kyujinpy/PlatYi-34B-Q

TEXT GENERATIONConcurrency Cost:2Model Size:34BQuant:FP8Ctx Length:32kPublished:Dec 1, 2023License:cc-by-nc-sa-4.0Architecture:Transformer Open Weights Cold

PlatYi-34B-Q is a 34 billion parameter auto-regressive language model developed by Kyujin Han (kyujinpy), based on the Yi-34B transformer architecture. This model is fine-tuned using QLoRA on the Open-Platypus dataset, demonstrating improved performance over its base model across various benchmarks including MMLU and GSM8K. It is designed for general text generation tasks, offering enhanced reasoning and problem-solving capabilities.

Loading preview...

Model Overview

PlatYi-34B-Q is a 34 billion parameter auto-regressive language model developed by Kyujin Han (kyujinpy). It is built upon the robust Yi-34B transformer architecture and has been fine-tuned using QLoRA on the Open-Platypus dataset.

Key Capabilities & Performance

This model demonstrates notable improvements over its base model, 01-ai/Yi-34B, across several benchmarks. On the Open LLM Leaderboard, PlatYi-34B-Q achieves an average score of 69.86, surpassing the base Yi-34B's 69.42. Specific benchmark scores include:

  • MMLU (5-Shot): 77.66 (vs. 76.35 for base Yi-34B)
  • GSM8K (5-Shot): 53.98 (vs. 50.64 for base Yi-34B)
  • ARC (25-Shot): 66.89
  • HellaSwag (10-Shot): 85.14
  • TruthfulQA (0-shot): 53.03
  • Winogrande (5-shot): 82.48

These results indicate enhanced reasoning and problem-solving abilities, particularly in multi-task language understanding and mathematical reasoning.

Use Cases

PlatYi-34B-Q is suitable for a variety of text generation tasks where improved general intelligence and benchmark performance are beneficial. Its fine-tuning on the Open-Platypus dataset suggests applicability in areas requiring strong instruction following and factual recall. Developers can integrate it using the Hugging Face transformers library with torch.float16 for efficient inference.