Dear friends,
Inexpensive token generation and agentic workflows for large language models (LLMs) open up intriguing new possibilities for training LLMs on synthetic data. Pretraining an LLM on its own directly generated responses to prompts doesn't help. But if an agentic workflow implemented with the LLM results in higher quality output than the LLM can generate directly, then training on that output becomes potentially useful.
Just as humans can learn from their own thinking, perhaps LLMs can, too. For example, imagine a math student who is learning to write mathematical proofs. By solving a few problems — even without external input — they can reflect on what does and doesn’t work and, through practice, learn how to more quickly generate good proofs.
Broadly, LLM training involves (i) pretraining (learning from unlabeled text data to predict the next word) followed by (ii) instruction fine-tuning (learning to follow instructions) and (iii) RLHF/DPO tuning to align the LLM’s output to human values. Step (i) requires many orders of magnitude more data than the other steps. For example, Llama 3 was pretrained on over 15 trillion tokens, and LLM developers are still hungry for more data. Where can we get more text to train on?
Many developers train smaller models directly on the output of larger models, so a smaller model learns to mimic a larger model’s behavior on a particular task. However, an LLM can’t learn much by training on data it generated directly, just like a supervised learning algorithm can’t learn from trying to predict labels it generated by itself. Indeed, training a model repeatedly on the output of an earlier version of itself can result in model collapse.
However, an LLM wrapped in an agentic workflow may produce higher-quality output than it can generate directly. In this case, the LLM’s higher-quality output might be useful as pretraining data for the LLM itself.
Efforts like these have precedents:
- When using reinforcement learning to play a game like chess, a model might learn a function that evaluates board positions. If we apply game tree search along with a low-accuracy evaluation function, the model can come up with more accurate evaluations. Then we can train that evaluation function to mimic these more accurate values.
- In the alignment step, Anthropic’s constitutional AI method uses RLAIF (RL from AI Feedback) to judge the quality of LLM outputs, substituting feedback generated by an AI model for human feedback.
A significant barrier to using LLMs prompted via agentic workflows to produce their own training data is the cost of generating tokens. Say we want to generate 1 trillion tokens to extend a pre-existing training dataset. Currently, at publicly announced prices, generating 1 trillion tokens using GPT-4-turbo ($30 per million output tokens), Claude 3 Opus ($75), Gemini 1.5 Pro ($21), and Llama-3-70B on Groq ($0.79) would cost, respectively, $30M, $75M, $21M and $790K. Of course, an agentic workflow that uses a design pattern like Reflection would require generating more than one token per token that we would use as training data. But budgets for training cutting-edge LLMs easily surpass $100M, so spending a few million dollars more for data to boost performance is quite feasible.
That’s why I believe agentic workflows will open up intriguing new opportunities for high-quality synthetic data generation.
Keep learning!
Andrew