Skip to content

Latest commit

 

History

History
364 lines (275 loc) · 15.2 KB

README.md

File metadata and controls

364 lines (275 loc) · 15.2 KB
Raw

LLMBox | Training | Utilization

LLMBox

LLMBox is a comprehensive library for implementing LLMs, including a unified training pipeline and comprehensive model evaluation. LLMBox is designed to be a one-stop solution for training and utilizing LLMs. Through a practical library design, we achieve a high-level of flexibility and efficiency in both training and utilization stages.

Key Features

Training

  • Diverse training strategies: We support multiple training strategies, including Supervised Fine-tuning (SFT), Pre-training (PT), PPO and DPO.
  • Comprehensive SFT datasets: We support 9 SFT datasets as the inputs for training.
  • Tokenizer Vocabulary Merging: We support the tokenizer merging function to expand the vocabulary.
  • Data Construction Strategies: We currently support merging multiple datasets for training. Self-Instruct and Evol-Instruct are also available to process the dataset.
  • Parameter Efficient Fine-Tuning: LoRA and QLoRA are supported in SFT or PT.
  • Efficient Training: We support Flash Attention and Deepspeed for efficient training.

Utilization

  • Blazingly Fast: By managing the KV Cache of prefixes, we can speed up local inference by up to 6x 🚀.
  • Comprehensive Evaluation: 59+ commonly used datasets and benchmarks in evaluating LLMs.
  • Evaluation Methods: Accurately reproduce results from original papers of OpenAI, LLaMA, Mistral, and other models.
  • In-Context Learning: We support various ICL strategies, including KATE, GlobalE, and APE.
  • Chain-of-Thought: For some datasets, we support three types of CoT evaluation: base, least-to-most, and pal.
  • vLLM and Flash Attention Support: We also support vLLM and Flash Attention for efficient inference.
  • Quantization: BitsAndBytes and GPTQ quantization are supported.

Documentations

See documentations for more details.

Quick Start

Install

git clone https://github.com/RUCAIBox/LLMBox.git && cd LLMBox
pip install -r requirements.txt

If you are only evaluating the OpenAI (or OpenAI compatible like DeepSeek, Perplexity) models, you can install the minimal requirements requirements-openai.txt.

For installation problem, see trouble shooting.

Update LLMBox

Currently, you can simply pull the latest repository from GitHub to update LLMBox.

git pull

If you are facing a merge conflict, please try to drop, stash, or commit your local changes first.

git checkout local_changes && git add -p && git commit -m "local changes"
git checkout main
git pull

The above commands show how to commit your local changes to a new branch, and then update the LLMBox.

Quick Start with Training

You can start with training a SFT model based on LLaMA-2 (7B) with deepspeed3:

cd training
bash download.sh
bash bash/run_ds3.sh

Quick Start with Utilization

To utilize your model, or evaluate an existing model, you can run the following command:

python inference.py -m gpt-3.5-turbo -d copa  # --num_shot 0 --model_type chat

This is default to run the OpenAI GPT 3.5 turbo model on the CoPA dataset in a zero-shot manner.

Training

LLMBox Training supports various training strategies and dataset construction strategies, along with some efficiency-improving modules. You can train your model with the following command:

python train.py \
    --model_name_or_path meta-llama/Llama-2-7b-hf \
    --data_path data/ \
    --dataset alpaca_data_1k.json \
    --output_dir $OUTPUT_DIR \
    --num_train_epochs 2 \
    --per_device_train_batch_size 8 \
    --gradient_accumulation_steps 2 \
    --save_strategy "epoch" \
    --save_steps 2 \
    --save_total_limit 2 \
    --learning_rate 1e-5 \
    --lr_scheduler_type "constant"

Alternatively, you can use the following preset bash scripts to train your model:

Merging Tokenizer

If you want to pre-train your models on corpora with languages or tokens not well-supported in original language mdoels(e.g., LLaMA), we provide the tokenizer merging function to expand the vocabulary based on the corpora by using sentencepiece. You can check merge_tokenizer.py for detailed information. Please follow the guide in Pre-train.

bash bash/run_7b_pt.sh

Merging Datasets

If you want to train your models with a mix of multiple datasets, you can pass a list of dataset files or names to LLMBox. LLMBox will transfer each file or name into a PTDataset or SFTDataset, and merge them together to construct a combined dataset. You can also set the merging ratio of each dataset by passing a list of floats to LLMBox. Please follow the guide in Merge Dataset.

bash bash/run_7b_hybrid.sh

Self-Instruct and Evol-Instruct

Since manually creating instruction data of high qualities to train the model is very time-consuming and labor-intensive, Self-Instruct and Evol-Instruct are proposed to create large amounts of instruction data with varying levels of complexity using LLM instead of humans. LLMBox support both Self-Instruct and Evol-Instruct to augment or enhance the input data files. Please follow the guide in Self-Insturct and Evol-Instruct

python self_instruct/self_instruct.py --seed_tasks_path=seed_tasks.jsonl

For more details, view the training documentation.

