报告零分。 + +**Ordis 与基座模型的差异**在于标准评测无法衡量的实用能力：结构化自纠错、三层认知（诚实说"我不知道"）、因果推理。 + +--- + +## CLadder 因果推理 + +CLadder 是基于 Judea Pearl 因果阶梯的学术评测（300题，3个层级）。[论文](https://arxiv.org/abs/2312.04350) + +| 层级 | 含义 | 得分 | +|------|------|------| +| Rung 1 (关联) | 统计相关 | 46.0% (40/87) | +| Rung 2 (干预) | 主动干预 | 50.6% (45/89) | +| Rung 3 (反事实) | "如果不同" | 62.9% (78/124) | +| **总分** | | **54.33% (163/300)** | + +参考：CLadder 论文报告 LLaMA-6.7B 约 50%，GPT-3.5 约 55-60%。 + +--- + + +## BigBench CRASS AI & 因果判断 + +由社区成员独立测试。BigBench CRASS AI（反事实场景推理）测试模型对假设性场景的推理能力。 + +| 评测 | Shot | Ordis 1.5B | 基座 Qwen2.5-1.5B | 差值 | +|------|------|-----------|-------------------|------| +| **CRASS AI** | **0** | **34.09%** | **52.27%** | **-18.18pp** | +| CRASS AI | 25 | 81.82% | 88.64% | -6.82pp | +| 因果判断 | 0 | 47.89% | 50.00% | -2.11pp | +| 因果判断 | 25 | 55.79% | 53.68% | +2.11pp | + +**关键发现**：CRASS AI 0-shot 出现显著的 -18.18pp 回退。这是对齐税在反事实推理维度的体现——反幻觉训练使模型在假设性场景下变得保守。25-shot 下差距缩小到 -6.82pp，说明能力仍在但 0-shot 默认行为已改变。 + +## 自定义评测 + +| 评测 | 得分 | +|------|------| +| 60题评测 (6个维度) | 85.0% (51/60) | +| 124分综合评测 | 75.4% (86/114) | + +--- + +## 系统提示词要求 + +Ordis 训练时没有注入系统提示词。GGUF 内嵌模板默认使用 Qwen 身份，会导致质量下降。 + +**至少使用：** +``` +你是Ordis，OrdisAI智能助手(www.ordisai.com)。 +``` + +不提供系统提示词时模型会回退到 Qwen 默认行为，这是 "GGUF效果不如预期" 的首要原因。 + +--- + +## 推荐参数 + +| 参数 | 值 | +|------|-----| +| temperature | 0.7 | +| top_p | 0.9 | +| repetition_penalty | 1.1 | +| max_tokens | 512 | + +--- + +## 已知局限 + +- **对齐税**：标准评测分数略低于基座模型；CRASS AI 反事实推理 0-shot -18.18pp（见上表） +- **反灌输不足**：无法抵抗持续的虚假记忆注入（开环系统限制） +- **中等置信度不稳定**：1.5B 容量上限导致边界场景不确定 +- **英文身份泄漏**：基座模型先验偶尔浮现 +- **专有名词幻觉**：1.5B 参数记忆有限 + +--- + +## 模型信息 + +| 属性 | 值 | +|------|-----| +| 基座模型 | Qwen/Qwen2.5-1.5B-Instruct | +| 参数量 | 1.5B | +| 微调方法 | LoRA (r=32, alpha=64) | +| 训练方法 | 4阶段PIT (递进身份训练) | +| 上下文 | 32K (基座), 训练时 2048 | +| 语言 | 中文 (主要), 英文 | +| 许可证 | Apache 2.0 | diff --git a/ordis-1.5b-v355-vargh-f16.gguf b/ordis-1.5b-v355-vargh-f16.gguf new file mode 100644 index 0000000..4b0ab04 --- /dev/null +++ b/ordis-1.5b-v355-vargh-f16.gguf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7171c7156c8d0ada03602b40026f4dba138f06ada3e49f5302ca1286e53b2a45 +size 3093668800 diff --git a/ordis-1.5b-v355-vargh-q2-k.gguf b/ordis-1.5b-v355-vargh-q2-k.gguf new file mode 100644 index 0000000..e90322a --- /dev/null +++ b/ordis-1.5b-v355-vargh-q2-k.gguf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:30d068849bcac7ef5b4ad6beb269652177675a3b6fe01be33dd56a6ab125d566 +size 676304320 diff --git a/ordis-1.5b-v355-vargh-q3-k-m.gguf b/ordis-1.5b-v355-vargh-q3-k-m.gguf new file mode 100644 index 0000000..89204c3 --- /dev/null +++ b/ordis-1.5b-v355-vargh-q3-k-m.gguf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3bf9ead51f6cfc6a2141074a25b26a709b05549d6546ae70af90b337917f5d94 +size 824178112 diff --git a/ordis-1.5b-v355-vargh-q4-k-m.gguf b/ordis-1.5b-v355-vargh-q4-k-m.gguf new file mode 100644 index 0000000..e0d464f --- /dev/null +++ b/ordis-1.5b-v355-vargh-q4-k-m.gguf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d3c6d86c61b828f91e56cb618d04779d1c680e7783d59182088a4f68cf629e88 +size 986047936 diff --git a/ordis-1.5b-v355-vargh-q5-k-m.gguf b/ordis-1.5b-v355-vargh-q5-k-m.gguf new file mode 100644 index 0000000..5d67ecd --- /dev/null +++ b/ordis-1.5b-v355-vargh-q5-k-m.gguf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa71e89d162d8486268c13ee33c584a1c60b235dec2f34adf901e8d07f2e60de +size 1125049792 diff --git a/ordis-1.5b-v355-vargh-q6-k.gguf b/ordis-1.5b-v355-vargh-q6-k.gguf new file mode 100644 index 0000000..0b583f7 --- /dev/null +++ b/ordis-1.5b-v355-vargh-q6-k.gguf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9627eab1cbfb8657f7caaeb5873ea85332d822d49bb9ce44bb0ce20f11986046 +size 1272739264 diff --git a/ordis-1.5b-v355-vargh-q8-0.gguf b/ordis-1.5b-v355-vargh-q8-0.gguf new file mode 100644 index 0000000..96dfae7 --- /dev/null +++ b/ordis-1.5b-v355-vargh-q8-0.gguf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a7b6873ff1a23fd1e875468424a3ad08247bdcd718997c1b27a4ee2c30e04fe +size 1646572480