WestSeverus-7B-DPO-v2 Overview

WestSeverus-7B-DPO-v2 is a 7 billion parameter model, part of the WestLake Family, developed by PetroGPT. It is an instruction-tuned model, specifically trained using Direct Preference Optimization (DPO) on various datasets, building upon the base model WestSeverus-7B. The model is noted for its 8192 token context length.

Key Capabilities & Performance

• Mathematical Proficiency: Demonstrates good performance on basic math problems.
• Benchmark Excellence: Achieves top scores on the YALL - Yet Another LLM Leaderboard, particularly outperforming on TruthfulQA (72.72) and BigBench (48.71) among comparable 7B models.
• Open LLM Leaderboard: Ranks highly among 7B models, with strong results in TruthfulQA (72.37) and GSM8K (71.65).
• Coding Potential: While HumanEval scores are 43.3, it is suggested for further research and reference in coding.

Use Cases

This model is suitable for applications requiring:

• Mathematical Problem Solving: Especially basic arithmetic and logical reasoning.
• Scientific Research: Applicable in chemistry, physics, and related fields for reference and analysis.
• Coding Assistance: Can be utilized for research and reference in programming tasks.

WestSeverus-7B-DPO-v2 uses the ChatML prompt format with system prompts, facilitating structured interactions.