LLMBox | Training | Utilization
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.
Training
- Diverse training strategies: We support multiple training strategies, including Supervised Fine-tuning (
SFT
), Pre-training (PT
),PPO
andDPO
. - 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
andEvol-Instruct
are also available to process the dataset. - Parameter Efficient Fine-Tuning:
LoRA
andQLoRA
are supported in SFT or PT. - Efficient Training: We support
Flash Attention
andDeepspeed
for efficient training.
Utilization
- Blazingly Fast: By managing the KV Cache of prefixes or using
vLLM
, we can speed up local inference by up to 6x π. - Comprehensive Evaluation: 59+ commonly used datasets and benchmarks in evaluating LLMs.
- Evaluation Methods: π Accurately reproduce results from original papers of OpenAI, LLaMA, Mistral, and other models.
- In-Context Learning: We support various ICL strategies, including
KATE
,GlobalE
, andAPE
. - Chain-of-Thought: For some datasets, we support three types of CoT evaluation:
base
,least-to-most
, andpal
. - Quantization: BitsAndBytes and GPTQ quantization are supported.
- Easy To Use: Detailed results are provided for users to debug or integrate new models/datasets/cot.
See documentations for more details.
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.
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
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.
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:
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
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
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.
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
-
π₯ Recently supported datasets:
imbue_code
,imbue_public
, andimbue_private
. -
π₯ 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 |
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
Various types of evaluation methods are supported:
Dataset | Evaluation Method | Instruction |
Generation
|
Generate based on the source text Example: ARC-Challenge (extract choice characters like A, B, C, D from model generation), GSM8K, HumanEval Notes: |
|
MultipleChoice
|
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 the probability of each option label (i.e. over choice characters) Example: MMLU, ARC-Challange |
|
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.
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.
LLMBox is developed and maintained by AI Box. See more details in change log
LLMBox uses MIT License.
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}
}