Skip to content

Qihoo360/360zhinao

Repository files navigation

中文 |   English 


360Zhinao (360智脑)

🤗 HuggingFace   |    🤖 ModelScope   |    💬 WeChat (微信)   |    📑 Technical Report  

Feel free to visit 360Zhinao's official website https://ai.360.com for more experience.


Introduction

🎉🎉🎉 We released the 360Zhinao model series:

  • 360Zhinao-7B-Base
  • 360Zhinao-7B-Chat-4K
  • 360Zhinao-7B-Chat-32K
  • 360Zhinao-7B-Chat-360K
  • 360Zhinao-search
  • 360Zhinao-1.8B-Reranking

Notable features of our 360Zhinao models are:

  • Base Model: Leveraging a high-quality corpus of 3.4 trillion tokens consisting of mainly Chinese, English and code, we achieved competitive performance on relevant benchmarks against other 7B models.
  • Chat Models: Powerful chat capabilities and three context lengths of 4K, 32K and 360K. 360K (around 500k Chinese characters) is the longest context length among Chinese open-sourced models upon release (Apr. 11, 2024).

News and Updates

  • [2024.05.23] We released two models, 360Zhinao-search and 360Zhinao-1.8B-Reranking, which ranked first respectively in the Retrieval and Reranking tasks of C-MTEB Leaderboard .
  • [2024.05.20] We extended llama3 and released llama3-8B-360Zhinao-360k-Instruct🤗 Details here.
  • [2024.04.12] We released 360Zhinao-7B v1.0, including the base model and three chat models with context lengths 4K, 32K and 360K. Technical report is here and on arXiv.

Table of contents


Download URL

Size Model BF16 Int4
7B 360Zhinao-7B-Base 🤖 🤗
7B 360Zhinao-7B-Chat-4K 🤖 🤗 🤖 🤗
7B 360Zhinao-7B-Chat-32K 🤖 🤗 🤖 🤗
7B 360Zhinao-7B-Chat-360K 🤖 🤗 🤖 🤗
325M 360Zhinao-search 🤗
1.8B 360Zhinao-1.8B-Reranking 🤗

Model Evaluation

Base Model

We evaluate our model on OpenCompass, more specifically on C-Eval, AGIEval, MMLU, CMMLU, HellaSwag, MATH, GSM8K, HumanEval, MBPP, BBH and LAMBADA. These benchmarks test the model on natural language understanding, knowledge, mathematics, code generation and logical reasoning, etc.

Results are listed as follows and could be viewed or reproduced on OpenCompass leaderboard.

Model
AVG CEval AGIEval MMLU CMMLU HellaSwag MATH GSM8K HumanEval MBPP BBH LAMBADA
Baichuan2-7B 41.49 56.3 34.6 54.7 57 67 5.4 24.6 17.7 24 41.8 73.3
Baichuan-7B 31.94 44.7 24.6 41.5 44.6 68.4 2.5 9.6 9.1 6.4 32.8 67.1
ChatGLM3-6B 58.67 67 47.4 62.8 66.5 76.5 19.2 61 44.5 57.2 66.2 77.1
DeepSeek-7B 39.8 45 24 49.3 46.8 73.4 4.2 18.3 25 36.4 42.8 72.6
InternLM2-7B 58.01 65.7 50.2 65.5 66.2 79.6 19.9 70.6 41.5 42.4 64.4 72.1
InternLM-7B 39.33 53.4 36.9 51 51.8 70.6 6.3 31.2 13.4 14 37 67
LLaMA-2-7B 33.27 32.5 21.8 46.8 31.8 74 3.3 16.7 12.8 14.8 38.2 73.3
LLaMA-7B 30.35 27.3 20.6 35.6 26.8 74.3 2.9 10 12.8 16.8 33.5 73.3
Mistral-7B-v0.1 47.67 47.4 32.8 64.1 44.7 78.9 11.3 47.5 27.4 38.6 56.7 75
MPT-7B 30.06 23.5 21.3 27.5 25.9 75 2.9 9.1 17.1 22.8 35.6 70
Qwen1.5-7B 55.12 73.57 50.8 62.15 71.84 72.62 20.36 54.36 53.05 36.8 40.01 70.74
Qwen-7B 49.53 63.4 45.3 59.7 62.5 75 13.3 54.1 27.4 31.4 45.2 67.5
XVERSE-7B 34.27 61.1 39 58.4 60.8 73.7 2.2 11.7 4.9 10.2 31 24
Yi-6B 47.8 73 44.3 64 73.5 73.1 6.3 39.9 15.2 23.6 44.9 68
360Zhinao-7B 56.15 74.11 49.49 67.44 72.38 83.05 16.38 53.83 35.98 42.4 43.95 78.59

