Edit model card

roberta-zh fine-tuned on human-annotated conversational model self-chat data. It supports 2-class calssification for multi-turn dialogue sensible detection. Usage example:

NOTE: it should be used under similar data distribution.

import torch
from transformers import BertTokenizer, BertForSequenceClassification

tokenizer = BertTokenizer.from_pretrained('thu-coai/roberta-zh-specific')
model = BertForSequenceClassification.from_pretrained('thu-coai/roberta-zh-specific', num_labels=2)
model.eva()

context = [
    "你大爱的冷门古诗词是什么?\t一枝红艳露凝香,云雨巫山枉断肠",
    "你大爱的冷门古诗词是什么?\t一枝红艳露凝香,云雨巫山枉断肠",
]

response = [
    "我也很喜欢,我觉得这句的意境很美",
    "我也很喜欢",
]

model_input = tokenizer(context, response, return_tensors='pt', padding=True)
with torch.no_grad():
    model_output = model(**model_input, return_dict=True)
logits = model_output.logits
preds_all = torch.argmax(logits, dim=-1).cpu()
print(preds_all) # 1 for specific response else 0

Downloads last month
37
Safetensors
Model size
102M params
Tensor type
I64
·
F32
·