LM-Cocktail: Resilient Tuning of Language Models via Model Merging

Make fine-tuning of language models akin to crafting a nuanced cocktail.

Model merging can be used to improve the performance of single model. We find this method is also useful for large language models and dense embedding model, and design the LM-Cocktail strategy which automatically merges fine-tuned models and base model using a simple function to compute merging weights. LM-Cocktail can be used to improve the performance on target domain without decrease the general capabilities beyond target domain. It also can be used to generate a model for new tasks without fine-tuning. For more details please refer to our report: LM-Cocktail.

The following are some application scenarios (Note that the models used to merge need to have the same architecture and the same initialization parameter):

Fine-tuning the base language model could lead to severe degeneration of model’s general capabilities beyond the targeted domain. By mixing the fine-tuned model and the base model (use function mix_models), LM-Cocktail can significantly enhance performance in downstream task while maintaining performance in other unrelated tasks.

If there are some available models fine-tuned on other tasks, you can further use them to enhance your fine-tuned model. Firstly, you need to collect five example data from your task, then employ function mix_models_with_data to compute weights and merge available models. In this way, it can assign lower weights to low-quality models, avoiding degrading the performance on your task. Finally, use mix_models to merge produced model and your fine-tuned model.

LM-Cocktail can improve the accuracy of the new task without a requisition to fine-tune a model. Give a few examples data (e.g., five examples), and some available models (from open-source community or pre-existing for other tasks), function mix_models_wit_data can automatically assign different merging weights for different model based their loss in example data, and then merge these available models to generate a task-specific new model.

If you have some models who are fine-tune on different tasks, you can merge them into one model to approximate multitask learning. The merged model can be used to perform multiple tasks.

Install the latest version from source (Recommended):

git clone https://github.com/FlagOpen/FlagEmbedding.git
cd FlagEmbedding/LM_Cocktail
pip install -e .

Install by pip:

pip install -U LM_Cocktail

There are two key functions in LM-Cocktail:

mix_models can merge models based on the given merging weights. An example is merging the fine-tuned model and the base model to mitigate Catastrophic Forgetting after fine-tuning:

from LM_Cocktail import mix_models, mix_models_with_data

# mix LLMs and save it to output_path: ./mixed_model_1
model = mix_models(
    model_names_or_paths=["meta-llama/Llama-2-7b-chat-hf", "Shitao/llama2-ag-news"], 
    model_type='decoder', 
    weights=[0.7, 0.3], 
    output_path='./mixed_llm')
# you can select a weight for your models to get a trade-off between generality and expertise.

# Mix Embedding Models
model = mix_models(
    model_names_or_paths=["BAAI/bge-base-en-v1.5", "Shitao/bge-hotpotqa"], 
    model_type='encoder', 
    weights=[0.5, 0.5],
    output_path='./mixed_embedder')

# Mix reranker Models
model = mix_models(
    model_names_or_paths=["BAAI/bge-reranker-base", "BAAI/bge-reranker-base"], 
    model_type='reranker', 
    weights=[0.5, 0.5],
    output_path="./mixed_reranker")

Note that the sum of weights should be equal to 1.

You also can merge multiple models:

from LM_Cocktail import mix_models, mix_models_with_data

model = mix_models(
    model_names_or_paths=["BAAI/bge-base-en-v1.5", "Shitao/bge-hotpotqa", "Shitao/bge-quora", "Shitao/bge-msmarco"], 
    model_type='encoder', 
    weights=[0.3, 0.2, 0.2, 0.3],
    output_path='./mixed_embedder_2')
# The sum of weights should be equal to 1.

mix_models_with_data can compute merging weights based on given data and merge models. It can be used to produce a model for a new task without training, or boost the performance for the downstream task by leveraging the knowledge in others models.

• For LLMs

The format of example_data for LLMs is a list, where each item is a dict like:

{"input": str, "output": str}

LM-cocktial will compute the loss of the output.