Chat Models

The 4K and 32K models are trained separately with the same 4K SFT data.

To train the long-context models, we adopted a two-stage approach.

First stage: We increased RoPE base and extended the context length to 32K.

  • Firstly, we performed Continual Pretraining on approximately 5B tokens with a 32K context window.
  • Then during the SFT stage, we finetuned the model using long data from various sources, including high-quality human-labeled 32K data.

Second stage: We extended the context length to 360K, training with the following data:

  • A small amount of high-quality human-labeled super-long data.
  • Due to the scarcity of annotated super-long data, we constructed various forms of synthetic data.
    • Multi-Doc QA: Similar to Ziya-Reader, we generated multi-document QA pairs based on 360's database. Multiple QA pairs are constructed for one row of Multi-Doc QA data input, resulting in a multi-turn format and significantly improving the training efficiency.
    • Single-Doc QA: Similar to LLama2 Long, we constructed multi-turn QA data based on different segments within one row of long-text input.

We evaluated our models across various lengths and benchmarks.

  • Long Context Benchmarks

    We evaluated our 32K and 360K models on LongBench, a multi-task bilingual benchmark for long contexts. We report results on Chinese tasks most relevant to downstream applications: Single/Multi-Doc QA, Summarization, Few-Shot Learning and Code Completion.

    Model Avg Single-Doc QA Multi-Doc QA Summarization Few-Shot Learning Code Completion
    GPT-3.5-Turbo-16k 37.84 61.2 28.7 16 29.2 54.1
    ChatGLM2-6B-32k 37.16 51.6 37.6 16.2 27.7 52.7
    ChatGLM3-6B-32k 44.62 62.3 44.8 17.8 42 56.2
    InternLM2-Chat-7B 42.20 56.65 29.15 17.99 43.5 63.72
    Qwen1.5-Chat-7B 36.75 52.85 30.08 14.28 32 54.55
    Qwen1.5-Chat-14B 39.80 60.39 27.99 14.77 37 58.87
    360Zhinao-7B-Chat-32K 45.18 57.18 48.06 15.03 44 61.64
  • 360Zhinao-7B-Chat-360K on "NeedleInAHaystack"

    NeedleInAHaystack places one small piece of information in different positions of long text and queries this information as a test of LLM's long-context capabilities.

    360Zhinao-7B-Chat-360K could achieve over 98% accuracy on both English and Chinese NeedleInAHaystack tasks.

    • English version(same as NeedleInAHaystack

      needle:The best thing to do in San Francisco is eat a sandwich and sit in Dolores Park on a sunny day.

      query:What is the best thing to do in San Francisco?

    • Chinese version

      We constructed the Chinese version following the SuperCLUE-200K benchmark:

      haystack:Chinese novels.

      needle:(in Chinese) 王莽是一名勤奋的店员,他每天凌晨就起床,赶在第一缕阳光照亮大地之前到达店铺,为即将开始的一天做准备。他清扫店铺,整理货架,为顾客提供方便。他对五金的种类和用途了如指掌,无论顾客需要什么,他总能准确地找到。\n然而,他的老板刘秀却总是对他吹毛求疵。刘秀是个挑剔的人,他总能在王莽的工作中找出一些小错误,然后以此为由扣他的工资。他对王莽的工作要求非常严格,甚至有些过分。即使王莽做得再好,刘秀也总能找出一些小问题,让王莽感到非常沮丧。\n王莽虽然对此感到不满,但他并没有放弃。他知道,只有通过自己的努力,才能获得更好的生活。他坚持每天早起,尽管他知道那天可能会再次被刘秀扣工资。他始终保持微笑,尽管他知道刘秀可能会再次对他挑剔。

      query:(in Chinese) 王莽在谁的手下工作?


