-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathindex.html
1062 lines (939 loc) · 44.8 KB
/
index.html
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
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<meta name="description" content="FuseChat-3.0: Preference Optimization for Implicit Model Fusion">
<meta property="og:title" content="FuseChat-3.0: Preference Optimization for Implicit Model Fusion"/>
<meta property="og:description" content="FuseChat-3.0 introduces implicit model fusion (IMF) with preference optimization to transfer capabilities of source models into target models."/>
<meta property="og:url" content="URL OF THE WEBSITE"/>
<meta property="og:image" content="static/images/your_banner_image.png" />
<meta property="og:image:width" content="1200"/>
<meta property="og:image:height" content="630"/>
<meta name="twitter:title" content="FuseChat-3.0: Preference Optimization for Implicit Model Fusion">
<meta name="twitter:description" content="FuseChat-3.0 introduces implicit model fusion (IMF) with preference optimization to transfer capabilities of source models into target models.">
<meta name="twitter:image" content="static/images/your_twitter_banner_image.png">
<meta name="twitter:card" content="summary_large_image">
<meta name="keywords" content="FuseChat-3.0, Implicit Model Fusion, Preference Optimization, LLMs, SFT, DPO">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>FuseChat-3.0: Preference Optimization for Implicit Model Fusion</title>
<link rel="icon" type="image/x-icon" href="static/images/fuseai.jpg">
<link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro" rel="stylesheet">
<link rel="stylesheet" href="static/css/bulma.min.css">
<link rel="stylesheet" href="static/css/bulma-carousel.min.css">
<link rel="stylesheet" href="static/css/bulma-slider.min.css">
<link rel="stylesheet" href="static/css/fontawesome.all.min.css">
<link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
<link rel="stylesheet" href="static/css/index.css">
<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
<script src="https://documentcloud.adobe.com/view-sdk/main.js"></script>
<script defer src="static/js/fontawesome.all.min.js"></script>
<script src="static/js/bulma-carousel.min.js"></script>
<script src="static/js/bulma-slider.min.js"></script>
<script src="static/js/index.js"></script>
</head>
<body>
<section class="hero">
<div class="hero-body">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column has-text-centered">
<h1 class="title is-1 publication-title">FuseChat-3.0: Preference Optimization for Implicit Model Fusion</h1>
<div class="is-size-5 publication-authors">
<span class="author-block">
<a href="https://github.com/yangzy39" target="_blank">Ziyi Yang</a><sup>*</sup>,
</span>
<span class="author-block">
<a href="https://fanqiwan.github.io/" target="_blank">Fanqi Wan</a><sup>*</sup>,
</span>
<span class="author-block">
<a href="https://github.com/MrGGLS" target="_blank">Longguang Zhong</a><sup>*</sup>,
</span>
<span class="author-block">
Canbin Huang,
</span>
<span class="author-block">
Guosheng Liang,
</span>
<span class="author-block">
<a href="https://sites.google.com/site/xiaojunquan" target="_blank">Xiaojun Quan</a><sup>†</sup>
</span>
</div>
<div class="is-size-5 publication-authors">
<span class="author-block">School of Computer Science and Engineering, Sun Yat-sen University, China</span>
<span class="eql-cntrb"><small><br><sup>*</sup>Main Contributors <sup>†</sup>Corresponding Author</small></span>
</div>
<div class="column has-text-centered">
<div class="publication-links">
<span class="link-block">
<a href="https://huggingface.co/FuseAI" target="_blank" class="external-link button is-normal is-rounded is-dark">
<span class="icon">🤗</span>
<span>Hugging Face</span>
</a>
</span>
<span class="link-block">
<a href="https://github.com/SLIT-AI/FuseChat-3.0" target="_blank" class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fab fa-github"></i>
</span>
<span>Code</span>
</a>
</span>
<span class="link-block">
<a href="https://arxiv.org/abs/2412.03187" target="_blank" class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="ai ai-arxiv"></i>
</span>
<span>arXiv</span>
</a>
</span>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<!-- </div>
<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column is-four-fifths">
<div class="content has-text-centered">
<img src="static/images/FuseChat-3.0.png" width=80%/>
</div>
</div>
</div>
</div> -->
<!-- Paper abstract -->
<section class="hero teaser">
<div class="container is-max-desktop">
<div class="hero-body">
<div align="center">
<img src="static/images/FuseChat-3.0.png" width=70%/>
</div>
<p>
In this blog post, we present FuseChat-3.0, a series of models crafted to enhance performance by integrating the strengths of multiple source LLMs into more compact target LLMs. To achieve this fusion, we utilized four powerful source LLMs: Gemma-2-27B-It, Mistral-Large-Instruct-2407, Qwen-2.5-72B-Instruct, and Llama-3.1-70B-Instruct. For the target LLMs, we employed three widely-used smaller models—Llama-3.1-8B-Instruct, Gemma-2-9B-It, and Qwen-2.5-7B-Instruct—along with two even more compact models—Llama-3.2-3B-Instruct and Llama-3.2-1B-Instruct. The implicit model fusion process involves a two-stage training pipeline comprising Supervised Fine-Tuning (SFT) to mitigate distribution discrepancies between target and source LLMs, and Direct Preference Optimization (DPO) for learning preferences from multiple source LLMs. The resulting FuseChat-3.0 models demonstrated substantial improvements in tasks related to general conversation, instruction following, mathematics, and coding. Notably, when Llama-3.1-8B-Instruct served as the target LLM, our fusion approach achieved an average improvement of 6.8 points across 14 benchmarks. Moreover, it showed significant improvements of 37.1 and 30.1 points on instruction-following test sets AlpacaEval-2 and Arena-Hard respectively. We have released the <a href="https://huggingface.co/FuseAI" target="_blank">FuseChat-3.0</a> models on Huggingface, stay tuned for the forthcoming dataset and code.
