BelleGroup/BELLE-7B-1M

Model Card for Model ID

Model description

BELLE is based on Bloomz-7b1-mt and finetuned with 1M Chinese data combined with 50,000 pieces of English data from the open source Stanford-Alpaca, resulting in good Chinese instruction understanding and response generation capabilities.

The code of Chinese data generation and other detailed information can be found in our Github project repository: https://github.com/LianjiaTech/BELLE.

We trained models using datasets of different sizes (200,000, 600,000, and 1,000,000 samples) for instruction learning, and we obtained different model versions as shown below: | Datasize| 200,000 | 600,000 | 1,000,000 | | ----- | ----- | ----- | ----- | | Finetuned Model | BELLE-7B-0.2M | BELLE-7B-0.6M | BELLE-7B-1M |

Training hyper-parameters

Use model

Please note that the input should be formatted as follows in both training and inference.

Human: {input} \n\nAssistant:

BELLE can be easily loaded with AutoModelForCausalLM.

from transformers import AutoTokenizer, AutoModelForCausalLM
import sys

model_path = "./" # You can modify the path for storing the local model
model =  AutoModelForCausalLM.from_pretrained(model_path)
tokenizer = AutoTokenizer.from_pretrained(model_path)
print("Human:")
line = input()
while line:
        inputs = 'Human: ' + line.strip() + '\n\nAssistant:'
        input_ids = tokenizer(inputs, return_tensors="pt").input_ids
        outputs = model.generate(input_ids, max_new_tokens=200, do_sample = True, top_k = 30, top_p = 0.85, temperature = 0.35, repetition_penalty=1.2)
        rets = tokenizer.batch_decode(outputs, skip_special_tokens=True)
        print("Assistant:\n" + rets[0].strip().replace(inputs, ""))
        print("\n------------------------------------------------\nHuman:")
        line = input()

The above example did not utilize the context from previous turns of the conversation. If you want to perform multi-turn dialogue understanding, you can concatenate the previous turns of the conversation in the format of Human:{input}\n\nAssistant:{output} and input it to the generate function

Limitations

There still exists a few issues in the model trained on current base model and data:

1. 1.The model might generate factual errors when asked to follow instructions related to facts.

1. 1.Occasionally generates harmful responses since the model still struggles to identify potential harmful instructions.

1. 1.Needs improvements on reasoning and coding.

Since the model still has its limitations, we require developers only use the open-sourced code, data, model and any other artifacts generated via this project for research purposes. Commercial use and other potential harmful use cases are not allowed.

Some use cases

模型描述

BELLE模型以Bloomz-7b1-mt为基础，在 1.0M 条中文数据上，结合Stanford-Alpaca开源的5万条英文数据上训练，取得了较好的中文instruction理解以及response生成能力。

中文数据生成代码及其他详细信息，位于我们的Github项目仓库: https://github.com/LianjiaTech/BELLE

我们采取了不同大小规模（20万、60万和100万样本）的指令学习的数据集训练模型，我们得到不同的模型版本如下所示: | Datasize| 200,000 | 600,000 | 1,000,000 | | ----- | ----- | ----- | ----- | | Finetuned Model | BELLE-7B-0.2M | BELLE-7B-0.6M | BELLE-7B-1M |

模型训练超参数

使用模型

请注意，在训练和推理时，模型的输入应该处理成如下形式：

Human: {input} \n\nAssistant:

通过AutoModelForCausalLM即可直接载入模型并使用。

from transformers import AutoTokenizer, AutoModelForCausalLM
import sys

model_path = "./" # You can modify the path for storing the local model
model =  AutoModelForCausalLM.from_pretrained(model_path)
tokenizer = AutoTokenizer.from_pretrained(model_path)
print("Human:")
line = input()
while line:
        inputs = 'Human: ' + line.strip() + '\n\nAssistant:'
        input_ids = tokenizer(inputs, return_tensors="pt").input_ids
        outputs = model.generate(input_ids, max_new_tokens=200, do_sample = True, top_k = 30, top_p = 0.85, temperature = 0.35, repetition_penalty=1.2)
        rets = tokenizer.batch_decode(outputs, skip_special_tokens=True)
        print("Assistant:\n" + rets[0].strip().replace(inputs, ""))
        print("\n------------------------------------------------\nHuman:")
        line = input()

上面样例没有利用对话上文，如果你想做多轮对话理解，可以将对话上文按照Human:{input}\n\nAssistant:{output} 方式拼接起来输入给generate函数

局限性和使用限制

基于当前数据和基础模型训练得到的SFT模型，在效果上仍存在以下问题：

1. 1.在涉及事实性的指令上可能会产生违背事实的错误回答。

1. 1.对于具备危害性的指令无法很好的鉴别，由此会产生危害性言论。

1. 1.在一些涉及推理、代码等场景下模型的能力仍有待提高。

基于以上模型局限性，我们要求开发者仅将我们开源的代码、数据、模型及后续用此项目生成的衍生物用于研究目的，不得用于商业，以及其他会对社会带来危害的用途。

一些示例