Quickstart

We provide simple examples illustrating the use of 360Zhinao-7B-Base and 360Zhinao-7B-Chat on 🤖ModelScope and 🤗Transformers.

Dependency Installation

  • python >= 3.8
  • pytorch >= 2.0
  • transformers >= 4.37.2
  • CUDA >= 11.4
pip install -r requirements.txt 

Optionally, we recommend installing Flash-Attention 2 to improve performance and reduce memory footprint.

flash-attn >= 2.3.6

FLASH_ATTENTION_FORCE_BUILD=TRUE pip install flash-attn==2.3.6

🤗 Transformers

Demonstration of Base Model Inference

from transformers import AutoTokenizer, AutoModelForCausalLM
from transformers.generation import GenerationConfig

MODEL_NAME_OR_PATH = "qihoo360/360Zhinao-7B-Base"

tokenizer = AutoTokenizer.from_pretrained(
    MODEL_NAME_OR_PATH, 
    trust_remote_code=True)

model = AutoModelForCausalLM.from_pretrained(
    MODEL_NAME_OR_PATH,
    device_map="auto",
    trust_remote_code=True)

generation_config = GenerationConfig.from_pretrained(
    MODEL_NAME_OR_PATH,
    trust_remote_code=True)

inputs = tokenizer('中国二十四节气\n1. 立春\n2. 雨水\n3. 惊蛰\n4. 春分\n5. 清明\n', return_tensors='pt')
inputs = inputs.to(model.device)

pred = model.generate(input_ids=inputs["input_ids"], generation_config=generation_config)
print("outputs:\n", tokenizer.decode(pred.cpu()[0], skip_special_tokens=True))

Demonstration of Chat Model Inference

from transformers import AutoTokenizer, AutoModelForCausalLM
from transformers.generation import GenerationConfig

MODEL_NAME_OR_PATH = "qihoo360/360Zhinao-7B-Chat-4K"

tokenizer = AutoTokenizer.from_pretrained(
    MODEL_NAME_OR_PATH, 
    trust_remote_code=True)

model = AutoModelForCausalLM.from_pretrained(
    MODEL_NAME_OR_PATH,
    device_map="auto",
    trust_remote_code=True)

generation_config = GenerationConfig.from_pretrained(
    MODEL_NAME_OR_PATH,
    trust_remote_code=True)

messages = []
#round-1
messages.append({"role": "user", "content": "介绍一下刘德华"})
response = model.chat(tokenizer=tokenizer, messages=messages, generation_config=generation_config)
messages.append({"role": "assistant", "content": response})
print(messages)

#round-2
messages.append({"role": "user", "content": "他有什么代表作?"})
response = model.chat(tokenizer=tokenizer, messages=messages, generation_config=generation_config)
messages.append({"role": "assistant", "content": response})
print(messages)

🤖 ModelScope

Demonstration of Base Model Inference

from modelscope import AutoModelForCausalLM, AutoTokenizer
from modelscope import GenerationConfig

MODEL_NAME_OR_PATH = "qihoo360/360Zhinao-7B-Base"

tokenizer = AutoTokenizer.from_pretrained(
    MODEL_NAME_OR_PATH, 
    trust_remote_code=True)

model = AutoModelForCausalLM.from_pretrained(
    MODEL_NAME_OR_PATH,
    device_map="auto",
    trust_remote_code=True)

generation_config = GenerationConfig.from_pretrained(
    MODEL_NAME_OR_PATH,
    trust_remote_code=True)

inputs = tokenizer('中国二十四节气\n1. 立春\n2. 雨水\n3. 惊蛰\n4. 春分\n5. 清明\n', return_tensors='pt')
inputs = inputs.to(model.device)

pred = model.generate(input_ids=inputs["input_ids"], generation_config=generation_config)
print("outputs:\n", tokenizer.decode(pred.cpu()[0], skip_special_tokens=True))

Demonstration of Chat Model Inference

from modelscope import AutoModelForCausalLM, AutoTokenizer
from modelscope import GenerationConfig

