File size: 20,781 Bytes
3d4d706 f2f302c e48003f f2f302c e48003f f2f302c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 |
---
license: apache-2.0
---
# FuseChat-7B-VaRM
## Description
This repo contains GGUF format model files for FuseChat-7B-VaRM.
## Files Provided
| Name | Quant | Bits | File Size | Remark |
| ----------------------------- | ------- | ---- | --------- | -------------------------------- |
| fusechat-7b-varm.IQ3_XXS.gguf | IQ3_XXS | 3 | 3.02 GB | 3.06 bpw quantization |
| fusechat-7b-varm.IQ3_S.gguf | IQ3_S | 3 | 3.18 GB | 3.44 bpw quantization |
| fusechat-7b-varm.IQ3_M.gguf | IQ3_M | 3 | 3.28 GB | 3.66 bpw quantization mix |
| fusechat-7b-varm.Q4_0.gguf | Q4_0 | 4 | 4.11 GB | 3.56G, +0.2166 ppl |
| fusechat-7b-varm.IQ4_NL.gguf | IQ4_NL | 4 | 4.16 GB | 4.25 bpw non-linear quantization |
| fusechat-7b-varm.Q4_K_M.gguf | Q4_K_M | 4 | 4.37 GB | 3.80G, +0.0532 ppl |
| fusechat-7b-varm.Q5_K_M.gguf | Q5_K_M | 5 | 5.13 GB | 4.45G, +0.0122 ppl |
| fusechat-7b-varm.Q6_K.gguf | Q6_K | 6 | 5.94 GB | 5.15G, +0.0008 ppl |
| fusechat-7b-varm.Q8_0.gguf | Q8_0 | 8 | 7.70 GB | 6.70G, +0.0004 ppl |
## Parameters
| path | type | architecture | rope_theta | sliding_win | max_pos_embed |
| ----------------------- | ------- | ------------------ | ---------- | ----------- | ------------- |
| FuseAI/FuseChat-7B-VaRM | mistral | MistralForCausalLM | 10000.0 | 4096 | 8192 |
## Benchmarks
## Specific Purpose Notes
This model has high MT-Bench score, comparable to GPT-3.5-Turbo. See original model card below.
# Original Model Card
---
license: apache-2.0
language:
- en
pipeline_tag: text-generation
tags:
- mistral
- mixtral
- solar
- model-fusion
- fusechat
library_name: transformers
---
<p align="center" width="100%">
</p>
<div id="top" align="center">
<p style="font-size: 32px; font-weight: bold;">FuseChat: Knowledge Fusion of Chat Models</p>
<h4> |<a href="https://arxiv.org/abs/2402.16107"> 📑 Paper </a> |
<a href="https://huggingface.co/FuseAI"> 🤗 Huggingface Repo </a> |
<a href="https://github.com/fanqiwan/FuseLLM"> 🐱 Github Repo </a> |
</h4>
<!-- **Authors:** -->
_**Fanqi Wan, Ziyi Yang, Longguang Zhong, Xiaojun Quan, Xinting Huang, Wei Bi**_
<!-- **Affiliations:** -->
_Sun Yat-sen University_
<p align="center">
<img src="./assets/fig_0.png" width="70%"> <br>
</p>
| Proprietary Models | #Params | MT-Bench | Open Source Models | #Params | MT-Bench |
|-----------------------------------------------------------------------|---------|----------|-----------------------------------------------------------------------|---------|----------|
| GPT-4-1106-preview | - | 9.32 | Qwen1.5-72B-Chat | 72B | 8.61 |
| GPT-4-0613 | - | 9.18 | Nous-Hermes-2-Mixtral-8x7B-DPO | 8x7B | 8.33 |
| GPT-4-0314 | - | 8.96 | Mixtral-8x7B-Instruct-v0.1 | 8x7B | 8.30 |
| Mistral Medium | - | 8.61 | 🤗 [FuseChat-7B-VaRM](https://huggingface.co/FuseAI/FuseChat-7B-VaRM) | 7B | 8.22 |
| GPT-3.5-Turbo-0613 | - | 8.39 | Starling-LM-7B-alpha | 7B | 8.09 |
| GPT-3.5-Turbo-1106 | - | 8.32 | Tulu-2-DPO-70B | 70B | 7.89 |
| 🤗 [FuseChat-7B-VaRM](https://huggingface.co/FuseAI/FuseChat-7B-VaRM) | 7B | 8.22 | OpenChat-3.5 | 7B | 7.81 |
| Claude-2.1 | - | 8.18 | OpenChat-3.5-0106 | 7B | 7.80 |
| Claude-2.0 | - | 8.06 | WizardLM-70B-v1.0 | 70B | 7.71 |
| GPT-3.5-Turbo-0314 | - | 7.94 | Yi-34B-Chat | 34B | 7.67 |
| Claude-1 | - | 7.90 | Nous-Hermes-2-SOLAR-10.7B | 10.7B | 7.66 |
</div>
