Beyond Causal Language Modeling

NeurIPS 2024 Spotlight: Optimizing Language Model Pretraining with Selective Language Modeling (SLM)

Recently, I presented a fascinating paper from NeurIPS 2024, "Not All Tokens Are What You Need for Pretraining," at a local reading group. This paper tackles a surprisingly simple yet impactful question: Is next-token prediction necessary for every token during language model pretraining?

The standard approach involves massive web-scraped datasets and applying causal language modeling (CLM) universally. This paper challenges that assumption, proposing that some tokens hinder, rather than help, the learning process. The authors demonstrate that focusing training on "useful" tokens significantly improves data efficiency and downstream task performance. This post summarizes their core ideas and key experimental findings.

The Problem: Noise and Inefficient Learning

Large web corpora inevitably contain noise. While document-level filtering helps, noise often resides within individual documents. These noisy tokens waste computational resources and potentially confuse the model.

The authors analyzed token-level learning dynamics, categorizing tokens based on their cross-entropy loss trajectory:

• L→L (Low to Low): Quickly learned, providing minimal further benefit.
• H→L (High to Low): Initially difficult, but eventually learned; representing valuable learning opportunities.
• H→H (High to High): Consistently difficult, often due to inherent unpredictability (aleatoric uncertainty).
• L→H (Low to High): Initially learned, but later become problematic, possibly due to context shifts or noise.

Their analysis reveals that only a small fraction of tokens provide meaningful learning signals.

The Solution: Selective Language Modeling (SLM)

The proposed solution, Selective Language Modeling (SLM), offers a more targeted approach:

1. Reference Model (RM) Training: A high-quality subset of the data is used to fine-tune a pre-trained base model, creating a reference model (RM). This RM acts as a benchmark for token "usefulness."

2. Excess Loss Calculation: For each token in the large corpus, the difference between the RM's loss and the current training model's loss (the "excess loss") is calculated. Higher excess loss indicates greater potential for improvement.

3. Selective Backpropagation: The full forward pass is performed on all tokens, but backpropagation only occurs for the top k% of tokens with the highest excess loss. This dynamically focuses training on the most valuable tokens.

Experimental Results: Significant Gains

SLM demonstrates significant advantages across various experiments:

• Math Domain: On OpenWebMath, SLM achieved up to 10% performance gains on GSM8K and MATH benchmarks compared to standard CLM, reaching baseline performance 5-10 times faster. A 7B model matched a state-of-the-art model using only 3% of its training tokens. Fine-tuning further boosted performance by over 40% for a 1B model.

• General Domain: Even with a strong pre-trained base model, SLM yielded approximately 5.8% average improvement across 15 benchmarks, particularly in challenging domains like code and math.

• Self-Referencing: Even a quickly trained RM from the raw corpus provided a 2-3% accuracy boost and a 30-40% reduction in tokens used.

Conclusion and Future Work

This paper offers valuable insights into token-level learning dynamics and introduces SLM, a highly effective technique for optimizing language model pretraining. Future research directions include scaling SLM to larger models, exploring API-based reference models, integrating reinforcement learning, using multiple reference models, and aligning SLM with safety and truthfulness considerations. This work represents a significant advancement in efficient and effective language model training.

The above is the detailed content of Beyond Causal Language Modeling. For more information, please follow other related articles on the PHP Chinese website!