MODEL_NAME_OR_PATH = "qihoo360/360Zhinao-7B-Chat-4K"

tokenizer = AutoTokenizer.from_pretrained(
    MODEL_NAME_OR_PATH, 
    trust_remote_code=True)

model = AutoModelForCausalLM.from_pretrained(
    MODEL_NAME_OR_PATH,
    device_map="auto",
    trust_remote_code=True)

generation_config = GenerationConfig.from_pretrained(
    MODEL_NAME_OR_PATH,
    trust_remote_code=True)

messages = []
#round-1
messages.append({"role": "user", "content": "介绍一下刘德华"})
response = model.chat(tokenizer=tokenizer, messages=messages, generation_config=generation_config)
messages.append({"role": "assistant", "content": response})
print(messages)

#round-2
messages.append({"role": "user", "content": "他有什么代表作?"})
response = model.chat(tokenizer=tokenizer, messages=messages, generation_config=generation_config)
messages.append({"role": "assistant", "content": response})
print(messages)

CLI Demo

Use terminal for command-line interface:

python cli_demo.py

Note: for Mac users, device = 'mps' is not supported yet.

Web Demo

streamlit run web_demo.py

API Demo

Launch api:

python openai_api.py

Then request with parameters:

curl 'http://localhost:8360/v1/chat/completions' \
-H 'Content-Type: application/json' \
-d '{
    "max_new_tokens": 200,
    "do_sample": true,
    "top_k": 0,
    "top_p": 0.8,
    "temperature": 1.0,
    "repetition_penalty": 1.0,
    "messages": [
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "你好"}
    ]
}'

Model Inference

Quantization

We provide quantization schemes based on AutoGPTQ and release the Int4 quantization models.

Deployment

vLLM Installation

We recommend using vLLM==0.3.3.

If you are using CUDA 12.1 and PyTorch 2.1, you can install vLLM directly with:

pip install vllm==0.3.3

Otherwise, please refer to the official vLLM Installation Instructions.

After installation, perform the following steps:

  1. Copy vllm/zhinao.py into vllm/model_executor/models in your vllm installation directory (in python/conda env).

  2. Copy vllm/serving_chat.py into vllm/entrypoints/openai in your vllm installation directory.

  3. Then add a line in vllm/model_executor/models/__init__.py

    "ZhinaoForCausalLM": ("zhinao", "ZhinaoForCausalLM"),

vLLM Service Start

Start the service:

python -m vllm.entrypoints.openai.api_server \
    --served-model-name 360Zhinao-7B-Chat-4K \
    --model qihoo360/360Zhinao-7B-Chat-4K \
    --trust-remote-code \
    --tensor-parallel-size 1 \
    --max-model-len 4096 \
    --host 0.0.0.0 \
    --port 8360

Use curl to request the service:

curl http://localhost:8360/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
    "model": "360Zhinao-7B-Chat-4K",
    "max_tokens": 200,
    "top_k": -1,
    "top_p": 0.8,
    "temperature": 1.0,
    "presence_penalty": 0.0,
    "frequency_penalty": 0.0,
    "messages": [
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "你好"}
    ],
    "stop": [
        "<eod>",
        "<|im_end|>",
        "<|im_start|>"
    ]
}'

Use python to request the service:

from openai import OpenAI
openai_api_key = "EMPTY"
openai_api_base = "http://localhost:8360/v1"

client = OpenAI(
    api_key=openai_api_key,
    base_url=openai_api_base,
)

chat_response = client.chat.completions.create(
    model="360Zhinao-7B-Chat-4K",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "你好"},
    ],
    stop=[
        "<eod>",
        "<|im_end|>",
        "<|im_start|>"
    ],
    presence_penalty=0.0,
    frequency_penalty=0.0
)
print("Chat response:", chat_response)

If you need to enable repetition penalty, we recommend setting presence_penalty and frequency_penalty instead of repetition_penalty.


Model Finetune

Training data

Training Data: data/training_data_sample.json. This example data has 10,000 rows sampled from multiturn_chat_0.8M with converted format.