## News
- **Feb 26, 2024:** 🔥 We release [FuseChat-7B-VaRM](https://huggingface.co/FuseAI/FuseChat-7B-VaRM), which is the fusion of three prominent chat LLMs with diverse architectures and scales, namely [NH2-Mixtral-8x7B](https://huggingface.co/NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO), [NH2-Solar-10.7B](https://huggingface.co/NousResearch/Nous-Hermes-2-SOLAR-10.7B), and [OpenChat-3.5-7B](https://huggingface.co/openchat/openchat_3.5). FuseChat-7B-VaRM achieves an average performance of **8.22** on MT-Bench, outperforming various powerful chat LLMs at 7B and 34B scales like [Starling-7B](https://huggingface.co/berkeley-nest/Starling-LM-7B-alpha) and [Yi-34B-Chat](https://huggingface.co/01-ai/Yi-34B-Chat), even surpassing [GPT-3.5 (March)](https://platform.openai.com/docs/models/gpt-3-5-turbo), [Claude-2.1](https://www.anthropic.com/news/claude-2-1), and approaching [Mixtral-8x7B-Instruct](https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1).
## Contents
- [Overview](#overview)
- [Model Release](#model-release)
- [Quick Start](#quick-start)
- [Data Construction](#data-construction)
- [Pairwise Knowledge Fusion](#pairwise-knowledge-fusion)
- [Model Merging](#model-merging)
- [Evaluation](#evaluation)
- [Citation](#citation)
## 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.
<p align="center">
<img src="./assets/fig_1.png" width="95%"> <br>
</p>
## Model Release
We release [FuseChat-7B-VaRM](https://huggingface.co/FuseAI/FuseChat-7B-VaRM), which is the fusion of three prominent chat LLMs with diverse architectures and scales, namely [NH2-Mixtral-8x7B](https://huggingface.co/NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO), [NH2-Solar-10.7B](https://huggingface.co/NousResearch/Nous-Hermes-2-SOLAR-10.7B), and [OpenChat-3.5-7B](https://huggingface.co/openchat/openchat_3.5). FuseChat-7B-VaRM achieves an average performance of **8.22** on MT-Bench, outperforming various powerful chat LLMs at 7B and 34B scales like [Starling-7B](https://huggingface.co/berkeley-nest/Starling-LM-7B-alpha) and [Yi-34B-Chat](https://huggingface.co/01-ai/Yi-34B-Chat), even surpassing [GPT-3.5 (March)](https://platform.openai.com/docs/models/gpt-3-5-turbo), [Claude-2.1](https://www.anthropic.com/news/claude-2-1), and approaching [Mixtral-8x7B-Instruct](https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1).
To support a plug-and-play fusion of new source LLM, we release our target LLMs: [OpenChat-3.5-7B-Solar](https://huggingface.co/FuseAI/OpenChat-3.5-7B-Solar) and [OpenChat-3.5-7B-Mixtral](https://huggingface.co/FuseAI/OpenChat-3.5-7B-Mixtral), 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.
We also release FuseChat with other merging methods: [FuseChat-7B-SLERP](https://huggingface.co/FuseAI/FuseChat-7B-SLERP) and [FuseChat-7B-TA](https://huggingface.co/FuseAI/FuseChat-7B-TA), which achieves an average performance of **8.19** and **8.20** on MT-Bench respectively.
Here are the evaluation results.
<p align="center">
<img src="./assets/tab_1.png" width="95%"> <br>
</p>
## Quick Start
### Setup
We use `python 3.11` in this project.
Then, we have to install all the libraries listed in `requirements.txt`.
```bash
pip install -r requirements.txt
```
### Usage
Here's how you can run the model using the 🤗 Transformers:
```python
import transformers
tokenizer = transformers.AutoTokenizer.from_pretrained("FuseAI/FuseChat-7B-VaRM")
# 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:
```python
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]
```
## Data Construction
We curated a comprehensive training dataset, [FuseChat-Mixture](https://huggingface.co/datasets/FuseAI/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.
Here we show the scripts to obtain representations from multiple source LLMs for model fusion.