</p>
</div>
</div>
</section>
<!-- End paper abstract -->
<!-- Paper overview -->
<section class="section hero is-light">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3">Overview</h2>
<div class="content has-text-justified">
<p>
Combining the strengths of multiple large language models (LLMs) represents a promising approach to enhance individual model capabilities. Model fusion is a technique that integrates the strengths of robust source LLMs into a target LLM.
</p>
<p>
Previous iterations of the <a href="https://arxiv.org/abs/2408.07990">FuseChat</a> series employed probabilistic distribution matrices generated by source models to transfer knowledge to target models. We refer to this method as <strong>explicit model fusion (EMF)</strong> because it involves a well-defined knowledge transfer process. While applicable to models with varying architectures and sizes, and without increasing memory overhead during inference, this approach presents notable challenges such as vocabulary alignment and the merging of distribution matrices from different LLMs. These issues complicate model fusion, reduce its efficiency, and may introduce noise and errors and affect the fusion results.
</p>
<p>
FuseChat-3.0, however, takes a different approach by enhancing a single LLM through implicit learning from robust open-source LLMs, a process we term <strong>implicit model fusion (IMF)</strong>. The concept of IMF has been widely utilized to improve the performance of weaker models. For instance, a weak model can be boosted through fine-tuning with outputs from stronger LLMs. Moreover, a reward model can be trained using outputs from various LLMs, enabling it to learn and capture the differences in capabilities between the LLMs. Zephyr further collects responses from multiple LLMs and ranks them with GPT-4 to obtain preference data for training the policy. Inspired by recent alignment techniques, we propose an IMF method to transfer the capabilities of source LLMs to a target LLM through preference optimization.
</p>
<p>
Our IMF method follows a three-stage process aimed at effectively transferring capabilities from source LLMs to a target LLM. First, during <strong>dataset construction</strong>, we sample N responses from each of the source LLMs and annotate these responses using an external reward model. Second, in the <strong>supervised fine-tuning (SFT)</strong> stage, we fine-tune the target model using the best responses, which not only enhances the target model's capabilities but also helps mitigate the distributional gap between the source and target models. Finally, in the <strong>direct preference optimization (DPO)</strong> stage, we optimize the target model by using the best and worst responses from the source models as preference pairs, further enhancing the target model's performance. The complete pipeline will be detailed in the following paragraph.
