Researchers have developed a new method for training Natural Language Inference (NLI) models using Group Relative Policy Optimization (GRPO), a reinforcement learning approach. This technique eliminates the need for human-labeled rationales, allowing models to be trained on challenging datasets like ANLI. When applied to 7B, 14B, and 32B language models using parameter-efficient methods such as LoRA and QLoRA, the GRPO-trained models demonstrated strong performance on standard and adversarial NLI benchmarks. Notably, the 32B model showed superior generalization on adversarial sets compared to supervised baselines, and with AWQ quantization, it fits within 22GB of CUDA memory. AI
IMPACT This research offers a more scalable and efficient way to train robust NLI models, potentially improving applications like fact-checking and information retrieval.
RANK_REASON The cluster contains a research paper detailing a new training methodology for NLI models. [lever_c_demoted from research: ic=1 ai=1.0]
- arXiv
- Chain-of-Thought
- CUDA
- Group Relative Policy Optimization
- GRPO
- Hugging Face
- LoRA
- QLoRA
- Natural Language Inference
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
