Model Card for Mistral-Crab-DPO

Large language models (LLMs) struggle to follow instructions with complex constraints in format, length, etc. Following the conventional instruction-tuning practice, previous works conduct post-training on complex instruction-response pairs generated by feeding complex instructions to advanced LLMs. However, even advanced LLMs cannot follow complex instructions well, thus limiting the quality of generated data. In this work, we find that existing datasets inherently contain implicit complex constraints and propose a novel data generation technique, constraint back-translation. Specifically, we take the high-quality instruction-response pairs in existing datasets and only adopt advanced LLMs to add complex constraints already met by the responses to the instructions, which naturally reduces costs and data noise. In the experiments, we adopt Llama3-70B-Instruct to back-translate constraints and create a high-quality complex instruction-response dataset, named CRAB. We present that post-training on CRAB improves multiple backbone LLMs' complex instruction-following ability, evaluated on extensive instruction-following benchmarks. We further find that constraint back-translation also serves as a useful auxiliary training objective in post-training.

• 📖 Paper: Constraint Back-translation Improves Complex Instruction Following of Large Language Models

• 🦀 Github: THU/Crab

Model Performance

Models	BaseModel	IFEval	FollowBench(HSR)			AVG
		AVG	L1-L2	L3-L5	AVG
GPT-3.5-turbo	GPT	66.3	74.2	61	66.2	66.3
GPT-4	GPT	81.3	80.4	69.4	73.8	77.6
Vicuna-13b-V1.5	Llama2	50.3	66.3	39.8	50.4	50.4
WizardLM-13B-V1.2	Llama2	51.4	56.5	36.9	44.7	48
Conifer-13B	Llama2	50.2	57.1	40.3	47	48.6
Zephyr-7B-beta	Mistral	45.4	54.8	38.2	44.8	45.1
Conifer-7B	Mistral	53.9	51.9	40.2	44.9	49.4
Conifer-7B-DPO	Mistral	55.7	57	45.4	50	52.9
Llama3 8B	Llama3	31.4	6.8	8.2	7.6	19.5
Llama3-crab	Llama3	46.9	51.2	26.7	36.5	41.7
Llama3-crab + DPO	Llama3	49.7	56.8	38.1	45.5	47.6
Mistral 7B	Mistral	25.2	15.5	6.5	10.1	17.7
Mistral-crab	Mistral	54.5	59.2	32.8	43.3	48.9
Mistral-crab + DPO	Mistral	59.4	59.9	42.5	49.4	54.4

Model Description

• Developed by: Yunjia Qi, Hao Peng, Xiaozhi Wang, Bin Xu, Lei Hou, Juanzi Li
• Model type: Text Generation
• Language(s) (NLP): English
• Finetuned from model [optional]: Mistral-7B-v0.3