|
--- |
|
language: |
|
- zh |
|
license: apache-2.0 |
|
metrics: |
|
- accuracy |
|
pipeline_tag: text-generation |
|
--- |
|
|
|
>> It's not a chat model, just using Wizard-LM-Chinese-instruct-evol datesets training with several steps for test the model typical Chinese skill， |
|
>> this is version1, will release version2 for more long context windows and Chat model |
|
>>____________________________ |
|
>>Train scenario: |
|
|
|
>>2k context |
|
|
|
>>datasets:Wizard-LM-Chinese-instruct-evol |
|
|
|
>>batchsize:8 |
|
|
|
>>steps:500 |
|
|
|
>>epchos:2 |
|
>>____________________________________________________ |
|
>>How to use? |
|
|
|
>>Follow common huggingface-api is enough or using other framework like VLLM, support continue training. |
|
|
|
____________________________________________________ |
|
|
|
>>import transformers |
|
>>import torch |
|
|
|
>>model_id = "BoyangZ/llama3-chinese" |
|
|
|
>>pipeline = transformers.pipeline( |
|
>> "text-generation", model=model_id, model_kwargs={"torch_dtype": torch.bfloat16}, device_map="auto" |
|
>> ) |
|
>>pipeline("川普和拜登谁能赢得大选??") |
|
|
|
|
|
>> [{'generated_text': '川普和拜登谁能赢得大选?](https://www.voachinese.com'}] |
|
>> |
|
>> import torch |
|
>> from transformers import AutoModelForCausalLM, AutoTokenizer |
|
>> torch.set_default_device("cuda") |
|
>> model = AutoModelForCausalLM.from_pretrained("BoyangZ/llama3-chinese", torch_dtype="auto", trust_remote_code=True) |
|
>> tokenizer = AutoTokenizer.from_pretrained("BoyangZ/llama3-chinese", trust_remote_code=True) |
|
>> inputs = tokenizer( |
|
>> "川普和拜登一起竞选，美国总统,谁获胜的几率大,分析一下?", |
|
>> return_tensors="pt", |
|
>> return_attention_mask=False |
|
>> ) |
|
>> outputs = model.generate(**inputs, max_length=200) |
|
>> text = tokenizer.batch_decode(outputs)[0] |
|
>> print(text) |
|
|
|
|
|
|
|
>>Wechat:18618377979, Gmail:zhouboyang1983@gmail.com |