</p>
</div>
</div>
</div>
</div>
</section>
<!-- End paper overview -->
<!-- Dataset Construction -->
<section class="section hero is-light">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3">Dataset</h2>
<div class="content has-text-justified">
<h3 class="title is-4">Prompt Selection</h3>
<p>
Our datasets were designed to enhance model's instruction following, general conversation, mathematics, coding, and Chinese-language capabilities. We selected data from open-source community datasets, applying targeted filtering and preprocessing. Key datasets and filtering criteria included:
</p>
<ul>
<li><strong>Instruction Following & General Conversation</strong>: Sourced from <a href="https://huggingface.co/datasets/openbmb/UltraFeedback">UltraFeedback</a>, <a href="https://huggingface.co/datasets/Magpie-Align/Magpie-Pro-DPO-100K-v0.1" target="_blank">Magpie-Pro-DPO-100K-v0.1</a>, and <a href="https://huggingface.co/datasets/nvidia/HelpSteer2" target="_blank">HelpSteer2</a>, excluding code and math data.</li>
<li><strong>Mathematics</strong>: Selected from <a href="https://huggingface.co/datasets/nvidia/OpenMathInstruct-2" target="_blank">OpenMathInstruct-2</a>, with nearly 60,000 unique samples.</li>
<li><strong>Coding</strong>: Curated from <a href="https://huggingface.co/datasets/greengerong/leetcode" target="_blank">leetcode</a> and <a href="https://huggingface.co/datasets/bigcode/self-oss-instruct-sc2-exec-filter-50k" target="_blank">self-oss-instruct-sc2-exec-filter-50k</a>, retaining prompts with test cases.</li>
<li><strong>Chinese Language</strong>: Integrated <a href="https://huggingface.co/datasets/llamafactory/alpaca_gpt4_zh" target="_blank">alpaca_gpt4_zh</a> and <a href="https://huggingface.co/datasets/Magpie-Align/Magpie-Qwen2-Pro-200K-Chinese" target="_blank">Magpie-Qwen2-Pro-200K-Chinese</a>, filtering out code and math prompts to retain approximately 10,000 high-quality samples.</li>
</ul>
<h3 class="title is-4">Response Sampling</h3>
<p>
For each dataset's prompts, we synthesized responses mainly from four different series of source models, specifically <a href="https://huggingface.co/google/gemma-2-27b-it" target="_blank">Gemma-2-27b-It</a>, <a href="https://huggingface.co/mistralai/Mistral-Large-Instruct-2407" target="_blank">Mistral-Large-Instruct-2407</a>, <a href="https://huggingface.co/Qwen/Qwen2-72B-Instruct" target="_blank">Qwen-2.5-72B-Instruct</a>, and <a href="https://huggingface.co/meta-llama/Llama-3.1-70B-Instruct" target="_blank">Llama-3.1-70B-Instruct</a>. Compared to the original approach in <a href="https://arxiv.org/abs/2412.03187" target="_blank">WRPO</a> using up to ten source LLMs, this streamlined selection ensures both representativeness and accessibility of the experiments.
</p>
<ul>
<li><strong>Instruction Following & General Conversation</strong>: We sampled each prompt five times from all the source models.</li>
<li><strong>Mathematics</strong>: We retained the responses generated by Llama-3.1-405B-Instruct from the original dataset (OpenMathInstruct-2) and additionally sampled responses using <a href="https://huggingface.co/Qwen/Qwen-2.5-Math-72B-Instruct" target="_blank">Qwen-2.5-Math-72B-Instruct</a>.</li>
<li><strong>Coding</strong>: We sampled each prompt eight times for all source models.</li>
<li><strong>Chinese Language</strong>: We included single response sampled exclusively from Qwen-2.5-72B-Instruct.</li>
</ul>
<p>
The sampling parameters for different models are detailed in Table below.
</p>
<table class="js-sort-table table hidden">
<tr>
<td class="js-sort-string"><strong>Source LLMs</strong></td>
<td class="js-sort-string"><strong>Sampling Params</strong></td>
</tr>
<tr>
<td>Gemma-2-27b-It</td>
<td>Temp 0.8 Top-p 0.95</td>
</tr>
<tr>
<td>Mistral-Large-Instruct-2407</td>
<td>Temp 0.8 Top-p 0.95</td>
</tr>
<tr>
<td>Qwen-2.5-(Math)-72B-Instruct</td>
<td>Temp 0.7 Top-p 0.8 Repetition penalty 1.05</td>
</tr>
<tr>
<td>Llama-3.1-70B-Instruct</td>
<td>Temp 0.8 Top-p 0.95</td>
</tr>
</table>
<h3 class="title is-4">Data Construction</h3>
<p>
Unlike the original approach in <a href="https://arxiv.org/abs/2412.03187" target="_blank">WRPO</a>, which constructs preference pairs from target model responses and treats source model responses as additional positive samples, our research in mathematics and coding domains revealed that sampling from multiple source models yields more and higher-quality preference pair data. Based on this insight, FuseChat-3.0 leverages the best and worst response pairs generated by source models as preference pairs to optimize the target model. This refined approach not only preserves the core advantages of implicit model fusion but also results in a more streamlined and practical implementation, making it particularly well-suited for real-world applications within the open-source community.
</p>
<p>
<strong>Instruction Following</strong>: To assign RM scores to the five responses generated by each source model, we employed <a href="https://huggingface.co/RLHFlow/ArmoRM-Llama3-8B-v0.1" target="_blank">ArmoRM</a> for annotation. We then divided the annotated data into SFT and DPO datasets using a 4:6 ratio. For the SFT phase, we selected the responses with the highest RM scores. During the DPO phase, we paired responses from the same source model, designating those with the highest RM scores as positive samples and those with the lowest RM scores as negative samples. We ensured that the RM score difference between the positive and negative samples in each pair ranged from 0.01 to 0.1.
</p>
<p>
<strong>Mathematics</strong>: We initially annotated the responses from all source models for correctness by comparing them with the gold labels and evaluating them using the RM scores provided by ArmoRM. We then strategically divided the dataset into SFT phase and DPO phase. In the SFT phase, we incorporated responses that were correct and had the highest RM scores. This selection ensured that the fine-tuning process was based on high-quality responses that aligned closely with the desired outcomes. For the DPO phase, we constructed paired samples from the same source model. The positive samples consisted of correct answers with the highest RM scores, while the negative samples were incorrect answers with the lowest RM scores. To ensure meaningful comparisons during optimization, we maintained an RM score differential between positive and negative pairs within the range of 0.01 to 0.1.
