Index-1.9B

README_zh | Online: Chat and Role-playing | QQ: QQ Group

Recent Updates

1. Adapted to llamacpp and Ollama, see Index-1.9B-Chat-GGUF
2. Open source Checkpoint before Decay available for research, see Index-1.9B-Constant-LR

Model Introduction

The Index-1.9B series is a lightweight version of the Index series models, including the following models:

• Index-1.9B base: Base model with 1.9 billion non-embedding parameters, pre-trained on 2.8T mainly Chinese and English corpus, leading in multiple evaluation benchmarks compared to models of the same level.
• Index-1.9B pure: Control version of the base model with the same parameters and training strategy, but strictly filtered out all instruction-related data from the corpus to verify the impact of instructions on benchmarks.
• Index-1.9B chat: A dialogue model aligned with SFT and DPO based on the index-1.9B base. We found that due to the introduction of a lot of internet community corpus in our pre-training, the model has significantly more interesting chat capabilities and strong multilingual (especially East Asian languages) translation abilities compared to models of the same level.
• Index-1.9B character: Introduces RAG on top of SFT and DPO to achieve fewshots role-playing customization.

Evaluation Results

Evaluation code is based on OpenCompass with compatibility modifications. See the evaluate folder for details.

Model Download

Usage Instructions

Environment Setup

1. Download this repository:

git clone https://github.com/bilibili/Index-1.9B
cd Index-1.9B

2. Install dependencies using pip:

pip install -r requirements.txt

Loading with Transformers

You can load the Index-1.9B-Chat model for dialogue using the following code:

import argparse
from transformers import AutoTokenizer, pipeline

# Attention! The directory must not contain "." and can be replaced with "_".
parser = argparse.ArgumentParser()
parser.add_argument('--model_path', default="./IndexTeam/Index-1.9B-Chat/", type=str, help="")
parser.add_argument('--device', default="cpu", type=str, help="") # also could be "cuda" or "mps" for Apple silicon
args = parser.parse_args()

tokenizer = AutoTokenizer.from_pretrained(args.model_path, trust_remote_code=True)
generator = pipeline("text-generation",
                    model=args.model_path,
                    tokenizer=tokenizer, trust_remote_code=True, 
                    device=args.device)

system_message = "你是由哔哩哔哩自主研发的大语言模型，名为“Index”。你能够根据用户传入的信息，帮助用户完成指定的任务，并生成恰当的、符合要求的回复。"
query = "续写 天不生我金坷垃"
model_input = []
model_input.append({"role": "system", "content": system_message})
model_input.append({"role": "user", "content": query})

model_output = generator(model_input, max_new_tokens=300, top_k=5, top_p=0.8, temperature=0.3, repetition_penalty=1.1, do_sample=True)

print('User:', query)
print('Model:', model_output)

Web Demo

Depends on Gradio, install with:

pip install gradio==4.29.0

Start a web server with the following code. After entering the access address in the browser, you can use the Index-1.9B-Chat model for dialogue:

python demo/web_demo.py --port='port' --model_path='/path/to/model/'

Terminal Demo

Start a terminal demo with the following code to use the Index-1.9B-Chat model for dialogue:

python demo/cli_demo.py  --model_path='/path/to/model/'

Openai Api Demo

Depends on Flask, install with:

pip install flask==2.2.5

Start a Flask API with the following code:

python demo/openai_demo.py --model_path='/path/to/model/'

You can conduct dialogues via command line:

curl http://127.0.0.1:8010/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
    "messages": [
    {"role": "system", "content": "你是由哔哩哔哩自主研发的大语言模型，名为“Index”。你能够根据用户传入的信息，帮助用户完成指定的任务，并生成恰当的、符合要求的回复。"},
    {"role": "user", "content": "花儿为什么这么红？"}
    ]
    }'

Index-1.9B-Chat Output Examples

• Below are some examples using web_demo.py to get Index-1.9B-Chat outputs.
• Change the System Message to role-play a stereotype of bilibili user!
• Translate Chinese to Japanese
• Translate Japanese to Chinese

Role Playing

We have simultaneously open-sourced the role-playing model and the accompanying framework.

• We currently have the character 三三 built-in.
• If you need to create your own character, please prepare a dialogue corpus similar to roleplay/character/三三.csv (note that the file name should match the name of the character you want to create) and a corresponding character description. Click 生成角色 to create it successfully.
• If the corresponding character has already been created, please directly enter the character you want to converse with in the Role name field, input your query, and click submit to start the conversation.

For detailed usage, please refer to the roleplay folder.

Quantization

Depends on bitsandbytes, installation command:

pip install bitsandbytes==0.43.0

You can use the following script to perform int4 quantization, which has less performance loss and further saves video memory usage.

import torch
import argparse
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    TextIteratorStreamer,
    GenerationConfig,
    BitsAndBytesConfig
)

parser = argparse.ArgumentParser()
parser.add_argument('--model_path', default="", type=str, help="")
parser.add_argument('--save_model_path', default="", type=str, help="")
args = parser.parse_args()

tokenizer = AutoTokenizer.from_pretrained(args.model_path, trust_remote_code=True)
quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.float16,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
    llm_int8_threshold=6.0,
    llm_int8_has_fp16_weight=False,
)
model = AutoModelForCausalLM.from_pretrained(args.model_path, 
                                             device_map="auto",
                                             torch_dtype=torch.float16,
                                             quantization_config=quantization_config,
                                             trust_remote_code=True)
model.save_pretrained(args.save_model_path)
tokenizer.save_pretrained(args.save_model_path)

Fine-tuning

Follow the steps in the fine-tuning tutorial to quickly fine-tune the Index-1.9B-Chat model. Give it a try and customize your exclusive Index model!

Limitations and Disclaimer

Index-1.9B may generate inaccurate, biased, or otherwise objectionable content in certain situations. The model cannot understand, express personal opinions, or make value judgments. Its outputs do not represent the views and positions of the model developers. Therefore, please use the generated content with caution. Users should independently evaluate and verify the content generated by the model and should not disseminate harmful content. Developers should conduct safety tests and fine-tuning according to specific applications before deploying any related applications.

We strongly advise against using these models to create or disseminate harmful information or engage in activities that may harm public, national, or social security or violate regulations. Do not use the models for internet services without proper safety review and filing. We have made every effort to ensure the compliance of the training data, but due to the complexity of the model and data, unforeseen issues may still exist. We will not be held responsible for any problems arising from the use of these models, whether related to data security, public opinion risks, or any risks and issues caused by misunderstanding, misuse, dissemination, or non-compliant use of the model.

Model Open Source License

Using the source code from this repository requires compliance with the Apache-2.0. The use of the Index-1.9B model weights requires compliance with the INDEX_MODEL_LICENSE.

The Index-1.9B model weights are fully open for academic research and support free commercial use.

Citation

If you think our work is helpful to you, please feel free to cite it!

@article{Index,
  title={Index1.9B Technical Report},
  year={2024}
}

Extended Works

libllm: https://github.com/ling0322/libllm/blob/main/examples/python/run_bilibili_index.py

chatllm.cpp：https://github.com/foldl/chatllm.cpp/blob/master/docs/rag.md#role-play-with-rag

ollama：https://ollama.com/milkey/bilibili-index

self llm: https://github.com/datawhalechina/self-llm/blob/master/bilibili_Index-1.9B/04-Index-1.9B-Chat%20Lora%20微调.md