Utilization

We provide a broad support on Huggingface models (e.g. LLaMA-3, Mistral, or the model you are building on), OpenAI, Anthropic, QWen and other OpenAI-compatible models for further utilization. Full list of model backends: here.

Currently a total of 59+ commonly used datasets are supported, including: HellaSwag, MMLU, GSM8K, GPQA, AGIEval, CEval, and CMMLU. Full list of datasets: here.

CUDA_VISIBLE_DEVICES=0 python inference.py \
  -m llama-2-7b-hf \
  -d mmlu agieval:[English] \
  --model_type chat \
  --num_shot 5 \
  --ranking_type ppl_no_option

  • 🔥 Recently supported datasets: imbue_code, imbue_public, and imbue_private.

  • 🔥 See benchmarking LLaMA3 for more examples.

Performance
Model get_ppl get_prob generation
Hellaswag (0-shot) MMLU (5-shot) GSM (8-shot)
GPT-3.5 Turbo 79.98 69.25 75.13
LLaMA-2 (7B) 76 45.95 14.63

Efficient Evaluation

We by default enable prefix caching for efficient evaluation. vLLM is also supported.

Time
Model Efficient Method get_ppl get_prob generation
Hellaswag (0-shot) MMLU (5-shot) GSM (8-shot)
LLaMA-2 (7B) Vanilla 0:05:32 0:18:30 2:10:27
vLLM 0:06:37 0:14:55 0:03:36
Prefix Caching 0:05:48 0:05:51 0:17:13

You can also use the following command to use vllm:

python inference.py -m ../Llama-2-7b-hf -d mmlu:abstract_algebra,anatomy --vllm True  # --prefix_caching False --flash_attention False

To evaluate with quantization, you can use the following command:

python inference.py -m model -d dataset --load_in_4bits  # --load_in_8_bits or --gptq

Evaluation Method

Various types of evaluation methods are supported:


Dataset Evaluation Method Instruction

Generation

{
  "question":
    "when was ...",
  "answer": [
    '14 December 1972',
    'December 1972'
  ]
}

generation

Generate based on the source text

Example: ARC-Challenge (extract choice characters like A, B, C, D from model generation), GSM8K, HumanEval

Notes: vLLM generally generates faster than transformers but has different default parameters. Users may need to adjust sampling parameters like temperature and length_penalty for optimal results.

generation

Q: When was ...?
A: ________

MultipleChoice

{
  "question":
    "What is the ...?",
  "choices": [
    "The first",
    "The second",
    ...
  ],
  "answer": 3
}

get_ppl

Calculate perplexity of the option text based on the source text (i.e. compute log-likelihood over the suffix)

Example: WinoGrande, BoolQ

Notes: Some datasets, such as ARC, OpenbookQA, and RACE use normalized accuracy when evaluated with the get_ppl method.

ppl_no_option

Q: What is ...?
A: The first
   └--ppl--┘

ppl

Q: What is ...?
A. The first
B. The second
C. ...
A: A. The first
   └----ppl---┘

get_prob

Get the probability of each option label (i.e. over choice characters)

Example: MMLU, ARC-Challange

prob

Q: What is ...?
A. The first
B. The second
C. ...
A: _
   └→ [A B C D]

You can find more evaluation details for each dataset at supported datasets.

You can use --instruction to pass a jinja template to override the default instruction.

By default, we use the get_ppl method with ppl_no_option ranking type for MultipleChoiceDataset and the generation method for GenerationDataset. You can also use the following command to use the get_prob method or ppl variant of get_ppl for MultipleChoiceDataset:

python inference.py -m model -d dataset --ranking_type prob  # or ppl

We also support In-Context Learning and Chain-of-Thought evaluation for some datasets:

python inference.py -m model -d dataset --kate  # --globale or --ape
python inference.py -m model -d dataset --cot least_to_most  # --base or --pal

For a more detailed instruction on model utilization, view the utilization documentation.

For a full list of evaluation results, see our paper LLMBox: A Comprehensive Library for Large Language Models.

Contributing

Please let us know if you encounter a bug or have any suggestions by filing an issue.

We welcome all contributions from bug fixes to new features and extensions.

We expect all contributions discussed in the issue tracker and going through PRs.

For more details, view the CONTRIBUTING documentation.

We thank the following contributors for their contributions to LLMBox:

  • @xansar for fixing multiple complex issues like batch sampler and self-consistency.

The Team

LLMBox is developed and maintained by AI Box. See more details in change log

License

LLMBox uses MIT License.

Reference

If you find LLMBox useful for your research or development, please cite the following papers:

@inproceedings{tang2024llmbox,
  title={LLMBox: A Comprehensive Library for Large Language Models},
  author={Tang, Tianyi and Yiwen, Hu and Li, Bingqian and Luo, Wenyang and Qin, ZiJing and Sun, Haoxiang and Wang, Jiapeng and Xu, Shiyi and Cheng, Xiaoxue and Guo, Geyang and others},
  booktitle={Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations)},
  pages={388--399},
  year={2024}
}