60 lines
1.7 KiB
Markdown
60 lines
1.7 KiB
Markdown
|
---
|
|||
|
|
license: apache-2.0
|
|||
|
|
base_model: Qwen/Qwen2.5-1.5B-Instruct
|
|||
|
|
tags:
|
|||
|
|
- finance
|
|||
|
|
- qwen2.5
|
|||
|
|
- lora
|
|||
|
|
- conversational
|
|||
|
|
- chinese
|
|||
|
|
language:
|
|||
|
|
- zh
|
|||
|
|
- en
|
|||
|
|
pipeline_tag: text-generation
|
|||
|
|
library_name: transformers
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
# Druckenmiller 1.5B v2（实验性）
|
|||
|
|
|
|||
|
|
斯坦利·德鲁肯米勒风格的**宏观交易认知陪练**小模型。基于 `Qwen/Qwen2.5-1.5B-Instruct` 做 QLoRA 微调后合并。
|
|||
|
|
|
|||
|
|
> **免责声明**：本模型仅供研究与教育演示，**不构成投资建议**。金融市场有风险，决策请自行负责。
|
|||
|
|
|
|||
|
|
## 训练概况
|
|||
|
|
|
|||
|
|
| 项 | 值 |
|
|||
|
|
|----|-----|
|
|||
|
|
| 基座 | Qwen2.5-1.5B-Instruct |
|
|||
|
|
| 数据 | ~284 条 ShareGPT（约 28% 多轮） |
|
|||
|
|
| 方法 | QLoRA r=16，4 epochs |
|
|||
|
|
| 多轮评测 | 6/10（门槛 8/10，未达标） |
|
|||
|
|
| 状态 | **Experimental preview** |
|
|||
|
|
|
|||
|
|
## 文件
|
|||
|
|
|
|||
|
|
- Transformers `safetensors` + tokenizer：HF Space CPU / 本地 GPU
|
|||
|
|
- `druckenmiller-1.5b-v2-q4_k_m.gguf`：推荐用 `llama.cpp` / `llama-cpp-python`（约 1GB）
|
|||
|
|
- `druckenmiller-1.5b-v2-f16.gguf`：全精度 GGUF（约 3GB）
|
|||
|
|
|
|||
|
|
## 本地快速试用
|
|||
|
|
|
|||
|
|
```bash
|
|||
|
|
# Transformers（需 GPU 更佳）
|
|||
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|||
|
|
import torch
|
|||
|
|
|
|||
|
|
repo = "januschoy/druckenmiller-1.5b-v2"
|
|||
|
|
tok = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
|
|||
|
|
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype=torch.float16, device_map="auto", trust_remote_code=True)
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
## 已知限制
|
|||
|
|
|
|||
|
|
- 多轮追问偶发跑题（分众粘性、共识黄金等场景）
|
|||
|
|
- 1.5B 容量有限，宏观细节可能幻觉
|
|||
|
|
- 公开 Demo 若跑在免费 CPU Space，延迟较高且有冷启动
|
|||
|
|
|
|||
|
|
## License
|
|||
|
|
|
|||
|
|
权重遵循 Apache-2.0；并请遵守 [Qwen2.5](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct) 相关条款。
|