InsTag基于Qwen-1_8B训练的中英文指令标签标注器

使用方法

参考通义千问官方repo使用vllm+fastchat进行推理，离线推理方法示例如下：

from vllm import LLM, SamplingParams
from fastchat.llm_judge.common import load_questions, temperature_config
from fastchat.model import load_model, get_conversation_template
import torch

model_id = "qwen-1p8b-tagger"
model_path = "/path/to/model/"

# Load model with vLLM
model = LLM(model=model_path, trust_remote_code=True)
# Setting greedy decoding with temperature=0
sampling_params = SamplingParams(temperature=0, max_tokens=512)


torch.manual_seed(0)
prompts = []
conv = get_conversation_template(model_id)
conv.append_message(conv.roles[0], "hi who are you")
conv.append_message(conv.roles[1], None)
prompt = conv.get_prompt()
prompts.append(prompt)

outputs = model.generate(prompts, sampling_params)
for output in outputs:
    output_ids = output.outputs[0].token_ids

    # be consistent with the template's stop_token_ids
    if conv.stop_token_ids:
        stop_token_ids_index = [
            i
            for i, id in enumerate(output_ids)
            if id in conv.stop_token_ids
        ]
        if len(stop_token_ids_index) > 0:
            output_ids = output_ids[: stop_token_ids_index[0]]

    output = model.get_tokenizer().decode(
        output_ids,
        spaces_between_special_tokens=False,
    )
    if conv.stop_str and output.find(conv.stop_str) > 0:
        output = output[: output.find(conv.stop_str)]
    for special_token in model.get_tokenizer().special_tokens_map.values():
        if isinstance(special_token, list):
            for special_tok in special_token:
                output = output.replace(special_tok, "")
        else:
            output = output.replace(special_token, "")
    output = output.strip()
    print(output)

Clone with HTTP

 git clone https://www.modelscope.cn/lukeminglkm/instagger_qwen1_8B.git