Conference Talk 9: Why Fine-Tuning is Dead

In this talk, Emmanuel Ameisen from Anthropic argues that fine-tuning LLMs is often less effective and efficient than focusing on fundamentals like data quality, prompting, and Retrieval Augmentation Generation (RAG).

This post is part of the following series:

• Mastering LLMs Course Notes: My notes from the course Mastering LLMs: A Conference For Developers & Data Scientists by Hamel Husain and Dan Becker.

• Trends in Machine Learning
• Performance Observations: Fine-tuning vs. RAG
• The Moving Target: Fine-tuning and Frontier Models
• The Difficulty of Fine-tuning: Prioritizing Fundamentals
• Extrapolating Trends: Context Size, Price, and Latency
• Conclusion
• Q&A Session

Trends in Machine Learning

• Focus on fundamentals: Throughout ML history, focusing on “boring” fundamentals like data quality and SQL queries often yielded better results than chasing the latest “cool” technology.
  • 2009: Data analysis and SQL were more impactful than training ML models.
  • 2012-2014: XGBoost outperformed deep learning for many tasks.
  • 2015: Data cleaning and error correction were more effective than inventing new loss functions.
  • 2023: Prompting and RAG often outperform fine-tuning.
• Fine-tuning’s uncertain future: While fine-tuning became popular with the rise of pre-trained models like ResNet and BERT, LLMs might be shifting the paradigm again towards prompt-based and RAG-augmented approaches.

Q&A on Trends and Embedding Models

• Fine-tuning’s future relevance: Fine-tuning might still become more valuable in the future, just like deep learning did after an initial period of limited practical use.
• Embedding models: While LLMs are currently the focus of improvement, embedding models might also see advancements that reduce the need for fine-tuning. However, domain-specific applications might still require fine-tuning or hybrid approaches like combining keyword search with embedding search.
• Domain-specific ranking and retrieval: RAG, combined with techniques like query expansion driven by LLMs, might be able to address the need for domain-specific ranking and retrieval without fine-tuning embedding models.
• Benchmarking prompting vs. fine-tuning: Limited data exists comparing prompting and RAG to fine-tuning, but existing research suggests RAG often outperforms fine-tuning.

Performance Observations: Fine-tuning vs. RAG

• RAG often outperforms fine-tuning: Several studies indicate that RAG, even without fine-tuning, often achieves comparable or better performance than fine-tuning alone, especially for larger models and knowledge-based tasks.
  • Forum Post: Fine-tuning vs Context-Injection (RAG)
• Prioritization over “versus”: While combining RAG and fine-tuning can yield incremental improvements, prioritizing RAG is crucial due to its efficiency and effectiveness.
  • Paper: Fine Tuning vs. Retrieval Augmented Generation for Less Popular Knowledge
  • Paper: Fine-Tuning or Retrieval? Comparing Knowledge Injection in LLMs
• Fine-tuning for style, not knowledge: Fine-tuning might be less suitable for incorporating domain knowledge into a model compared to RAG, which directly provides relevant context.
  • Paper: RAG vs Fine-tuning: Pipelines, Tradeoffs, and a Case Study on Agriculture
• The evolving definition of “knowledge”: Distinguishing between “knowledge” and “style” in LLMs is complex and changes with each model generation. What previously required fine-tuning (e.g., writing style) might be achievable through prompting in newer models.

Q&A on Fine-tuning, RAG, and Knowledge

• Fine-tuning effectiveness and model size: Fine-tuning might be more beneficial for smaller models compared to larger ones, which are already more capable of learning from context.
• Domain-specific knowledge and RAG: Even for smaller models, RAG remains crucial for tasks involving domain-specific knowledge.
• Evaluating fine-tuning success: The choice of evaluation metric significantly impacts the perceived effectiveness of fine-tuning. For tasks like style adherence, fine-tuning might appear more beneficial than RAG.
• The blurry line between style and content: The distinction between “style” and “content” can be ambiguous, making it difficult to definitively determine when fine-tuning is beneficial.

