Hugging Face's logo

FuseAI
/
OpenChat-3.5-7B-Starling-v2.0

Text Generation
Transformers
Safetensors
English
mistral
model-fusion
fusechat
conversational
Eval Results
text-generation-inference
Inference Endpoints
Model card Files Community

FuseChat: Knowledge Fusion of Chat Models

| 📑 Paper | 🤗 HuggingFace Repo | 🐱 GitHub Repo |

Fanqi Wan, Longguang Zhong, Ziyi Yang, Ruijun Chen, Xiaojun Quan

Sun Yat-sen University


News

Contents

Overview

In this work, we propose an extended framework of FuseLLM to integrate the collective knowledge and individual strengths of multiple structure and scale-varied chat LLMs into a more powerful chat LLM, resulting in FuseChat. FuseChat adopts a fuse-then-merge strategy with two main stages. Firstly, it undertakes pairwise knowledge fusion for source LLMs to derive multiple target LLMs of identical structure and size via lightweight fine-tuning. Then, these target LLMs are merged within the parameter space, wherein we propose a novel method VaRM for determining the merging weights based on the variation ratio of parameter matrices before and after fine-tuning.

Moreover, we argue that the concept of knowledge fusion adopted by both FuseChat and FuseLLM shares a fundamentally similar purpose with other related topics, such as the recently popular topic of mixture of experts (MoEs), because they all aim to leverage the strengths of multiple models (experts). However, while MoEs require loading multiple experts during inference, which has higher memory requirements, knowledge fusion supports the integration of multiple LLMs with diverse architectures into a single LLM without any additional memory requirement, making it more memory-efficient.


Model Release

We release FuseChat-7B-v2.0, which is the fusion of six prominent chat LLMs with diverse architectures and scales, namely OpenChat-3.5-7B, Starling-LM-7B-alpha, NH2-Solar-10.7B, InternLM2-Chat-20B, Mixtral-8x7B-Instruct, and Qwen1.5-Chat-72B. FuseChat-7B-v2.0 achieves an average performance of 7.38 on MT-Bench (GPT-4-0125-Preview as judge LLM), which is comparable to Mixtral-8x7B-Instruct and approaches GPT-3.5-Turbo-1106.

To support a plug-and-play fusion of new source LLM, we release our target LLMs: OpenChat-3.5-7B-Starling-v2.0, OpenChat-3.5-7B-SOLAR-v2.0, OpenChat-3.5-7B-InternLM-v2.0, OpenChat-3.5-7B-Mixtral-v2.0, and OpenChat-3.5-7B-Qwen-v2.0, which are obtained from pair-wise knowledge fusion. Integrating a new source LLM at any scale requires only obtaining a target LLM from the new source LLM and merging it with the existing target LLMs.

Here are the evaluation results.


Quick Start

Setup

We use python 3.11 in this project.

Then, we have to install all the libraries listed in requirements.txt.

pip install -r requirements.txt

Usage

Here's how you can run the model using the 🤗 Transformers:

import transformers
tokenizer = transformers.AutoTokenizer.from_pretrained("FuseAI/FuseChat-7B-v2.0")
# Single-turn
tokens = tokenizer("GPT4 Correct User: Hello<|end_of_turn|>GPT4 Correct Assistant:").input_ids
assert tokens == [1, 420, 6316, 28781, 3198, 3123, 1247, 28747, 22557, 32000, 420, 6316, 28781, 3198, 3123, 21631, 28747]
# Multi-turn
tokens = tokenizer("GPT4 Correct User: Hello<|end_of_turn|>GPT4 Correct Assistant: Hi<|end_of_turn|>GPT4 Correct User: How are you today?<|end_of_turn|>GPT4 Correct Assistant:").input_ids
assert tokens == [1, 420, 6316, 28781, 3198, 3123, 1247, 28747, 22557, 32000, 420, 6316, 28781, 3198, 3123, 21631, 28747, 15359, 32000, 420, 6316, 28781, 3198, 3123, 1247, 28747, 1602, 460, 368, 3154, 28804, 32000, 420, 6316, 28781, 3198, 3123, 21631, 28747]