You can use the example data to merge models as following:

from LM_Cocktail import mix_models, mix_models_with_data

example_data = [
    {"input": "Question: when was the last time anyone was on the moon? Answer:\n", "output": "14 December 1972 UTC"},
    {"input": "Review: \"it 's a charming and often affecting journey . \" Is this movie review sentence negative or positive?\n", "output": "Positive"}
]

model = mix_models_with_data(
    model_names_or_paths=["meta-llama/Llama-2-7b-chat-hf", "Shitao/llama2-ag-news", "Shitao/llama2-nq"], 
    model_type='decoder', 
    example_data=example_data, 
    temperature=5.0)
# you can set the temperature argument to adjust the distribution of mixing weights

• For Embedder

The format of example_data for LLMs is a list, where each item is a dict like:

{"query": str, "pos": List[str], 'neg': List[str]}

where pos is a list of positive text and neg is a list of negative text. LM-Cocktail will compute the contrastive loss.

You can use the example data to merge models as following:

from LM_Cocktail import mix_models, mix_models_with_data

example_data = [
    {"query": "How does one become an actor in the Telugu Film Industry?", "pos": [" How do I become an actor in Telugu film industry?"], "neg": [" What is the story of Moses and Ramesses?", " Does caste system affect economic growth of India?"]}, 
    {"query": "Why do some computer programmers develop amazing software or new concepts, while some are stuck with basic programming work?", "pos": [" Why do some computer programmers develops amazing softwares or new concepts, while some are stuck with basics programming works?"], "neg": [" When visiting a friend, do you ever think about what would happen if you did something wildly inappropriate like punch them or destroy their furniture?", " What is the difference between a compliment and flirting?"]}
]

model = mix_models_with_data(
    model_names_or_paths=["BAAI/bge-base-en-v1.5", "Shitao/bge-hotpotqa", "Shitao/bge-quora"], 
    model_type='encoder', 
    example_data=example_data,
    temperature=5.0,
    max_input_length=512,
    neg_number=2)

The function mix_models_by_layers creates temporary directories to store weights of individual models and then merges them layer by layer.

This approach helps in reducing the memory consumption.

Once the merging process is completed, the temporary directories and files will be automatically removed.

from LM_Cocktail import mix_models_by_layers

# Mix Large Language Models (LLMs) and save the combined model to the path: ./mixed_llm
model = mix_models_by_layers(
    model_names_or_paths=["meta-llama/Llama-2-7b-chat-hf", "Shitao/llama2-ag-news"], 
    model_type='decoder', 
    weights=[0.7, 0.3], 
    output_path='./mixed_llm')

Detailed results please refer to our report: LM-Cocktail

• LM-Cocktail for Catastrophic Forgetting

[1]: merge 2 models: fine-tuned model and the base model

[2]: merge 10 models based on five examples: fine-tuned model, the base model, and 8 models fine-tuned on other tasks

• LM-Cocktail for new tasks without fine-tuning

Merge 10 models fine-tuned on other tasks based on five examples for new tasks:

• Models: we fine-tune the meta-llama/Llama-2-7b-chat-hf on 9 tasks, and you can find the fine-tuned models at this link. Note that the most of fine-tuned models has a poor performance on other unrelated tasks.
• Examples Data for dataset from FLAN: ./llm_examples.json
• MMLU dataset: https://huggingface.co/datasets/cais/mmlu (use the example in dev set to do in-context learning)

You can use these models and our code to produce a new model and evaluate its performance using the llm-embedder script as following:

# for 30 tasks from FLAN
torchrun --nproc_per_node 8 -m evaluation.eval_icl \
--retrieval_method no \
--few_shot 0 \
--data_root /data/llm-embedder \
--model_name_or_path ./mixed_model_1

# for MMLU datasets
torchrun --nproc_per_node 8 -m evaluation.eval_mmlu \
--retrieval_method no \
--few_shot 0 \
--data_root /data/llm-embedder \
--model_name_or_path ./mixed_model_2

• Models: we fine-tune the bge-base-en-v1.5 on 9 tasks, and you can find the fine-tuned models at this link.
• Examples Data: ./embedder_examples.json

Use MTEB script to evaluate the mixed embedding model:

python eval_MTEB.py --model_name_or_path mixed_model --task_type Retrieval

This project is inspired by previous researches on model merging, including LoraHub, model soups, and PAINT .

The Llama is fine-tuned using the FastChat scripts. Fine-tuning datasets are from sentence-transformers/embedding-training-data and intfloat/llm-retriever-tasks.

If you find this repository useful, please consider giving a star ⭐ and citation

@misc{cocktail,
      title={LM-Cocktail: Resilient Tuning of Language Models via Model Merging}, 
      author={Shitao Xiao and Zheng Liu and Peitian Zhang and Xingrun Xing},
      year={2023},
      eprint={2311.13534},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}