Data Format:

[
  {
    "id": 1,
    "conversations": [
        {
            "from": "system",
            "value": "You are a helpful assistant."
        },
        {
            "from": "user",
            "value": "您好啊"
        },
        {
            "from": "assistant",
            "value": "你好!我今天能为您做些什么?有什么问题或需要帮助吗? 我在这里为您提供服务。"
        }
    ]
  }
]

Finetuning scripts

set -x

HOSTFILE=hostfile
DS_CONFIG=./finetune/ds_config_zero2.json

# PARAMS
LR=5e-6
EPOCHS=3
MAX_LEN=4096
BATCH_SIZE=4
NUM_NODES=1
NUM_GPUS=8
MASTER_PORT=29500

IS_CONCAT=False # Whether to concatenate to maximum length (MAX_LEN)

DATA_PATH="./data/training_data_sample.json"
MODEL_PATH="qihoo360/360Zhinao-7B-Base"
OUTPUT_DIR="./outputs/"

deepspeed --hostfile ${HOSTFILE} \
        --master_port ${MASTER_PORT} \
        --num_nodes ${NUM_NODES} \
        --num_gpus ${NUM_GPUS} \
        finetune.py \
        --report_to "tensorboard" \
        --data_path ${DATA_PATH} \
        --model_name_or_path ${MODEL_PATH} \
        --output_dir ${OUTPUT_DIR} \
        --model_max_length ${MAX_LEN} \
        --num_train_epochs ${EPOCHS} \
        --per_device_train_batch_size ${BATCH_SIZE} \
        --gradient_accumulation_steps 1 \
        --save_strategy steps \
        --save_steps 200 \
        --learning_rate ${LR} \
        --lr_scheduler_type cosine \
        --adam_beta1 0.9 \
        --adam_beta2 0.95 \
        --adam_epsilon 1e-8 \
        --max_grad_norm 1.0 \
        --weight_decay 0.1 \
        --warmup_ratio 0.01 \
        --gradient_checkpointing True \
        --bf16 True \
        --tf32 True \
        --deepspeed ${DS_CONFIG} \
        --is_concat ${IS_CONCAT} \
        --logging_steps 1 \
        --log_on_each_node False
bash finetune/ds_finetune.sh
  • Configuring HOSTFILE switches between single-machine and multi-machine training.
  • configuring ds_config switches between zero1, zero2 and zero3.
  • fp16, bf16 could configure mixed precision training. bf16 is recommended to be consistent with the pretrained model.
  • is_concat configures whether the training data is concatenated or not.

360Zhinao-search Model Introduction

360Zhinao-search uses the self-developed BERT model as the base for multi-task fine-tuning, which has an average score of 75.05 on the Retriev al task on the C-MTEB-Retrieval benchmark, currently ranking first. C-MTEB-Retrieval leaderboard contains a total of 8 [query, passage] similarity retrieval sub tasks in different fields, using NDCG@10 (Normalized Discounted Cumulative Gain @ 10) as the evaluation index.

Model T2Retrieval MMarcoRetrieval DuRetrieval CovidRetrieval CmedqaRetrieval EcomRetrieval MedicalRetrieval VideoRetrieval Avg
360Zhinao-search 87.12 83.32 87.57 85.02 46.73 68.9 63.69 78.09 75.05
AGE_Hybrid 86.88 80.65 89.28 83.66 47.26 69.28 65.94 76.79 74.97
OpenSearch-text-hybrid 86.76 79.93 87.85 84.03 46.56 68.79 65.92 75.43 74.41
piccolo-large-zh-v2 86.14 79.54 89.14 86.78 47.58 67.75 64.88 73.1 74.36
stella-large-zh-v3-1792d 85.56 79.14 87.13 82.44 46.87 68.62 65.18 73.89 73.6

Optimization points

  1. Data filtering: Strictly prevent the C-MTEB-Retrieval test data from leaking, and clean all queries and passages in the test set;
  2. Data source enhancement: Use open source data and LLM synthetic data to improve data diversity;
  3. Negative example mining: Use multiple methods to deeply mine difficult-to-distinguish negative examples to improve information gain;
  4. Training efficiency: multi-machine multi-CPU + Deepspeed method to optimize GPU memory utilization.