</p>
<p>
<strong>Coding</strong>: We employed a dual-scoring system comprising correctness scores and RM scores for coding evaluation. The correctness scores assessed whether the code passed both static analysis and test cases, ensuring functional accuracy. The RM scores were used for preference evaluation, gauging the quality of responses based on predefined criteria. During the SFT phase, we included responses that not only passed all test cases but also achieved the highest RM scores. This selection ensured that the model was fine-tuned on exemplary code that met both correctness and preference standards. In the DPO phase, we contrasted positive samples—high-scoring responses that passed the tests—with negative samples—low-scoring responses that failed the tests. This comparison aimed to optimize the model's ability to prefer higher-quality code during training. We excluded any instances where all model responses failed to meet the testing criteria. This exclusion was necessary to maintain the integrity of the evaluation process, as such cases did not provide meaningful data for assessing and improving the model's performance.
</p>
<p>
<strong>Chinese</strong>: We exclusively utilized responses sampled from Qwen-2.5-72B-Instruct during the SFT phase, due to its strong performance in the Chinese language.
</p>
<p>
Our final dataset comprised 158,784 total entries, with 94,539 entries for the SFT phase and 64,245 preference pairs for the DPO phase. The overall composition of the datasets is shown below.
</p>
<table class="js-sort-table table hidden">
<tr>
<td class="js-sort-string"><strong>Dataset</strong></td>
<td class="js-sort-number"><strong>Total Count</strong></td>
<td class="js-sort-number"><strong>SFT Count</strong></td>
<td class="js-sort-number"><strong>DPO Count</strong></td>
<td class="js-sort-string"><strong>Category</strong></td>
</tr>
<tr>
<td><a href="https://huggingface.co/datasets/openbmb/UltraFeedback" target="_blank">UltraFeedback</a></td>
<td>51098</td>
<td>20439</td>
<td>30659</td>
<td>Instruction following</td>
</tr>
<tr>
<td><a href="https://huggingface.co/datasets/Magpie-Align/Magpie-Pro-DPO-100K-v0.1" target="_blank">Magpie-Pro-DPO</a></td>
<td>20374</td>
<td>8149</td>
<td>12225</td>
<td>Instruction following</td>
</tr>
<tr>
<td><a href="https://huggingface.co/datasets/nvidia/HelpSteer2" target="_blank">HelpSteer2</a></td>
<td>9435</td>
<td>3774</td>
<td>5661</td>
<td>Instruction following</td>
</tr>
<tr>
<td><a href="https://huggingface.co/datasets/nvidia/OpenMathInstruct-2" target="_blank">OpenMathInstruct-2</a></td>
<td>58546</td>
<td>40188</td>
<td>11615</td>
<td>Mathematics</td>
</tr>
<tr>
<td><a href="https://huggingface.co/datasets/greengerong/leetcode" target="_blank">leetcode</a></td>
<td>3113</td>
<td>1877</td>
<td>1236</td>
<td>Coding</td>
</tr>
<tr>
<td><a href="https://huggingface.co/datasets/bigcode/self-oss-instruct-sc2-exec-filter-50k" target="_blank">self-oss-instruct-sc2</a></td>
<td>13696</td>
<td>10160</td>
<td>2849</td>
<td>Coding</td>
</tr>
<tr>
<td><a href="https://huggingface.co/datasets/llamafactory/alpaca_gpt4_zh" target="_blank">alpaca_gpt4_zh</a></td>
<td>2471</td>
<td>2471</td>
<td>0</td>
<td>Chinese Language</td>
</tr>
<tr>
<td><a href="https://huggingface.co/datasets/Magpie-Align/Magpie-Qwen2-Pro-200K-Chinese" target="_blank">Magpie-Qwen2-Pro</a></td>
<td>7481</td>
<td>7481</td>
<td>0</td>
<td>Chinese Language</td>
</tr>
<tr>
<td><strong>Total</strong></td>
<td>158784</td>
<td>94539</td>
<td>64245</td>
<td>All</td>
</tr>
</table>
</div>
</div>
</div>
</div>
</section>
<!-- End Dataset Construction -->
<!-- Training Pipeline -->
<section class="section hero is-light">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3">Training</h2>
<div class="content has-text-justified">
<p>
The implicit model fusion process involves a two-stage training pipeline comprising Supervised Fine-Tuning (SFT) to mitigate distribution discrepancies between target and source LLMs, and Direct Preference Optimization (DPO) for learning preferences from multiple source LLMs.
