Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

OpenChat: Advancing Open-source Language Models with Mixed-Quality Data

About

Nowadays, open-source large language models like LLaMA have emerged. Recent developments have incorporated supervised fine-tuning (SFT) and reinforcement learning fine-tuning (RLFT) to align these models with human goals. However, SFT methods treat all training data with mixed quality equally, while RLFT methods require high-quality pairwise or ranking-based preference data. In this study, we present a novel framework, named OpenChat, to advance open-source language models with mixed-quality data. Specifically, we consider the general SFT training data, consisting of a small amount of expert data mixed with a large proportion of sub-optimal data, without any preference labels. We propose the C(onditioned)-RLFT, which regards different data sources as coarse-grained reward labels and learns a class-conditioned policy to leverage complementary data quality information. Interestingly, the optimal policy in C-RLFT can be easily solved through single-stage, RL-free supervised learning, which is lightweight and avoids costly human preference labeling. Through extensive experiments on three standard benchmarks, our openchat-13b fine-tuned with C-RLFT achieves the highest average performance among all 13b open-source language models. Moreover, we use AGIEval to validate the model generalization performance, in which only openchat-13b surpasses the base model. Finally, we conduct a series of analyses to shed light on the effectiveness and robustness of OpenChat. Our code, data, and models are publicly available at https://github.com/imoneoi/openchat and https://huggingface.co/openchat.

Guan Wang, Sijie Cheng, Xianyuan Zhan, Xiangang Li, Sen Song, Yang Liu• 2023

Related benchmarks

TaskDatasetResultRank
Jailbreak AttackHarmBench--
624
SummarizationXsum--
108
General Language IntelligenceMMLU, GSM8K, BBH, TriviaQA, NQ latest available (test)
MMLU63.87
26
Multi-Robot Task AllocationLEGO-MRTA (test)
BLEU-422
18
Prompt Injection Attack EvaluationAdvBench
ASP94
14
Prompt Injection Attack EvaluationJailbreakBench
Attack Success Probability (ASP)92
14
Prompt Injection Attack EvaluationWalledEval
ASP78.7
14
Prompt Injection Attack EvaluationSAP10
Attack Success Percentage (ASP)96.3
14
Prompt Injection Attack EvaluationHarmBench
ASP86.1
14
Instruction FollowingAlpacaEval, MT-bench, Vicuna-bench
AlpacaEval Score89.5
13
Showing 10 of 16 rows

Other info

Code

Follow for update