Instructions to use liuchen88/CLA1.1-27B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use liuchen88/CLA1.1-27B with MLX:

# Make sure mlx-lm is installed
# pip install --upgrade mlx-lm

# Generate text with mlx-lm
from mlx_lm import load, generate

model, tokenizer = load("liuchen88/CLA1.1-27B")

prompt = "Write a story about Einstein"
messages = [{"role": "user", "content": prompt}]
prompt = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True
)

text = generate(model, tokenizer, prompt=prompt, verbose=True)

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use liuchen88/CLA1.1-27B with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf liuchen88/CLA1.1-27B:F16
# Run inference directly in the terminal:
llama cli -hf liuchen88/CLA1.1-27B:F16

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf liuchen88/CLA1.1-27B:F16
# Run inference directly in the terminal:
llama cli -hf liuchen88/CLA1.1-27B:F16

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf liuchen88/CLA1.1-27B:F16
# Run inference directly in the terminal:
./llama-cli -hf liuchen88/CLA1.1-27B:F16

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf liuchen88/CLA1.1-27B:F16
# Run inference directly in the terminal:
./build/bin/llama-cli -hf liuchen88/CLA1.1-27B:F16

Use Docker

docker model run hf.co/liuchen88/CLA1.1-27B:F16

Llama new
LM Studio
Jan

vLLM

How to use liuchen88/CLA1.1-27B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "liuchen88/CLA1.1-27B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "liuchen88/CLA1.1-27B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/liuchen88/CLA1.1-27B:F16

Ollama
How to use liuchen88/CLA1.1-27B with Ollama:
```
ollama run hf.co/liuchen88/CLA1.1-27B:F16
```
Unsloth Desktop

How to use liuchen88/CLA1.1-27B with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "liuchen88/CLA1.1-27B"

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "mlx-lm": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "liuchen88/CLA1.1-27B"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

MLX LM

How to use liuchen88/CLA1.1-27B with MLX LM:

Generate or start a chat session

# Install MLX LM
uv tool install mlx-lm
# Interactive chat REPL
mlx_lm.chat --model "liuchen88/CLA1.1-27B"

Run an OpenAI-compatible server

# Install MLX LM
uv tool install mlx-lm
# Start the server
mlx_lm.server --model "liuchen88/CLA1.1-27B"
# Calling the OpenAI-compatible server with curl
curl -X POST "http://localhost:8000/v1/chat/completions" \
   -H "Content-Type: application/json" \
   --data '{
     "model": "liuchen88/CLA1.1-27B",
     "messages": [
       {"role": "user", "content": "Hello"}
     ]
   }'

Docker Model Runner
How to use liuchen88/CLA1.1-27B with Docker Model Runner:
```
docker model run hf.co/liuchen88/CLA1.1-27B:F16
```

Lemonade

How to use liuchen88/CLA1.1-27B with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull liuchen88/CLA1.1-27B:F16

Run and chat with the model

lemonade run user.CLA1.1-27B-F16

List all available models

lemonade list

Hermes Agent

How to use liuchen88/CLA1.1-27B with Hermes Agent:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "liuchen88/CLA1.1-27B"

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default liuchen88/CLA1.1-27B

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use liuchen88/CLA1.1-27B with OpenClaw:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "liuchen88/CLA1.1-27B"

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "liuchen88/CLA1.1-27B" \
  --custom-provider-id mlx-lm \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

CLA 1.0-27B

Commercial Loan Approval v1.0 — 基于 Qwen3.8-27B (27B 参数, bfloat16 原版) 通过 QLoRA 微调的企业信贷审批专用大模型，支持多模态输入（财报截图、票据、营业执照等图片）。

本仓库提供 GGUF 多模态版：CLA1.1-27B-f16.gguf（文本塔 F16）+ mmproj-CLA1.1-27B-f16.gguf（视觉投影塔）+ preprocessor_config.json + video_preprocessor_config.json，适用于 llama.cpp / Ollama / LM Studio。

模型概览

项目	内容
模型名称	CLA 1.0-27B
基座模型	Qwen3.8-27B (27B 参数, bfloat16 原版)
微调方式	QLoRA (LoRA rank=32, alpha=1024, dropout=0.2, ckpt-900)
上下文窗口	262,144 tokens
量化精度	F16 (文本塔) / BF16 (视觉塔)
适用平台	llama.cpp / Ollama / LM Studio / MLX

模型架构

项目	规格
架构	Qwen3_5ForCausalLM (混合注意力)
参数量	27B (bfloat16 约 54 GB, F16 导出约 53.8 GB)
隐藏维度	5120
层数	64 (48 SSM/线性注意力 + 16 Full Attention)
词表大小	248,320
注意力头	24 Q heads + 4 KV heads (GQA), head_dim=256
MLP 中间层	17,408 (SwiGLU)
RoPE	θ=10,000,000, mrope sections=[11,11,10,0]
上下文长度	262,144 tokens (原生支持)
视觉塔	Qwen3-VL 编码器 (27 层 ViT, 1152 dim, patch 16, merge 2)
投影器	Qwen3VL Merger (2 层 MLP, 4608→5120)

混合注意力机制

48 层 SSM/线性注意力：无 KV cache，推理恒定内存
**16 层 Full Attention (每 4 层 1 层)**：标准因果注意力，需 KV cache
实际 KV cache 仅 16 层：llama.cpp 读取 qwen35.attention.recurrent_layers 自动跳过 SSM 层

