Alibaba-NLP/ERank-14B
ERank-14B is a 14 billion parameter pointwise reranker developed by Alibaba-NLP, designed for effective and efficient text reranking with a 32K context length. It utilizes a novel two-stage training pipeline involving Supervised Fine-Tuning (SFT) for generative integer score output and Reinforcement Learning (RL) with a listwise derived reward. This model excels in diverse relevance scenarios, including challenging reasoning-intensive tasks, while maintaining low latency compared to listwise rerankers.
Loading preview...
ERank-14B: An Efficient and Effective Text Reranker
ERank-14B is a 14 billion parameter pointwise reranker developed by Alibaba-NLP, designed to provide highly effective and efficient text reranking across various relevance scenarios. It stands out by outperforming many recent listwise rerankers on complex, reasoning-intensive tasks while maintaining low latency.
Key Capabilities and Innovations
- Novel Two-Stage Training: ERank is trained using a unique pipeline combining Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL). The SFT stage trains the model to generatively output fine-grained integer scores, departing from traditional binary relevance classification.
- Listwise Awareness in Pointwise Architecture: The RL stage introduces a novel listwise derived reward, which instills global ranking awareness into the efficient pointwise architecture, enhancing its reranking capabilities.
- Instruction Awareness: The model supports customizing input instructions, allowing for flexible adaptation to different reranking tasks.
- Strong Performance: ERank-14B demonstrates competitive performance across various benchmarks, including BRIGHT, FollowIR (reasoning-intensive), BEIR, and TREC DL (semantic relevance). It achieves strong results, particularly on the BRIGHT benchmark, with top-100 documents retrieved by ReasonIR-8B using GPT-4 reason-query, and further improved with BM25 hybrid scores.
- Low Latency: As a pointwise reranker, ERank-14B offers significantly lower latency compared to listwise models, making it suitable for applications requiring fast response times.
When to Use ERank-14B
- Text Reranking: Ideal for improving the relevance of retrieved documents or passages in search, question answering, and recommendation systems.
- Reasoning-Intensive Tasks: Particularly effective for scenarios requiring deep understanding and reasoning to rank documents, as evidenced by its performance on BRIGHT and FollowIR benchmarks.
- Low-Latency Applications: Suitable for use cases where quick reranking is crucial, thanks to its efficient pointwise architecture.
- Customizable Reranking: Its instruction-aware design allows developers to tailor the reranking behavior to specific task requirements.