Audience Questions and Examples of Fine-tuning Success

• Complex knowledge bases and fine-tuning: When dealing with large, curated knowledge bases, it’s crucial to evaluate whether the next generation of LLMs, combined with RAG, might be sufficient without fine-tuning.
• Adding knowledge via prompting and RAG: In many cases, adding knowledge to the model can be achieved through prompting, RAG, or a combination of both, eliminating the need for fine-tuning.
• Fine-tuning for multilingual models: Fine-tuning might be beneficial for improving the performance of multilingual models on languages with limited training data, as it leverages the model’s existing understanding of language mapping.
• Fine-tuning for code generation: While fine-tuning can be used to adapt code generation models to specific styles and conventions, RAG remains highly effective for providing codebase context.
• Contextual learning vs. fine-tuning: LLMs are demonstrating impressive abilities to learn from context, potentially replacing the need for fine-tuning in scenarios where sufficient context can be provided.

The Moving Target: Fine-tuning and Frontier Models

• Rapid LLM advancements: The rapid pace of LLM development makes fine-tuning a moving target, as newer models often surpass the performance of previously fine-tuned models.
• Bloomberg GPT example: Bloomberg GPT, a large language model pre-trained on financial data, initially outperformed existing models on financial tasks. However, its performance was subsequently surpassed by newer models like GPT-4.
  • Paper: BloombergGPT: A Large Language Model for Finance
  • Paper: Are ChatGPT and GPT-4 General-Purpose Solvers for Financial Text Analytics? A Study on Several Typical Tasks
• The cost of keeping up: Continuously fine-tuning on new model releases can be prohibitively expensive, especially for large datasets. Prompt-based and RAG-based pipelines offer more flexibility and cost-effectiveness.
• Fine-tuning effectiveness and model scale: Fine-tuning might become less effective as models grow larger and more capable.

The Difficulty of Fine-tuning: Prioritizing Fundamentals

• The 80/20 rule of ML: Similar to traditional ML, most effort in LLM development should be dedicated to data work (80%), followed by engineering (18%), debugging (2%), and architecture research (0%).
• Fine-tuning as a last resort: Fine-tuning should only be considered after thoroughly addressing fundamentals like data quality, evaluation, prompting, and RAG.
• Hierarchy of needs: Prioritize building a solid ML system with robust evaluation, prompting, and RAG before attempting to fine-tune.
  • Book: Building Machine Learning Powered Applications: Going from Idea to Product
  • Continuous Integration
    • Model Backtesting
    • Model Evaluation
    • Experimentation Framework
  • Application Logic
    • Input Validation → Filtering Logic → Model Code → Output Validation → Displaying Logic
  • Monitoring
    • Monitoring Input Distribution
    • Monitoring Latency
    • Monitoring Output Distribution

Extrapolating Trends: Context Size, Price, and Latency

• Decreasing costs and increasing capabilities: LLM costs are rapidly decreasing, while context windows and processing speeds are increasing.
• The impact of future trends: If these trends continue, providing sufficient context to highly capable LLMs might become more efficient than fine-tuning for many use cases.
• Context window limitations: Fine-tuning might still be necessary for applications requiring context exceeding the limits of available LLMs. However, techniques like prefix caching could mitigate this need.

Conclusion

• Finetuning is:
  • expensive and complex
  • has become less valuable
  • often underperforms simpler approaches
• Models are continuously becoming:
  • cheaper
  • smarter
  • faster
  • longer context
• Always start with:
  • prompting
  • making a train/test set
  • RAG
• Treat finetuning as a niche/last resort solution
  • like cloud vs on prem

Q&A Session

Question #1

• Context window size and computational cost: Passing large amounts of context through an LLM for every request can be computationally expensive. However, increasing LLM efficiency and advancements like prefix caching could mitigate this cost.
• Fine-tuning complexity: Fine-tuning increasingly complex and larger LLMs might become more challenging, potentially outweighing the benefits compared to context-based approaches.

Question #2

• Dynamic few-shot learning: Dynamically selecting and providing relevant few-shot examples from a database is a powerful technique for improving LLM performance without fine-tuning.
• Iterative prompt and example improvement: Invest time in iteratively refining prompts and curating effective few-shot examples before considering fine-tuning.

---

About Me:

I’m Christian Mills, a deep learning consultant specializing in practical AI implementations. I help clients leverage cutting-edge AI technologies to solve real-world problems.

Interested in working together? Fill out my Quick AI Project Assessment form or learn more about me.