</p>
<h3 class="title is-4">SFT</h3>
<p>
We used <a href="https://github.com/hiyouga/LLaMA-Factory" target="_blank">Llama-Factory</a> as our fine-tuning library. For all target models, we fine-tuned for 3 epochs, with a batch size of 128 and a maximum sequence length of 2048 tokens. A cosine learning rate schedule with a warmup ratio of 0.1 is employed. Different models' learning rates are shown in the table below.
</p>
<table class="js-sort-table table hidden">
<tr>
<td class="js-sort-string"><strong>Target Models</strong></td>
<td class="js-sort-string"><strong>Learning rate</strong></td>
</tr>
<tr>
<td>Llama-3.1-8B-Instruct</td>
<td>5e-6</td>
</tr>
<tr>
<td>Qwen-2.5-7B-Instruct</td>
<td>2e-6</td>
</tr>
<tr>
<td>Gemma-2-9B-It</td>
<td>2e-6</td>
</tr>
<tr>
<td>Llama-3.2-(1/3)B-Instruct</td>
<td>5e-6</td>
</tr>
</table>
<h3 class="title is-4">DPO</h3>
<p>
We used <a href="https://github.com/huggingface/alignment-handbook" target="_blank">alignment-handbook</a> as our DPO training library. For all Target SFT models, we trained for 1 epoch with a batch size of 128, set maximum sequence length to 2048, used cosine learning rate with a warmup ratio of 0.1. We saved checkpoints every 100 steps and selected the best from the last two checkpoints. For Llama-3.1 and Llama-3.2 series models, we introduced length normalization in DPO training, as shown in the formula below.
</p>
<p>
<img src="https://latex.codecogs.com/svg.image?\mathcal{L}_{\text{LN-DPO}}=-\log\sigma\left(\frac{\beta}{|y_w|}\log\frac{\pi_\theta(y_w|x)}{\pi_{\text{ref}}(y_w|x)}-\frac{\beta}{|y_l|}\log\frac{\pi_\theta(y_l|x)}{\pi_{\text{ref}}(y_l|x)}\right)" alt="Length Normalized DPO Formula" />
</p>
<p>
Different models' hyperparameters are shown in the table below.
</p>
<table class="js-sort-table table hidden">
<tr>
<td class="js-sort-string"><strong>Target SFT Models</strong></td>
<td class="js-sort-string"><strong>Learning rate</strong></td>
<td class="js-sort-string"><strong>β</strong></td>
<td class="js-sort-string"><strong>Length normalize</strong></td>
</tr>
<tr>
<td>FuseChat-Llama-3.1-8B-SFT</td>
<td>8e-7</td>
<td>10</td>
<td>Yes</td>
</tr>
<tr>
<td>FuseChat-Qwen-2.5-7B-SFT</td>
<td>3e-7</td>
<td>0.01</td>
<td>No</td>
</tr>
<tr>
<td>FuseChat-Gemma-2-9B-SFT</td>
<td>5e-7</td>
<td>0.01</td>
<td>No</td>
</tr>
<tr>
<td>FuseChat-Llama-3.2-(1/3)B-SFT</td>
<td>1e-6</td>
<td>10</td>
<td>Yes</td>
</tr>
</table>
</div>
</div>
</div>
</div>
</section>
<!-- End Training Pipeline -->
<!-- Evaluation -->
<section class="section hero is-light">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3">Evaluation</h2>
<div class="content has-text-justified">
<p>
The evaluation of instruction-tuned models mainly focuses on the model performance of instruction following, natural language understanding, general question answering, reasoning, mathematics, coding, etc. For the evaluation of FuseChat-3.0, we include 14 benchmarks and organize them into four categories:
</p>
<ul>
<li><strong>Instruction Following</strong> Tasks: AlpacaEval-2, Arena-Hard, MTbench, AlignBench v1.1 (Chinese).</li>
<li><strong>General</strong> Tasks: LiveBench-0831, MMLU-Pro, MMLU-redux, GPQA-Diamond.</li>
<li><strong>Mathematics</strong> Tasks: GSM8K, MATH, AMC 23.</li>
<li><strong>Coding</strong> Tasks: HumanEval, MBPP, LiveCodeBench 2408-2411.</li>
</ul>
<p>
We include more details and release our evaluation code at <a href="https://github.com/SLIT-AI/FuseChat-3.0/tree/main/FuseEval" target="_blank">FuseEval</a>.
</p>
<p>
The evaluation results of five series fused models are as follows, showing that our FuseChat-3.0 models achieved varying degrees of improvement across different target models. When selecting Llama-3.1-8B-Instruct as the target model, our fusion model <strong>FuseChat-Llama-3.1-8B-Instruct achieved an average performance improvement of 6.8 points across 14 benchmarks. Notably, it showed significant improvements of 37.1 and 30.1 points on instruction-following test sets AlpacaEval-2 and Arena-Hard respectively</strong>. Additionally, FuseChat-Llama-3.1-8B-Instruct outperformed AllenAI's recently released Llama-3.1-Tulu-3-8B model on all benchmarks except GSM8K and GPQA-Diamond. All these results demonstrate the effectiveness and success of FuseChat-3.0.