The GPT4 template is also available as the integrated tokenizer.chat_template, which can be used instead of manually specifying the template:

messages = [
    {"role": "user", "content": "Hello"},
    {"role": "assistant", "content": "Hi"},
    {"role": "user", "content": "How are you today?"}
]
tokens = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
assert tokens == [1, 420, 6316, 28781, 3198, 3123, 1247, 28747, 22557, 32000, 420, 6316, 28781, 3198, 3123, 21631, 28747, 15359, 32000, 420, 6316, 28781, 3198, 3123, 1247, 28747, 1602, 460, 368, 3154, 28804, 32000, 420, 6316, 28781, 3198, 3123, 21631, 28747]

Supported Models

Model Model size Chat Template
openchat_3.5/Starling-LM-7B-alpha 7B openchat_3.5
Mistral/Mixtral 7B/8x7B/8x22B mistral
Llama 3/Llama 3.1 8B/70B llama-3
Gemma/Gemma 2/CodeGemma 2B/7B/9B/27B gemma
Phi-3 4B/7B/14B phi-3
Qwen1.5/Qwen2 0.5B/1.5B/4B/7B/14B/32B/72B/110B qwen
InternLM2/InternLM2.5 7B/20B internlm2
Yi/Yi-1.5 6B/9B/34B yi

To support a new model in FuseChat, you'll need to follow these steps:

  1. Implement a conversation template for the new model at conversation.py. You can follow existing examples and use register_conv_template to add a new one.
  2. Implement a model adapter for the new model at model/model_adapter.py. You can follow existing examples and use register_model_adapter to add a new one.
  3. Modify the preprocess() Function in train.py.

Experiment Preparation

Source LLMs

We conduct experiments using six representative chat LLMs as the source LLMs, including OpenChat-3.5-7B, Starling-LM-7B-alpha, NH2-SOLAR-10.7B, InternLM2-Chat-20B, Mixtral-8x7B-Instruct, Qwen-1.5-Chat-72B. As for the pivot LLM, which also serves as the starting point for the target LLMs, we opt for OpenChat-3.5-7B due to its balanced scale and performance. You should download all these models and place them into /models before experiments.

Training Dataset

We curated a comprehensive training dataset, FuseChat-Mixture, from various sources. This dataset covers different styles and capabilities, featuring both human-written and model-generated, and spanning general instruction-following and specific skills. You should download the dataset and place it into data/fusechat_v1_clean_split_2048_filter_wrong.json before experiments.

Data Construction

Here we show the scripts to obtain representations from multiple source LLMs for model fusion with the following three steps.

1. Get Representations

Here we show the scripts to obtain representations from multiple source LLMs.

# We split the dataset into 4 splits, then process each split on one or multiple GPUs.
# OpenChat-3.5-7B Starling-LM-7B-alpha Nous-Hermes-2-SOLAR-10.7B internlm2-chat-20b Mixtral-8x7B-Instruct-v0.1 Qwen1.5-72B-Chat
export CUDA_VISIBLE_DEVICES=0 # specify one or multiple GPUs
PROJ_PATH=FuseChat # specify your own project path
DATA_NAME="fusechat_v1_clean_split_2048_filter_wrong" 
MODEL_NAME=openchat_3.5 # model to get representation
CONV_TEMP=openchat_3.5 # conversation template, should be the same for all models, see more template names in train/conversation.py

for i in {0..3}; do
python ${PROJ_PATH}/train/get_data_representation.py \
  --model_name_or_path ${PROJ_PATH}/models/${MODEL_NAME} \
  --data_path ${PROJ_PATH}/data/${DATA_NAME}.json \
  --dataset_save_dir ${PROJ_PATH}/representations/${MODEL_NAME}_representation_split${i} \
  --tknz_dataset_path ${PROJ_PATH}/representations/${MODEL_NAME}_representation_tknz_split${i} \
  --cache_dir ${PROJ_PATH}/.cache/huggingface/datasets \
  --model_max_length 2048 \
  --load_in_half bf16 \
  --batch_size 32 \
  --top_k_logits 10 \
  --save_per_token_metric \
  --no_assert \
  --conv_temp ${CONV_TEMP} \
  --mask_instruction \
  --dataset_split_num 4 \
  --dataset_index ${i} \
  --get_representation \
  --device_map "auto"
done

2. Align Representations

Here we show the scripts to align representations from different source LLMs to pivot LLM.