GGUF 文件

文件	大小	SHA256	说明
`CLA1.1-27B-f16.gguf`	~50 GB	`084446b4f812275c6d8dc6495bec777cfe45c0a1a298a3a4694b215b3bb3dbff`	文本塔 F16，含 64 层混合注意力权重
`mmproj-CLA1.1-27B-f16.gguf`	~888 MB	`113b5f8473ec754fe00abff6c3ae0c965f9cbee2e9ccd653740dbe39f29e8669`	视觉投影塔 BF16，含 ViT 编码器 + Merger

关键元数据（自动写入，无需手动配置）

qwen35.block_count = 64
qwen35.full_attention_interval = 4
qwen35.attention.recurrent_layers = [True, True, True, False, ...]  # 48 True + 16 False
general.architecture = qwen35
general.file_type = 1 (F16)
general.quantization_version = 2

内存需求估算（F16 + 262K ctx）

组件	大小
模型权重	~53.8 GB
KV cache = 16 层 × 4 KV heads(GQA) × 256 head_dim × 2(K+V) × 2 bytes(f16) × 262,144	~17.2 GB
总计 (Flash Attention 开启)	~71 GB
总计 (Flash Attention 关闭, f32 KV)	~137 GB (guardrail 显示值，非实际占用)

LM Studio 显示 ~137 GB 是估算器 Bug，实际运行仅需 ~71 GB。解决：开 Flash Attention 或降 n_ctx。

训练信息

超参数	值
LoRA target_modules	`q_proj`, `v_proj`, `k_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj` (7 类投影层)
rank / scale (MLX) / alpha (PEFT等价)	32 / 32 / 1024
dropout	0.2
mask_prompt	true (仅计算 assistant 回复 loss)
学习率	1.4e-6 (cosine decay, warmup=50 steps)
序列长度	2304 (训练) / 262,144 (推理原生支持)
训练框架	mlx-lm (Apple MLX / Metal)
平台	macOS (Apple Silicon)
混合精度	bfloat16 (训练) / F16 (导出)

损失函数

标准因果语言建模 loss (cross-entropy)
mask_prompt=true：system + user 部分 loss 置零，仅反传 assistant token

多模态用法

能力范围

财报截图/PDF 渲染页 → 表格结构化提取
营业执照/许可证 → 关键字段识别
银行回单/票据 → 金额、日期、付款方核验
诉讼文书/判决书 → 案由、金额、当事人抽取

加载要求（三件套缺一不可）

CLA1.1-27B-f16.gguf          # 文本塔
mmproj-CLA1.1-27B-f16.gguf   # 视觉投影
preprocessor_config.json     # 图片预处理配置
video_preprocessor_config.json # 视频预处理配置

示例 Prompt

# 图片 + 文本混合输入
messages = [
    {"role": "system", "content": "你是一位具有20年企业信贷审批经验的商业银行对公信贷审批专家。"},
    {"role": "user", "content": [
        {"type": "image", "image": "financial_statement.png"},
        {"type": "text", "text": "该企业申请授信1000万元，请结合财报图片给出审批结论。"}
    ]}
]

部署指引

llama.cpp (CPU/Metal)

# 仅文本
./llama-cli -m CLA1.1-27B-f16.gguf -c 32768 -p "..." -n 2048

# 多模态 (需 --mmproj)
./llama-cli \
  -m CLA1.1-27B-f16.gguf \
  --mmproj mmproj-CLA1.1-27B-f16.gguf \
  -c 32768 \
  -p "你是信贷专家...企业申请授信..." \
  -n 2048

Ollama

# Modelfile
FROM ./CLA1.1-27B-f16.gguf
TEMPLATE "{{ .System }}\n\n{{ .Prompt }}\n\n{{ .Response }}"
PARAMETER num_ctx 32768
PARAMETER num_gpu 99

ollama create cla-27b -f Modelfile
ollama run cla-27b "企业申请授信500万，资产负债率65%，流动比率1.3，利息保障倍数2.1，无不良记录。"

LM Studio

将四个文件放入同一目录
点击 "Add Model" 选择 CLA1.1-27B-f16.gguf
关键设置（Settings → Model → GPU Offload）：
- ✅ Flash Attention (KV cache f16，内存减半)
- n_ctx: 32768 (生产建议，估算进一步下降)
- GPU Offload: 全层 (Metal 统一内存)
模型列表会显示 Vision 标签，支持拖入图片

限制与免责声明

数据局限

非中国司法管辖区企业、特殊行业（金融/军工/烟草等）适用性未验证
宏观周期、政策突变等外生冲击未显式建模

模型局限

三态分类存在误判风险（尤其边界样本：cond_approve ↔ approve/reject）
长上下文（>32K）推理注意力稀疏化可能遗漏关键细节
视觉塔仅支持定长 patch，超长报表需分页/切片

部署局限

需 ≥64GB 统一内存 (Mac) 或 ≥80GB VRAM (NVIDIA F16)
GGUF 面向 llama.cpp 生态，推理引擎支持 CPU/Metal/CUDA/Vulkan；原生 PyTorch 推理需转回 HF safetensors 格式
多模态需完整三件套，缺一不可

合规声明

本模型仅供信贷审批辅助决策参考，不可替代人工审批岗
审批结论存在误判风险，请结合尽调报告、实地走访、行业研判综合决策
使用者须遵守《商业银行资本管理办法》《流动资金贷款管理办法》等监管要求
本项目不构成任何金融建议、授信承诺或法律意见

引用与致谢

Qwen Team: Qwen3.8-27B 基座模型 (Apache-2.0)

Downloads last month: -

GGUF

Model size

27B params

Architecture

qwen35

Hardware compatibility

16-bit

Model tree for liuchen88/CLA1.1-27B

Base model

Qwen/Qwen3.8-27B

Adapter

(268)

this model