</p>
<h3 class="title is-4">FuseChat-Llama-3.1-8B-Instruct Performance</h3>
<table class="js-sort-table table hidden">
<tr>
<td class="js-sort-string"><strong>Benchmarks</strong></td>
<td class="js-sort-string"><strong>Llama-3.1-8B-Instruct</strong></td>
<td class="js-sort-string"><strong>Llama-3.1-Tulu-3-8B</strong></td>
<td class="js-sort-string"><strong>FuseChat-Llama-3.1-8B-SFT</strong></td>
<td class="js-sort-string"><strong>FuseChat-Llama-3.1-8B-Instruct</strong></td>
</tr>
<tr>
<td style="white-space: nowrap;">AlpacaEval-2 (LC %)</td>
<td>28.3</td>
<td>33.4</td>
<td>41.3</td>
<td><strong>65.4</strong></td>
</tr>
<tr>
<td>Arena-Hard (WR %)</td>
<td>28.1</td>
<td>45.6</td>
<td>38.7</td>
<td><strong>58.2</strong></td>
</tr>
<tr>
<td>MT-Bench</td>
<td>8.4</td>
<td>8.3</td>
<td>8.5</td>
<td><strong>9.0</strong></td>
</tr>
<tr>
<td>AlignBench v1.1</td>
<td>4.6</td>
<td>6.2</td>
<td>6.3</td>
<td><strong>6.7</strong></td>
</tr>
<tr>
<td>GSM8K</td>
<td>85.9</td>
<td><strong>88.6</strong></td>
<td>87.0</td>
<td>88.0</td>
</tr>
<tr>
<td>MATH</td>
<td>50.7</td>
<td>47.5</td>
<td>54.7</td>
<td><strong>55.2</strong></td>
</tr>
<tr>
<td>AMC 23</td>
<td>25.0</td>
<td>25.0</td>
<td>30.0</td>
<td><strong>37.5</strong></td>
</tr>
<tr>
<td>LiveBench 0831</td>
<td>27.6</td>
<td>30.1</td>
<td>30.2</td>
<td><strong>32.0</strong></td>
</tr>
<tr>
<td>MMLU-Pro</td>
<td><strong>50.0</strong></td>
<td>42.9</td>
<td>47.8</td>
<td>49.2</td>
</tr>
<tr>
<td>MMLU-redux</td>
<td>67.2</td>
<td>66.3</td>
<td>68.4</td>
<td><strong>69.2</strong></td>
</tr>
<tr>
<td>GPQA-Diamond</td>
<td>33.8</td>
<td>35.9</td>
<td><strong>37.9</strong></td>
<td>34.9</td>
</tr>
<tr>
<td>HumanEval</td>
<td>69.5</td>
<td>66.5</td>
<td>69.5</td>
<td><strong>71.3</strong></td>
</tr>
<tr>
<td>MBPP</td>
<td><strong>75.4</strong></td>
<td>56.3</td>
<td>71.4</td>
<td>72.0</td>
</tr>
<tr>
<td>LiveCodeBench<br>2408-2411</td>
<td>12.3</td>
<td>10.6</td>
<td>12.6</td>
<td><strong>13.1</strong></td>
</tr>
<tr>
<td>Average</td>
<td>40.5</td>
<td>40.2</td>
<td>43.2</td>
<td><strong>47.3</strong></td>
</tr>
</table>
<h3 class="title is-4">FuseChat-Qwen-2.5-7B-Instruct Performance</h3>
<table class="js-sort-table table hidden">
<tr>
<td class="js-sort-string"><strong>Datasets</strong></td>
<td class="js-sort-string"><strong>Qwen-2.5-7B-Instruct</strong></td>
<td class="js-sort-string"><strong>FuseChat-Qwen-2.5-7B-SFT</strong></td>
<td class="js-sort-string"><strong>FuseChat-Qwen-2.5-7B-Instruct</strong></td>
</tr>
<tr>
<td style="white-space: nowrap;">AlpacaEval-2 (LC %)</td>
<td>33.2</td>
<td>34.2</td>
<td><strong>63.6</strong></td>
</tr>
<tr>
<td>Arena-Hard (WR %)</td>
<td>50.7</td>
<td>45.2</td>
<td><strong>61.4</strong></td>
</tr>
<tr>
<td>MT-Bench</td>
<td>8.4</td>
<td>8.5</td>
<td><strong>9.0</strong></td>
</tr>
<tr>
<td>AlignBench v1.1</td>
<td>7.5</td>
<td>7.4</td>
<td><strong>7.6</strong></td>
</tr>
<tr>
<td>GSM8K</td>
<td>91.7</td>
<td><strong>92.3</strong></td>
