Files

157 lines
4.0 KiB
Markdown
Raw Permalink Normal View History

---
license: Apache License 2.0
tags:
- 政衡
- 政企大模型
- 政务 AI 大模型
- JavaFamily
- 微调
- LLaMA-Factory
- LoRA
- SFT
- 政务领域
- 中文对话
tasks:
- text-generation
#model-type:
##如 gpt、phi、llama、chatglm、baichuan 等
#- gpt
model_type: llama3
#domain:
##如 nlp、cv、audio、multi-modal
#- nlp
#language:
##语言代码列表 https://help.aliyun.com/document_detail/215387.html?spm=a2c4g.11186623.0.0.9f8d7467kni6Aa
#- cn
#metrics:
##如 CIDEr、Blue、ROUGE 等
#- CIDEr
#tags:
##各种自定义,包括 pretrained、fine-tuned、instruction-tuned、RL-tuned 等训练方法和其他
#- pretrained
#tools:
##如 vllm、fastchat、llamacpp、AdaSeq 等
#- vllm
base_model:
- LLM-Research/Llama-3.2-3B-Instruct
base_model_relation: finetune
language:
- zh
metrics:
- accuracy
---
# Corvus-Gov-3B
当大模型从实验室涌向政务大厅,考验的不是才华,而是分寸。行政边界一字之差,政策口径一厘之偏,足以让信任归零。
「政衡」corvus-gov 基于 Llama3/Qwen3 基座模型,在百万级中文政务领域对话数据集上完成 SFT 微调。训练采用 LoRA 轻量化微调方案。 以衡为名,百万数据精训为每一句政务回答校准刻度。这不是参数的竞赛,而是关于可信的承诺。政企 AI 落地,自此有衡。
- 训练框架:LLaMA Factory
- 对话模板:`llama3`/`qwen3`
## Training Loss
训练 loss 曲线如下图所示,loss 随训练 step 平稳下降,模型收敛效果良好:
![training loss](https://www.modelscope.cn/models/JavaFamily/corvus-gov-3B/resolve/main/loss_plot.png)
## Training Data
- 数据规模:百万级中文政务领域对话样本
- 训练方式:LoRA SFT 有监督微调
- 训练框架:LLaMA Factory
## Quickstart
### ModelScope Python SDK
```python
from modelscope import AutoTokenizer, AutoModelForCausalLM
model_id = "JavaFamily/corvus-gov-3B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype="auto",
device_map="auto"
)
messages = [
{"role": "system", "content": "你是政务领域智能助手,请严谨、准确回答用户问题。"},
{"role": "user", "content": "请介绍一下政务咨询相关政策。"}
]
text = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to("cuda")
generated_ids = model.generate(
**model_inputs,
max_new_tokens=512,
temperature=0.7,
top_p=0.9
)
generated_ids = [
output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)
```
### vLLM 快速推理(推荐部署)
```python
from vllm import LLM, SamplingParams
model_id = "JavaFamily/corvus-gov-3B"
sampling_params = SamplingParams(temperature=0.7, top_p=0.9, max_tokens=512)
llm = LLM(model=model_id)
prompt = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True
)
outputs = llm.generate(prompt, sampling_params)
print(outputs[0].outputs[0].text)
```
#### 您可以通过如下git clone命令,或者ModelScope SDK来下载模型
1)ModelScope 下载
```bash
#安装ModelScope
pip install modelscope
```
* CLI 命令下载
```shell
# cli 下载
modelscope download --model JavaFamily/corvus-gov-3B --local_dir ./llama3/corvus‑gov‑3B
```
* SDK下载
```python
#SDK模型下载
from modelscope import snapshot_download
model_dir = snapshot_download('JavaFamily/corvus-gov-3B')
```
2) Git下载
```
#Git模型下载
git clone https://www.modelscope.cn/JavaFamily/corvus-gov-3B.git
```
<p style="color: lightgrey;">如果您是本模型的贡献者,我们邀请您根据<a href="https://modelscope.cn/docs/ModelScope%E6%A8%A1%E5%9E%8B%E6%8E%A5%E5%85%A5%E6%B5%81%E7%A8%8B%E6%A6%82%E8%A7%88" style="color: lightgrey; text-decoration: underline;">模型贡献文档</a>,及时完善模型卡片内容。</p>