Langboat_bloom-6b4-zh-instruct_finetune-chat

是基于Langboat_bloom-6b4-zh模型，在firefly-train-1.1M和Belle-train_2m_cn数据集上采用的QLoRA方法微调的对话模型。
在CEVAL上的评测结果：

STEM	Social Sciences	Humanities	Others	Average	AVG(Hard)
27.9	27.2	24.8	26.4	26.8	28.0

使用

单轮指令生成

from transformers import AutoTokenizer, AutoModelForCausalLM

device = "cuda"
model = AutoModelForCausalLM.from_pretrained("SmilePanda/Langboat_bloom-6b4-zh-instruct_finetune-chat", device_map=device)
tokenizer = AutoTokenizer.from_pretrained("SmilePanda/Langboat_bloom-6b4-zh-instruct_finetune-chat", use_fast=False)

source_prefix = "human"
target_prefix = "assistant"
query = "你好"
sentence = f"{source_prefix}: \n{query}\n\n{target_prefix}: \n"
print("query: ", sentence)
input_ids = tokenizer(sentence, return_tensors='pt').input_ids.to(device)
outputs = model.generate(input_ids=input_ids, max_new_tokens=500,
                         do_sample=True,
                         top_p=0.8,
                         temperature=0.35,
                         repetition_penalty=1.2,
                         eos_token_id=tokenizer.eos_token_id)
rets = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0].strip()
response = rets.replace(sentence, "")
print(response)

多轮对话

import os
from transformers import AutoTokenizer, AutoModelForCausalLM

device = "cuda"
model = AutoModelForCausalLM.from_pretrained("SmilePanda/Langboat_bloom-6b4-zh-instruct_finetune-chat", device_map=device)
tokenizer = AutoTokenizer.from_pretrained("SmilePanda/Langboat_bloom-6b4-zh-instruct_finetune-chat", use_fast=False)

source_prefix = "human"
target_prefix = "assistant"

history = ""

while True:
    query = input("user: ").strip()
    if not query:
        continue
    if query == 'q' or query == 'stop':
        break
    if history:
        sentence = history + f"\n{source_prefix}: \n{query}\n\n{target_prefix}: \n"
    else:
        sentence = f"{source_prefix}: \n{query}\n\n{target_prefix}: \n"
    input_ids = tokenizer(sentence, return_tensors='pt').input_ids.to(device)
    outputs = model.generate(input_ids=input_ids, max_new_tokens=1024,
                             do_sample=True,
                             top_p=0.90,
                             temperature=0.1,
                             repetition_penalty=1.0,
                             eos_token_id=tokenizer.eos_token_id)
    rets = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0].strip()
    print("bloom: {}".format(rets.replace(sentence, "")))
    history = rets