For source LLMs share the same vocab as pivot LLM, we only merge their representations into a single dataset.

# Pivot LLM:OpenChat-3.5-7B <-> Source LLMs: Starling-LM-7B-alpha Nous-Hermes-2-SOLAR-10.7B Mixtral-8x7B-Instruct-v0.1
PROJ_PATH=FuseChat # specify your own project path
PIVOT_NAME=openchat_3.5  # Pivot LLM
SOURCE_NAME=Starling-LM-7B-alpha # Source LLMs with the same vocab as Pivot

for i in {0..3}; do
python ${PROJ_PATH}/train/replace_model.py \
  --dataset_dir ${PROJ_PATH}/representations/${PIVOT_NAME}_representation_split${i} \
  --replace_dataset_dir ${PROJ_PATH}/representations/${SOURCE_NAME}_representation_split${i} \
  --dataset_save_dir ${PROJ_PATH}/representations/${PIVOT_NAME}_${SOURCE_NAME}_representation_split${i} \
  --preprocessing_num_workers 32 \
  --batch_size 1000 
done

For source LLMs have different vocabs with pivot LLM, we need to do token alignment and distribution alignment.

# Pivot LLM:OpenChat-3.5-7B <->Source LLMs: internlm2-chat-20b Qwen1.5-72B-Chat
PROJ_PATH=FuseChat # specify your own project path
PIVOT_NAME=openchat_3.5 # Pivot LLM
SOURCE_NAME=internlm2-chat-20b # Source LLMs have different vocab with Pivot
align_type="default" # different alignment method hard -> EM, soft -> MinED, default -> MS
token_alignment_matrix_file=${PROJ_PATH}/aligned_representations/${PIVOT_NAME}_${SOURCE_NAME}_token_sparse_matrix_${align_type}.npz
blending_to_base_file=${PROJ_PATH}/aligned_representations/${PIVOT_NAME}_${SOURCE_NAME}_token_mapping_${align_type}.json

# token alignment
python ${PROJ_PATH}/train/align_token_and_vocab.py \
  --align_type ${align_type} \
  --base_model_name_or_path ${PROJ_PATH}/models/${PIVOT_NAME} \
  --blending_model_name_or_path ${PROJ_PATH}/models/${SOURCE_NAME} \
  --base_dataset_dir "${PROJ_PATH}/representations/${PIVOT_NAME}_representation_tknz_split0,${PROJ_PATH}/representations/${PIVOT_NAME}_representation_tknz_split1,${PROJ_PATH}/representations/${PIVOT_NAME}_representation_tknz_split2,${PROJ_PATH}/representations/${PIVOT_NAME}_representation_tknz_split3" \
  --blending_dataset_dir "${PROJ_PATH}/representations/${SOURCE_NAME}_representation_tknz_split0,${PROJ_PATH}/representations/${SOURCE_NAME}_representation_tknz_split1,${PROJ_PATH}/representations/${SOURCE_NAME}_representation_tknz_split2,${PROJ_PATH}/representations/${SOURCE_NAME}_representation_tknz_split3" \
  --aligned_dataset_tknz_save_dir ${PROJ_PATH}/aligned_representations/${PIVOT_NAME}_${SOURCE_NAME}_representation_tknz \
  --model_max_length 2048 \
  --preprocessing_num_workers 32 \
  --batch_size 16 \
  --token_alignment_matrix_file ${token_alignment_matrix_file} \
  --blending_to_base_file ${blending_to_base_file} \
  --do_token_alignment \
  --metric_level "sequence" \
  --use_token_alignment_matrix