1. Get representations for each source LLM
```bash
# We split the dataset into 4 splits, then process each split on one or multiple GPU.
# OpenChat-3.5-7B
export CUDA_VISIBLE_DEVICES=0
for i in {0..3}; do
python /train/get_data_representation.py \
--model_name_or_path "openchat/openchat_3.5" \
--data_path "/data/fusechat_v1_clean_split_2048_filter_wrong.json" \
--dataset_save_dir "<${i}_4_path_to_openchat_representation>" \
--tknz_dataset_path "<${i}_4_path_to_openchat_tknz>" \
--cache_dir "/.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 "openchat" \
--flash_attn_transformers \
--mask_instruction \
--dataset_split_num 4 \
--dataset_index ${i}
done
# NH2-Mixtral-8x7B
export CUDA_VISIBLE_DEVICES=0,1,2
for i in {0..3}; do
python /train/get_data_representation.py \
--model_name_or_path "NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO" \
--data_path "/data/fusechat_v1_clean_split_2048_filter_wrong.json" \
--dataset_save_dir "<${i}_4_path_to_mixtral_representation>" \
--tknz_dataset_path "<${i}_4_path_to_mixtral_tknz>" \
--cache_dir "/.cache/huggingface/datasets" \
--model_max_length 2048 \
--load_in_half bf16 \
--batch_size 4 \
--top_k_logits 10 \
--save_per_token_metric \
--no_assert \
--conv_temp "openchat" \
--flash_attn_transformers \
--mask_instruction \
--device_map "auto" \
--dataset_split_num 4 \
--dataset_index ${i}
done
# NH2-Solar-10.7B
export CUDA_VISIBLE_DEVICES=0
for i in {0..3}; do
python /train/get_data_representation.py \
--model_name_or_path "NousResearch/Nous-Hermes-2-SOLAR-10.7B" \
--data_path "/data/fusechat_v1_clean_split_2048_filter_wrong.json" \
--dataset_save_dir "<${i}_4_path_to_solar_representation>" \
--tknz_dataset_path "<${i}_4_path_to_solar_tknz>" \
--cache_dir "/.cache/huggingface/datasets" \
--model_max_length 2048 \
--load_in_half bf16 \
--batch_size 8 \
--top_k_logits 10 \
--save_per_token_metric \
--no_assert \
--conv_temp "openchat" \
--flash_attn_transformers \
--mask_instruction \
--dataset_split_num 4 \
--dataset_index ${i}
done
```
2. Align representations from different source LLMs
```bash
# Since the tokenizers and vocabularies of these source LLMs are identical, we do not align.
# OpenChat-3.5-7B <-> NH2-Mixtral-8x7B
for i in {0..3}; do
python /train/replace_model.py \
--dataset_dir "<${i}_4_path_to_openchat_representation>" \
--replace_dataset_dir "<${i}_4_path_to_mixtral_representation>" \
--dataset_save_dir "<${i}_4_path_to_openchat_mixtral_representation>" \
--preprocessing_num_workers 64 \
--batch_size 1000 \
--replace_model model_0
done
# OpenChat-3.5-7B <-> NH2-Solar-10.7B
for i in {0..3}; do
python /train/replace_model.py \
--dataset_dir "<${i}_4_path_to_openchat_mixtral_representation>" \
--replace_dataset_dir "<${i}_4_path_to_solar_representation>" \
--dataset_save_dir "<${i}_4_path_to_openchat_mixtral_solar_representation>" \
--preprocessing_num_workers 64 \
--batch_size 1000 \
--replace_model model_1
done
```
3. Filter instances with NaN loss in the dataset
```bash
for i in {0..3}; do
python /train/filter_nan.py \
--input_data_dir "<${i}_4_path_to_openchat_mixtral_solar_representation>" \
--output_data_dir "<${i}_4_path_to_openchat_mixtral_solar_representation_fnan>"
done
```
The final processed data is at `<${i}_4_path_to_openchat_mixtral_solar_representation_fnan>`.
## Pairwise Knowledge Fusion
We show the scripts for pairwise knowledge fusion.