Environmental requirements

cd Retrieval
pip install -r requirements.txt

Training script

cd Retrieval/finetune
sh train.sh

Inference script

cd Retrieval/eval
python test_model.py

C-MTEB test script

cd Retrieval/eval
sh eval.sh

Reference

bge fine-tuning code C-MTEB official test script

360Zhinao-1.8B-Reranking Model Introduction

The 360Zhinao-1.8B-Reranking model utilizes the self-developed 360Zhinao_1.8B_base model as its foundation. Our self-developed unidirectional generative model, 360Zhinao_1.8B_reranking, achieved an average score of 70.13, currently ranking first overall and first among open-source mo dels, opening up new possibilities for generative models to undertake discriminative tasks.

C-MTEB-Reranking leaderboard contains four subtasks, which are tasks of judging the similari ty of user questions and answers in different fields. It uses MAP (Mean-average-precision) as the evaluation index. Currently, the open-source models on this leaderboard are primarily bidirectional discriminative models (BERT-like models). The only unidirectional generative model (GP T-like model) is gte-Qwen1.5-7B-instruct, which has an average score of 66.38, ranking 25th, with less than ideal results.

Model T2Reranking MMarcoReranking CMedQAv1 CMedQAv2 Avg
360Zhinao-1.8B-Reranking 68.55 37.29 86.75 87.92 70.13
piccolo-large-zh-v2 67.15 33.39 90.14 89.31 70
Baichuan-text-embedding 67.85 34.3 88.46 88.06 69.67
stella-mrl-large-zh-v3.5-1792d 66.43 28.85 89.18 89.33 68.45
PEG 69.43 33.55 86.56 84.09 68.41
bge-reranker-base 67.28 35.46 81.27 84.1 67.03
bge-reranker-large 67.6 37.17 82.14 84.19 67.78

Optimization points

Through iterative discovery and resolution of the following technical issues, it continuously stimulates the world knowledge inherent in the l arge model during the pre-training phase, better bridging the gap between generative models and discriminative tasks.

  1. Data Processing: The model training did not utilize world knowledge, meaning it neither continued pre-training with domain-specific data no r fine-tuned datasets outside of the four datasets on the leaderboard. It only used the four datasets within the leaderboard, carefully iterat ing through data perception, and targeting different datasets for data cleaning and mining to ensure that the ranking in individual tasks coul d reach the top three.
  2. Resolving Task Conflicts: When merging four tasks, due to different data domain distributions, answer patterns, training data volumes, conv ergence steps, and even sequence lengths, conflicts exist between different tasks. Deeply resolving these conflict issues is crucial to obtain ing a universal model with the best comprehensive indicators across different tasks.
  3. Resolving Training Instability: Unlike generative tasks that produce multiple characters, using generative models for discriminative tasks requires the model to output a continuous value. Therefore, there is an oscillation problem during the training process. Deeply analyzing and resolving training instability can result in a model with better generalization and robustness.

Environmental requirements

cd Reranking
pip install -r requirements.txt

If your GPU supports fp16 or bf16 precision, we also recommend installing flash-attention (n ow with support for flash attention 2) to improve your runtime efficiency and reduce memory usage. (flash-attention is optional and not re quired for running this project)

git clone https://github.com/Dao-AILab/flash-attention
cd flash-attention && pip install .
# The installation below is optional and might be slow.
# pip install csrc/layer_norm
# No need to install the following if the flash-attn version is above 2.1.1.
# pip install csrc/rotary

Input Format

[
  {
    "id": "identity_0",
    "conversations": [
      {
        "from": "user",
        "value": "What Color Is the Sky\n\nBlue"
      },
      {
        "from": "assistant",
        "value": "3"
      }
    ]
  }
]

Training Script

cd Reranking
sh finetune/finetune_ds.sh

Inference Script

cd Reranking
python test_model.py

Citation

If you find our work helpful, feel free to cite as:

@article{qwen,
  title={360Zhinao Technical Report},
  author={360Zhinao-Team},
  journal={arXiv preprint arXiv:2405.13386},
  year={2024}
}

License

The source code of this repository follows the open-source license Apache 2.0.

360​Zhinao open-source models support commercial use. If you wish to use these models or continue training them for commercial purposes, please contact us via email ([email protected]) to apply. For the specific license agreement, please see <<360 Zhinao Open-Source Model License>>.