😎 Who knew you could create an amazing LLM that outperforms all other open-source models with just 20,000 human-labeled training samples?
💡 The coolest thing about Nvidia's Nematron is that it became the state-of-the-art open-source LLM using 98% synthetic data and only 20k human-labeled samples! Nematron not only beat all other open-source LLMs on standard benchmarks but also topped the LLMsys leaderboard.
What's special about their synthetic data generation?
⛳ Prompt Generation
They generate a range of prompts that span various tasks, topics, and instructions using Mixtral-8x7B-Instruct-v0.1
👉 Synthetic Single-Turn Prompts
👉 Synthetic Instruction-Following Prompts
👉 Synthetic Two-Turn Prompts
👉 Real-World LMSYS-Chat-1M Prompts
⛳ Synthetic Dialogue Generation
For synthetic dialogue generation, supervised fine-tuning is used to enable models to learn interaction in a dialogue format. Each conversation comprises three turns for a dynamic and interactive flow. The quality of these dialogues is controlled using the Nemotron-4-340B-Reward model, which filters out lower-quality samples.
⛳ Synthetic Preference Data Generation
👉 The synthetic data includes single-turn, instruction-following, and two-turn prompts, as well as real-world prompts from various datasets.
👉 Responses are generated using multiple random intermediate models to ensure diversity.
👉 For judging preference, they use ground-truth labels when available. Otherwise, we employ two methods: LLM-as-Judge, where a language model compares responses, and Reward-Model-as-Judge, where the reward model predicts the reward for each response.
⛳ Iterative Weak-to-Strong Alignment
👉An iterative approach is developed that combines alignment training and data synthesis to refine data towards optimality, enhancing each other and driving continuous improvement.
👉 The approach involves starting with an initial aligned model as the generator, aligning a better base model using supervised fine-tuning and preference tuning, and achieving significant improvement in model performance through iterations.
#genaish #llms