NLPinas/yi-bagel-2x34b

TEXT GENERATIONConcurrency Cost:2Model Size:34BQuant:FP8Ctx Length:32kPublished:Jan 11, 2024License:yi-licenseArchitecture:Transformer0.0K Cold

NLPinas/yi-bagel-2x34b is a 34 billion parameter merged language model created by NLPinas, combining jondurbin/bagel-dpo-34b-v0.2 and jondurbin/nontoxic-bagel-34b-v0.2 using a linear merge method. This model debuted at rank #4 on the Open LLM Leaderboard (January 11, 2024), demonstrating strong performance across various reasoning and commonsense benchmarks. It is designed to assess the impact of DPO on censoring and offers a balanced approach to general language understanding and generation.

Loading preview...

Model Overview

NLPinas/yi-bagel-2x34b is a 34 billion parameter language model developed by NLPinas, created through an experimental linear merge of two distinct models: jondurbin/bagel-dpo-34b-v0.2 and jondurbin/nontoxic-bagel-34b-v0.2. This merge aims to evaluate the influence of DPO (Direct Preference Optimization) on model behavior, particularly concerning censoring.

Key Capabilities & Performance

This model achieved a notable rank #4 on the Hugging Face Open LLM Leaderboard as of January 11, 2024, showcasing robust performance across a suite of benchmarks:

  • MMLU (5-shot): 76.60
  • ARC (25-shot): 72.70
  • HellaSwag (10-shot): 85.44
  • TruthfulQA (0-shot): 71.42
  • Winogrande (5-shot): 82.72
  • GSM8K (5-shot): 60.73

These metrics indicate strong abilities in multitask accuracy, scientific reasoning, commonsense inference, truthfulness, and mathematical problem-solving.

Unique Characteristics

  • Merged Architecture: Utilizes a linear merge of two specialized models, offering a unique blend of their respective strengths.
  • DPO Impact Assessment: Specifically designed to explore the effects of DPO on model outputs, particularly regarding content moderation and bias.

Use Cases

Given its balanced performance and experimental nature, yi-bagel-2x34b is suitable for:

  • General-purpose language generation and understanding tasks.
  • Research into model merging techniques and the impact of DPO.
  • Applications requiring a model with strong reasoning and commonsense capabilities.