<td>91.7</td>
</tr>
<tr>
<td>MATH</td>
<td><strong>75.0</strong></td>
<td>72.7</td>
<td>73.6</td>
</tr>
<tr>
<td>AMC 23</td>
<td>52.5</td>
<td>45.0</td>
<td><strong>57.5</strong></td>
</tr>
<tr>
<td>LiveBench-0831</td>
<td><strong>35.4</strong></td>
<td>33.7</td>
<td>33.2</td>
</tr>
<tr>
<td>MMLU-Pro</td>
<td><strong>54.1</strong></td>
<td>51.7</td>
<td>53.0</td>
</tr>
<tr>
<td>MMLU-redux</td>
<td><strong>75.1</strong></td>
<td>72.7</td>
<td>74.4</td>
</tr>
<tr>
<td>GPQA-Diamond</td>
<td>34.9</td>
<td><strong>38.4</strong></td>
<td>33.8</td>
</tr>
<tr>
<td>HumanEval</td>
<td><strong>85.4</strong></td>
<td>81.7</td>
<td>79.9</td>
</tr>
<tr>
<td>MBPP</td>
<td>80.2</td>
<td><strong>84.1</strong></td>
<td>83.1</td>
</tr>
<tr>
<td>LiveCodeBench<br>2408-2411</td>
<td>15.8</td>
<td>17.3</td>
<td><strong>18.9</strong></td>
</tr>
<tr>
<td>Average</td>
<td>50.0</td>
<td>48.9</td>
<td><strong>52.9</strong></td>
</tr>
</table>
<h3 class="title is-4">FuseChat-Gemma-2-9B-Instruct Performance</h3>
<table class="js-sort-table table hidden">
<tr>
<td class="js-sort-string"><strong>Datasets</strong></td>
<td class="js-sort-string"><strong>Gemma-2-9B-It</strong></td>
<td class="js-sort-string"><strong>FuseChat-Gemma-2-9B-SFT</strong></td>
<td class="js-sort-string"><strong>FuseChat-Gemma-2-9B-Instruct</strong></td>
</tr>
<tr>
<td style="white-space: nowrap;">AlpacaEval-2 (LC %)</td>
<td>51.1</td>
<td>49.8</td>
<td><strong>70.2</strong></td>
</tr>
<tr>
<td>Arena-Hard (WR %)</td>
<td>40.8</td>
<td>44.5</td>
<td><strong>63.4</strong></td>
</tr>
<tr>
<td>MT-Bench</td>
<td>8.5</td>
<td><strong>8.7</strong></td>
<td>8.6</td>
</tr>
<tr>
<td>AlignBench v1.1</td>
<td>7.0</td>
<td>7.1</td>
<td><strong>7.4</strong></td>
</tr>
<tr>
<td>GSM8K</td>
<td>88.5</td>
<td>90.5</td>
<td><strong>91.0</strong></td>
</tr>
<tr>
<td>MATH</td>
<td>49.6</td>
<td><strong>58.0</strong></td>
<td>57.8</td>
</tr>
<tr>
<td>AMC 23</td>
<td>20.0</td>
<td>27.5</td>
<td><strong>35.0</strong></td>
</tr>
<tr>
<td>LiveBench-0831</td>
<td>31.6</td>
<td><strong>33.3</strong></td>
<td>33.2</td>
</tr>
<tr>
<td>MMLU-Pro</td>
<td>50.5</td>
<td>52.5</td>
<td><strong>52.9</strong></td>
</tr>
<tr>
<td>MMLU-redux</td>
<td>72.8</td>
<td>72.8</td>
<td><strong>73.7</strong></td>
</tr>
<tr>
<td>GPQA-Diamond</td>
<td><strong>39.4</strong></td>
<td>33.3</td>
<td>35.4</td>
</tr>
<tr>
<td>HumanEval</td>
<td><strong>67.1</strong></td>
<td>65.9</td>
<td>64.0</td>
</tr>
<tr>
<td>MBPP</td>
<td><strong>75.1</strong></td>
<td>70.6</td>
<td>71.7</td>
</tr>
<tr>
<td>LiveCodeBench<br>2408-2411</td>
<td><strong>11.9</strong></td>
<td>11.0</td>
<td>10.1</td>
</tr>
<tr>
<td>Average</td>
<td>43.9</td>
<td>44.7</td>
<td><strong>48.2</strong></td>
</tr>
</table>
<h3 class="title is-4">FuseChat-Llama-3.2-3B-Instruct Performance</h3>
<table class="js-sort-table table hidden">
<tr>
<td class="js-sort-string"><strong>Benchmarks</strong></td>
<td class="js-sort-string"><strong>Llama-3.2-3B-Instruct</strong></td>
