VitalContribution/Evangelion-7B

TEXT GENERATIONConcurrency Cost:1Model Size:7BQuant:FP8Ctx Length:8kPublished:Jan 11, 2024License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

VitalContribution/Evangelion-7B is a 7 billion parameter language model developed by VitalContribution, built upon the Weyaxi/OpenHermes-2.5-neural-chat-v3-3-Slerp architecture. This model was created by merging DPO-optimized and non-DPO-optimized models and fine-tuned on a high-quality DPO dataset, resulting in strong performance across various reasoning and language understanding benchmarks. It is particularly suited for general conversational AI and tasks requiring nuanced response generation.

Loading preview...

Overview

VitalContribution/Evangelion-7B is a 7 billion parameter language model that explores the impact of combining DPO-optimized and non-DPO-optimized models, further enhanced by fine-tuning on a high-quality DPO dataset. The base model for this experiment was Weyaxi/OpenHermes-2.5-neural-chat-v3-3-Slerp.

Key Capabilities

  • Enhanced Reasoning: Achieves 68.94 on AI2 Reasoning Challenge (25-Shot) and 66.94 on GSM8k (5-shot).
  • Language Understanding: Demonstrates strong performance with 86.45 on HellaSwag (10-Shot) and 79.95 on Winogrande (5-shot).
  • Instruction Following: Fine-tuned using a carefully filtered, high-quality DPO dataset (/argilla/distilabel-intel-orca-dpo-pairs) to improve response quality.
  • ChatML Format: Utilizes the ChatML template, integrated into the model's tokenizer, for structured conversational interactions.

Performance Highlights

Evaluated on the Open LLM Leaderboard, Evangelion-7B achieved an average score of 71.71. Notable scores include:

  • MMLU (5-Shot): 63.97
  • TruthfulQA (0-shot): 64.01

Good For

  • General-purpose conversational AI applications.
  • Tasks requiring robust reasoning and language comprehension.
  • Developers looking for a 7B model with strong instruction-following capabilities derived from DPO fine-tuning.

Popular Sampler Settings

Top 3 parameter combinations used by Featherless users for this model. Click a tab to see each config.

temperature
top_p
top_k
frequency_penalty
presence_penalty
repetition_penalty
min_p