语言:
📝 描述
基于俄语训练并在我个人电报聊天记录上微调的DialoGPT模型。
该模型由Sberbank-AI创建,并在俄语论坛数据上训练(参见Grossmend的模型)。训练细节可查阅habr文章(俄语)。我建立了简易处理流程,并基于个人导出的电报聊天记录(约30MB的json文件)对该模型进行了微调。实际上从电报获取数据并微调模型非常简单,为此我制作了Colab教程: https://colab.research.google.com/drive/1fnAVURjyZRK9VQg1Co_-SKUQnRES8l9R?usp=sharing
⚠️ 由于数据特殊性,托管推理API可能无法正常工作 ⚠️
🤗试用请访问我的Spaces演示🤗
❓ 代码调用方法
checkpoint = "Kirili4ik/ruDialoGpt3-medium-finetuned-telegram"
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForCausalLM.from_pretrained(checkpoint)
model.eval()
def get_length_param(text: str, tokenizer) -> str:
tokens_count = len(tokenizer.encode(text))
if tokens_count <= 15:
len_param = '1'
elif tokens_count <= 50:
len_param = '2'
elif tokens_count <= 256:
len_param = '3'
else:
len_param = '-'
return len_param
def get_user_param(text: dict, machine_name_in_chat: str) -> str:
if text['from'] == machine_name_in_chat:
return '1'
else:
return '0'
chat_history_ids = torch.zeros((1, 0), dtype=torch.int)
while True:
next_who = input("发言者?\t")
if next_who == "H" or next_who == "Human":
input_user = input("===> 人类: ")
new_user_input_ids = tokenizer.encode(f"|0|{get_length_param(input_user, tokenizer)}|" \
+ input_user + tokenizer.eos_token, return_tensors="pt")
chat_history_ids = torch.cat([chat_history_ids, new_user_input_ids], dim=-1)
if next_who == "G" or next_who == "GPT":
next_len = input("语句长度?1/2/3/-\t")
new_user_input_ids = tokenizer.encode(f"|1|{next_len}|", return_tensors="pt")
chat_history_ids = torch.cat([chat_history_ids, new_user_input_ids], dim=-1)
input_len = chat_history_ids.shape[-1]
chat_history_ids = model.generate(
chat_history_ids,
num_return_sequences=1,
max_length=512,
no_repeat_ngram_size=3,
do_sample=True,
top_k=50,
top_p=0.9,
temperature = 0.6,
mask_token_id=tokenizer.mask_token_id,
eos_token_id=tokenizer.eos_token_id,
unk_token_id=tokenizer.unk_token_id,
pad_token_id=tokenizer.pad_token_id,
device='cpu'
)
print(f"===> GPT-3回复: {tokenizer.decode(chat_history_ids[:, input_len:][0], skip_special_tokens=True)}")