NVIDIA, in collaboration with Mistral AI, has announced the release of the Mistral-NeMo-Minitron 8B model, a highly advanced open-access large language model (LLM). According to the NVIDIA Technical Blog, this model surpasses other models of a similar size in terms of accuracy on nine popular benchmarks.
Advanced Model Pruning and Distillation
The Mistral-NeMo-Minitron 8B model was developed by width-pruning the larger Mistral NeMo 12B model, followed by a light retraining process using knowledge distillation. This methodology, originally proposed by NVIDIA in their paper on Compact Language Models via Pruning and Knowledge Distillation, has been validated through multiple successful implementations, including the NVIDIA Minitron 8B and 4B models, as well as the Llama-3.1-Minitron 4B model.
Model pruning involves reducing the size and complexity of a model by either dropping layers (depth pruning) or neurons and attention heads (width pruning). This process is often paired with retraining to recover any lost accuracy. Model distillation, on the other hand, transfers knowledge from a large, complex model (the teacher model) to a smaller, simpler model (the student model), aiming to retain much of the predictive power of the original model while being more efficient.
The combination of pruning and distillation allows for the creation of progressively smaller models from a large pretrained model. This approach significantly reduces the computational cost, as only 100-400 billion tokens are needed for retraining, compared to the much larger datasets required for training from scratch.
Mistral-NeMo-Minitron 8B Performance
The Mistral-NeMo-Minitron 8B model demonstrates leading accuracy on several benchmarks, outperforming other models in its class, including the Llama 3.1 8B and Gemma 7B models. The table below highlights the performance metrics:
Training tokens | Wino-Grande 5-shot | ARC Challenge 25-shot | MMLU 5-shot | Hella Swag 10-shot | GSM8K 5-shot | TruthfulQA 0-shot | XLSum en (20%) 3-shot | MBPP 0-shot | Human Eval 0-shot | ||
Llama 3.1 8B | 15T | 77.27 | 57.94 | 65.28 | 81.80 | 48.60 | 45.06 | 30.05 | 42.27 | 24.76 | |
Gemma 7B | 6T | 78 | 61 | 64 | 82 | 50 | 45 | 17 | 39 | 32 | |
Mistral-NeMo-Minitron 8B | 380B | 80.35 | 64.42 | 69.51 | 83.03 | 58.45 | 47.56 | 31.94 | 43.77 | 36.22 | |
Mistral NeMo 12B | N/A | 82.24 | 65.10 | 68.99 | 85.16 | 56.41 | 49.79 | 33.43 | 42.63 | 23.78 |
Implementation and Future Work
Following the best practices of structured weight pruning and knowledge distillation, the Mistral-NeMo 12B model was width-pruned to yield the 8B target model. The process involved fine-tuning the unpruned Mistral NeMo 12B model using 127 billion tokens to correct for distribution shifts, followed by width-only pruning and distillation using 380 billion tokens.
The Mistral-NeMo-Minitron 8B model showcases superior performance and efficiency, making it a significant advancement in the field of AI. NVIDIA plans to continue refining the distillation process to produce even smaller and more accurate models. The implementation of this technique will be gradually integrated into the NVIDIA NeMo framework for generative AI.
For further details, visit the NVIDIA Technical Blog.
Image source: Shutterstock
Credit: Source link