```bash
# OpenChat-3.5-7B <-> NH2-Mixtral-8x7B
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
torchrun --nproc_per_node=8 --master_port=20001 /train/train.py \
--model_name_or_path "openchat/openchat_3.5" \
--data_path "<0_4_path_to_openchat_mixtral_solar_representation_fnan>,<1_4_path_to_openchat_mixtral_solar_representation_fnan>,<2_4_path_to_openchat_mixtral_solar_representation_fnan>,<3_4_path_to_openchat_mixtral_solar_representation_fnan>" \
--bf16 True \
--output_dir "<path_to_save_openchat_mixtral_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" \
--lazy_preprocess True \
--flash_attn_transformers True \
--do_train \
--do_distill \
--distill_with_ref_model True \
--distill_with_aligned_model_0 True \
--distill_with_aligned_model_1 False \
--distill_loss_type "ce" \
--distill_teacher_temperature 1.0 \
--lm_loss_weight 0.9 \
--distill_greater_as_gt True \
--distill_greater_as_gt_type hard \
--dataloader_num_workers 8 \
--remove_unused_columns False
# OpenChat-3.5-7B <-> NH2-Solar-10.7B
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
torchrun --nproc_per_node=8 --master_port=20001 /train/train.py \
--model_name_or_path "openchat/openchat_3.5" \
--data_path "<0_4_path_to_openchat_mixtral_solar_representation_fnan>,<1_4_path_to_openchat_mixtral_solar_representation_fnan>,<2_4_path_to_openchat_mixtral_solar_representation_fnan>,<3_4_path_to_openchat_mixtral_solar_representation_fnan>" \
--bf16 True \
--output_dir "<path_to_save_openchat_solar_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" \
--lazy_preprocess True \
--flash_attn_transformers True \
--do_train \
--do_distill \
--distill_with_ref_model True \
--distill_with_aligned_model_0 False \
--distill_with_aligned_model_1 True \
--distill_loss_type "ce" \
--distill_teacher_temperature 1.0 \
--lm_loss_weight 0.9 \
--distill_greater_as_gt True \
--distill_greater_as_gt_type hard \
--dataloader_num_workers 8 \
--remove_unused_columns False
```
## Model Merging
We show the scripts to obtain the final FuseChat using different merging methods.
```bash
# For "slerp", "ta", "ties", and "dare" methods (Please install "mergekit")
export CUDA_VISIBLE_DEVICES=0
mergekit-yaml merge/mergekit_configs/fusechat-slerp.yml "<path_to_save_fusechat_7b_slerp>"
mergekit-yaml merge/mergekit_configs/fusechat-ta.yml "<path_to_save_fusechat_7b_ta>"
mergekit-yaml merge/mergekit_configs/fusechat-ties.yml "<path_to_save_fusechat_7b_ties>"
mergekit-yaml merge/mergekit_configs/fusechat-dare.yml "<path_to_save_fusechat_7b_dare>"
# For "linear" method
python merge/VaRM/merge.py \
--merged_model_names "FuseAI/OpenChat-3.5-7B-Mixtral,FuseAI/OpenChat-3.5-7B-Solar" \
--merged_model_save_dir "<path_to_save_fusechat_7b_linear>" \
--merge_method "linear" \
--linear_weights "1,2"
# For our "varm" method
python merge/VaRM/analysis.py \
--model1_path "FuseAI/OpenChat-3.5-7B-Mixtral" \
--model2_path "FuseAI/OpenChat-3.5-7B-Solar" \
--save_path "<path_to_save_analysis_result>/analysis.json" \
--merge_type "square"
python merge/VaRM/merge.py \
--merged_model_names "FuseAI/OpenChat-3.5-7B-Mixtral,FuseAI/OpenChat-3.5-7B-Solar" \
--analysis_result "<path_to_save_analysis_result>/analysis.json" \
--merged_model_save_dir "<path_to_save_fusechat_7b_varm>" \
--merge_method "avg_param" \
--merge_type "square"
```
## Evaluation
We evaluate FuseChat on MT-Bench, which comprises 80 multi-turn dialogues spanning writing, roleplay, reasoning, math, coding, stem, and humanities domains. Please download the [official code](https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge) and follow the guidelines for evaluation. We provide the scripts for our evaluation.
```bash
# Step 1. Generate model answers to MT-bench questions
export CUDA_VISIBLE_DEVICES=0,1
python gen_model_answer.py \
--model-path "FuseAI/FuseChat-7B-VaRM" \
--model-id "openchat_3.5_fusechat_7b_varm" \
--num-gpus-per-model 1 \
--num-gpus-total 2
# Step 2. Generate GPT-4 judgments
export OPENAI_API_KEY=XXXXXX # set the OpenAI API key
python gen_judgment.py \
--parallel 2
# Step 3. Show MT-bench scores
python show_result.py
```
## 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 Ziyi Yang and Longguang Zhong and Xiaojun Quan and Xinting Huang and Wei Bi},
journal={arXiv preprint arXiv:2402.16107},
year={2024}
}
``` |