InteRank: Reasoning-Intensive Document Ranking with Small Language Models

InteRank is a novel approach for training compact language models (< 3B parameters) to perform reasoning-intensive document ranking with performance comparable to models over 20x larger. Our methodology combines knowledge distillation from a large teacher model with reinforcement learning optimization to create efficient yet powerful ranking models that can explain their decisions.

Overview

Key features:

• Achieves state-of-the-art performance on the BRIGHT benchmark using only a 3B parameter model
• Generates natural language explanations to justify ranking decisions
• Uses a two-stage training approach combining knowledge distillation and reinforcement learning
• Requires no human annotations for training
• Supports both query generation and document relevance assessment

Training Process

1. Synthetic Data Generation: Automatically generates training data from StackExchange question-answer pairs using a large teacher model (Llama 3.3 70B)

2. Knowledge Distillation: Transfers initial reasoning capabilities from the teacher to a compact student model (Llama 3.2 3B)

3. Reinforcement Learning: Refines reasoning capabilities by rewarding high-quality explanations and accurate relevance predictions

Model Architecture

• Base Model: Llama 3.2 3B
• Training: QLoRA with 4-bit quantization and rank-64 adapters
• Context Length: 4K tokens
• Hardware Requirements: Single A100 GPU

Results

Our 3B parameter model achieves:

• 27.4% average nDCG@10 across all domains on BRIGHT benchmark
• 3rd place on BRIGHT leaderboard
• Outperforms recent approaches like Reason-to-Rank (nDCG@5 26.2 vs 19.6)
• Performance comparable to ensemble models using 70B+ parameters