InstructionGPT-4: A 200-Instruction Paradigm for Fine-Tuning MiniGPT-4 (2308.12067v2)

Abstract: Multimodal LLMs are typically trained in two stages: first pre-training on image-text pairs, and then fine-tuning using supervised vision-language instruction data. Recent studies have shown that LLMs can achieve satisfactory results even with a limited amount of high-quality instruction-following data. In this paper, we introduce InstructionGPT-4, which is fine-tuned on a small dataset comprising only 200 examples, amounting to approximately 6\% of the instruction-following data used in the alignment dataset for MiniGPT-4. To achieve this, we first propose several metrics to access the quality of multimodal instruction data. Based on these metrics, we present an effective and trainable data selector to automatically identify and filter low-quality vision-language data. By employing this method, InstructionGPT-4 outperforms the original MiniGPT-4 on various evaluations. Overall, our findings demonstrate that less but high-quality instruction tuning data is efficient in enabling multimodal LLMs to generate better output. Our code is available at https://github.com/waltonfuture/InstructionGPT-4.

• The paper demonstrates that a limited set of 200 high-quality instruction examples can outperform full training in MiniGPT-4 using strategic data curation.
• The methodology employs novel metrics like CLIP, GPT, Reward, and Length Scores, along with multimodal features and a self-attention network for automatic data selection.
• Empirical results show significant gains across benchmarks, with +23 on MME, +1.55 on MMBench, and a +1.76% boost on VQA datasets compared to MiniGPT-4.

An Examination of InstructionGPT-4: Enhancing Multimodal Models Through Strategic Data Selection

The paper presents an investigation into the potential of strategically curated data in enhancing the performance of multimodal LLMs, particularly through the introduction of InstructionGPT-4. This model is a variant of MiniGPT-4, fine-tuned meticulously with only a small subset of high-quality instruction-following data, amounting to 200 examples, or roughly 6% of the initial data used for MiniGPT-4's alignment.

Data Selection and Methodology

A core element of this paper is the proposal of a robust, trainable data selector designed to identify and filter low-quality vision-language data efficiently. The data selection process revolves around several novel metrics tailored to assess the quality of multimodal instruction data. These include the CLIP Score, GPT Score, Reward Score, Length Score, and Multimodal Features, each offering a distinct perspective on the data's potential utility for fine-tuning.

Central to the paper is the principle that less but high-quality data can yield superior model performance. This aligns with findings from studies such as LIMA, which advocate for a data selection approach that prioritizes quality over quantity. Interestingly, the paper ventures beyond mere data accumulation to explore automatic data selection using a self-attention network, trained to map proposed quality metrics to actual task performance in a validation set.

Empirical Results

The practical efficacy of InstructionGPT-4 is evidenced across numerous benchmarks, including MME, MMBench, and various VQA datasets. Notably, InstructionGPT-4 outperformed MiniGPT-4 across all tested metrics. It achieved a +23 score improvement on MME, +1.55 on MMBench, and a +1.76% boost on VQA datasets over MiniGPT-4.

A critical revelation from these results is the role of data quality in enabling more efficient and effective model fine-tuning. The paper concludes that 200 handpicked data points, derived from their proposed selection mechanism, can suffice to exceed MiniGPT-4’s benchmark performance. This insight could significantly impact future approaches to fine-tuning, providing a more efficient framework for resource utilization in training multimodal models.

Implications and Speculation on Future Directions

The implications of this research are manifold. Practically, it suggests that institutions with limited datasets can leverage high-quality selection methods to achieve competitive results with fewer resources. Theoretically, it prompts a re-evaluation of the relationship between data scaling laws and model performance, reinforcing the notion that strategic data curation can compete with raw data volume in certain contexts.

Looking ahead, the framework presents an opportunity to explore multimodal instruction mining more broadly. This could involve refining selection metrics or expanding the model applicability to other architectures beyond MiniGPT-4. Additionally, future research might investigate further dimensions of data quality, encompassing syntactic and semantic factors, leveraging more advanced evaluation models, or even integrating human preferences more directly into machine-guided data selection processes.

In conclusion, the introduction of InstructionGPT-4 underscores a shift towards more sophisticated data curation techniques in AI, advocating for the prioritization of quality over mere quantity in instruction datasets. The insights derived from this work could pave the way for new paradigms in multimodal model training, potentially catalyzing advancements in artificial general intelligence by fostering more robust, adaptable, and efficient learning frameworks.