# distribution alignment
for i in {0..3}; do
python ${PROJ_PATH}/train/align_token_and_vocab.py \
  --align_type ${align_type} \
  --base_model_name_or_path ${PROJ_PATH}/models/${PIVOT_NAME} \
  --blending_model_name_or_path ${PROJ_PATH}/models/${SOURCE_NAME} \
  --base_dataset_dir ${PROJ_PATH}/representations/${PIVOT_NAME}_representation_split${i} \
  --blending_dataset_dir ${PROJ_PATH}/representations/${SOURCE_NAME}_representation_split${i} \
  --aligned_dataset_save_dir ${PROJ_PATH}/aligned_representations/${PIVOT_NAME}_${SOURCE_NAME}_representation_split${i} \
  --model_max_length 2048 \
  --preprocessing_num_workers 32 \
  --batch_size 16 \
  --temperature 0.5 \
  --token_alignment_matrix_file ${token_alignment_matrix_file} \
  --blending_to_base_file ${blending_to_base_file} \
  --do_distribution_alignment \
  --metric_level "sequence" \
  --use_token_alignment_matrix
done

3. Filter NaN

Here we show the scripts to filter instance with NaN "metric_ce" in the dataset.

for i in {0..3}; do
python ${PROJ_PATH}/train/filter_nan.py \
  --input_data_dir ${PROJ_PATH}/aligned_representations/${PIVOT_NAME}_${SOURCE_NAME}_representation_split${i} \
  --output_data_dir ${PROJ_PATH}/aligned_representations/${PIVOT_NAME}_${SOURCE_NAME}_representation_split${i}_fnan \
done

The processed representations data are in the following path:

${PROJ_PATH}/aligned_representations/${PIVOT_NAME}_${SOURCE_NAME}_representation_split${i}_fnan

Pairwise Knowledge Fusion

We show the scripts for pairwise knowledge fusion with the processed representations.

# OpenChat-3.5-7B <-> Starling-LM-7B-alpha Nous-Hermes-2-SOLAR-10.7B internlm2-chat-20b Mixtral-8x7B-Instruct-v0.1 Qwen1.5-72B-Chat
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
torchrun --nproc_per_node=8 --master_port=20001 ${PROJ_PATH}/train/train.py \
  --model_name_or_path "openchat/openchat_3.5" \
  --data_path "${PROJ_PATH}/aligned_representations/${PIVOT_NAME}_${SOURCE_NAME}_representation_split0_fnan,${PROJ_PATH}/aligned_representations/${PIVOT_NAME}_${SOURCE_NAME}_representation_split1_fnan,${PROJ_PATH}/aligned_representations/${PIVOT_NAME}_${SOURCE_NAME}_representation_split2_fnan,${PROJ_PATH}/aligned_representations/${PIVOT_NAME}_${SOURCE_NAME}_representation_split3_fnan" \
  --bf16 True \
  --output_dir "${PROJ_PATH}/checkpoints/${PIVOT_NAME}_${SOURCE_NAME}_pairwise_fusion_ckpt" \
  --num_train_epochs 3 \
  --per_device_train_batch_size 4 \
  --per_device_eval_batch_size 4 \
  --gradient_accumulation_steps 4 \
  --evaluation_strategy "no" \
  --save_strategy "epoch" \
  --save_steps 10000 \
  --save_total_limit 5 \
  --learning_rate 5e-6 \
  --weight_decay 0. \
  --warmup_ratio 0.03 \
  --lr_scheduler_type "cosine" \
  --logging_steps 1 \
  --fsdp "full_shard auto_wrap" \
  --fsdp_transformer_layer_cls_to_wrap 'MistralDecoderLayer' \
  --tf32 True \
  --model_max_length 2048 \
  --gradient_checkpointing True \
  --conv_temp "openchat_3.5" \
  --lazy_preprocess True \
  --flash_attn_transformers True \
  --do_train \
  --do_fuse \
  --fuse_with_ref_model True \
  --fuse_loss_type "ce" \
  --fuse_temperature 1.0 \
  --lm_loss_weight 0.9 \
  --dataloader_num_workers 8 \
  --remove_unused_columns False

Model Merging

We show the scripts on how to get FuseChat from target LLMs using different merging methods.

Before merging, please install our modified "mergekit".

cd mergekit
pip install -e .