<td class="js-sort-string"><strong>FuseChat-Llama-3.2-3B-SFT</strong></td>
<td class="js-sort-string"><strong>FuseChat-Llama-3.2-3B-Instruct</strong></td>
</tr>
<tr>
<td style="white-space: nowrap;">AlpacaEval-2 (LC %)</td>
<td>21.4</td>
<td>31.1</td>
<td><strong>54.0</strong></td>
</tr>
<tr>
<td>Arena-Hard (WR %)</td>
<td>16.6</td>
<td>21.3</td>
<td><strong>30.2</strong></td>
</tr>
<tr>
<td>MT-Bench</td>
<td>6.9</td>
<td>7.3</td>
<td><strong>7.7</strong></td>
</tr>
<tr>
<td>AlignBench v1.1</td>
<td>3.8</td>
<td>5.5</td>
<td><strong>5.9</strong></td>
</tr>
<tr>
<td>GSM8K</td>
<td>82.0</td>
<td><strong>82.8</strong></td>
<td>82.0</td>
</tr>
<tr>
<td>MATH</td>
<td>51.4</td>
<td>52.9</td>
<td><strong>53.1</strong></td>
</tr>
<tr>
<td>AMC23</td>
<td>22.5</td>
<td>20.0</td>
<td><strong>35.0</strong></td>
</tr>
<tr>
<td>LiveBench 0831</td>
<td>23.4</td>
<td>24.5</td>
<td><strong>24.9</strong></td>
</tr>
<tr>
<td>MMLU-Pro</td>
<td>39.3</td>
<td><strong>40.3</strong></td>
<td>40.3</td>
</tr>
<tr>
<td>MMLU-redux</td>
<td>58.5</td>
<td>58.2</td>
<td><strong>59.0</strong></td>
</tr>
<tr>
<td>GPQA-Diamond</td>
<td>29.8</td>
<td>33.3</td>
<td><strong>33.8</strong></td>
</tr>
<tr>
<td>HumanEval</td>
<td>61.0</td>
<td><strong>62.8</strong></td>
<td>60.4</td>
</tr>
<tr>
<td>MBPP</td>
<td><strong>68.5</strong></td>
<td>67.5</td>
<td>67.5</td>
</tr>
<tr>
<td>LiveCodeBench<br>2408-2411</td>
<td>8.3</td>
<td>7.1</td>
<td><strong>9.0</strong></td>
</tr>
<tr>
<td>Average</td>
<td>35.2</td>
<td>36.8</td>
<td><strong>40.2</strong></td>
</tr>
</table>
<h3 class="title is-4">FuseChat-Llama-3.2-1B-Instruct Performance</h3>
<table class="js-sort-table table hidden">
<tr>
<td class="js-sort-string"><strong>Benchmarks</strong></td>
<td class="js-sort-string"><strong>Llama-3.2-1B-Instruct</strong></td>
<td class="js-sort-string"><strong>FuseChat-Llama-3.2-1B-SFT</strong></td>
<td class="js-sort-string"><strong>FuseChat-Llama-3.2-1B-Instruct</strong></td>
</tr>
<tr>
<td style="white-space: nowrap;">AlpacaEval-2 (LC %)</td>
<td>9.7</td>
<td>14.0</td>
<td><strong>25.3</strong></td>
</tr>
<tr>
<td>Arena-Hard (WR %)</td>
<td>5.1</td>
<td>6.0</td>
<td><strong>8.6</strong></td>
</tr>
<tr>
<td>MT-Bench</td>
<td>4.7</td>
<td>5.2</td>
<td><strong>5.7</strong></td>
</tr>
<tr>
<td>AlignBench v1.1</td>
<td>2.9</td>
<td>3.9</td>
<td><strong>4.3</strong></td>
</tr>
<tr>
<td>GSM8K</td>
<td>46.3</td>
<td><strong>55.6</strong></td>
<td>54.5</td>
</tr>
<tr>
<td>MATH</td>
<td>32.7</td>
<td><strong>34.7</strong></td>
<td>33.6</td>
</tr>
<tr>
<td>AMC23</td>
<td>17.5</td>
<td>15.0</td>
<td><strong>20.0</strong></td>
</tr>
<tr>
<td>LiveBench 0831</td>
<td>14.0</td>
<td>13.9</td>
<td><strong>15.8</strong></td>
</tr>
<tr>
<td>MMLU-Pro</td>
<td><strong>22.3</strong></td>
<td>21.5</td>
<td>21.3</td>
</tr>
<tr>
<td>MMLU</td>
<td><strong>45.8</strong></td>
<td>45.0</td>
<td>44.8</td>
</tr>
<tr>
<td>GPQA-Diamond</td>
<td>21.2</td>