Our SCE method

The implementation of our method is in mergekit/mergekit/merge_methods/sce_merging.py.

model_save_dir=xx # specify your path to save the merged models
mergekit-yaml mergekit/fusechat_configs/fusechat-sce.yml ${model_save_dir}/FuseChat-7B-SCE

Other merging methods 

model_save_dir=xx # your path to save the merged models
mergekit-yaml mergekit/fusechat_configs/fusechat-linear.yml ${model_save_dir}/FuseChat-7B-LINEAR

mergekit-yaml mergekit/fusechat_configs/fusechat-ta.yml ${model_save_dir}/FuseChat-7B-TA

mergekit-yaml mergekit/fusechat_configs/fusechat-ties.yml ${model_save_dir}/FuseChat-7B-TIES

mergekit-yaml mergekit/fusechat_configs/fusechat-dare.yml ${model_save_dir}/FuseChat-7B-DARE

Evaluation

We conduct experiments on two representative benchmarks named AlpacaEval 2.0 and MT-Bench to evaluate the instruction-following and multi-turn conversation capabilities.

MT-Bench

MT-Bench comprises 80 multi-turn dialogues spanning writing, roleplay, reasoning, math, coding, stem, and humanities domains.The original benchmark uses GPT-4-0613 as the evaluator to provide a scalar score ranging from 1 (lowest) to 10 (highest) for the generated responses. However, due to inaccuracies in the reference responses generated by the old GPT-4-0613, we follow the latest works to adopt an updated GPT-4-0125-Preview to correct these errors and evaluate the generated responses.

Please download the official code and follow the guidelines for evaluation. To use GPT-4-0125-Preview as judge model, you should download gpt-4-0125-preview.jsonl, and place it in llm_judge/data/mt_bench/reference_answer. Then, add "gpt-4-0125-preview" as a valid judge model in common.py.

# Step 1. Generate model answers to MT-bench questions
export CUDA_VISIBLE_DEVICES=0,1
python gen_model_answer.py \
  --model-path "FuseChat-7B-v2.0" \
  --model-id "openchat_3.5_fusechat_7b_sce" \
  --num-gpus-per-model 1 \
  --num-gpus-total 2

# Step 2. Generate GPT-4-0125-Preview judgments
export OPENAI_API_KEY=XXXXXX  # set the OpenAI API key
python gen_judgment.py \
  --model-list "openchat_3.5_fusechat_7b_sce" \
  --judge-model "gpt-4-0125-preview" \
  --parallel 8

# Step 3. Show MT-bench scores
python show_result.py --model-list "openchat_3.5_fusechat_7b_sce"

AlpacaEval 2.0

AlpacaEval 2.0, contains 805 instructions from five test subsets. This benchmark compares the Win Rate and Length-Controlled Win Rate (LC Win Rate) against GPT-4. We follow the default settings to employ GPT-4-1106-Preview to assess the quality of generated responses.

Please download the official code and follow the guidelines. We use the default alpaca_eval_gpt4_turbo_fn for evaluation. The prompt for generation is:

GPT4 Correct User: {instruction}<|end_of_turn|>GPT4 Correct Assistant:

Citation

If you find this work is relevant with your research or applications, please feel free to cite our work!

@article{wan2024fusechat,
  title={FuseChat: Knowledge Fusion of Chat Models},
  author={Fanqi Wan and Longguang Zhong and Ziyi Yang and Ruijun Chen and Xiaojun Quan},
  journal={arXiv preprint arXiv:2408.07990},
  year={2024}
}
Downloads last month
289
Safetensors
Model size
7.24B params
Tensor type
BF16
·
Inference Examples
Text Generation
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Model tree for FuseAI/OpenChat-3.5-7B-Starling-v2.0

Base model

openchat/openchat_3.5
Finetuned
(26)
this model
Quantizations
2 models

Dataset used to train FuseAI/OpenChat-3.5-7B-Starling-v2.0

Collection including FuseAI/OpenChat-3.5-7B-Starling-v2.0

Evaluation results

  • score on MT-Bench (GPT-4-0125-Preview)
    source
    7.220
  • lc win rate on AlpacaEval 2.0 (GPT-4-1106-Preview)
    source
    16.200
View on Papers With Code