Model Overview

macadeliccc/WestLake-7B-v2-laser-truthy-dpo is a 7 billion parameter language model built upon the cognitivecomputations/WestLake-7B-v2-laser base model. It underwent a fine-tuning process using the jondurbin/truthy-dpo-v0.1 dataset over two epochs with a learning rate of 2e-5, aiming to enhance its truthfulness and alignment.

Key Capabilities & Performance

This model demonstrates strong performance across various benchmarks, as evaluated on the Open LLM Leaderboard:

• Average Score: 75.37
• AI2 Reasoning Challenge (25-Shot): 73.89
• HellaSwag (10-Shot): 88.85
• MMLU (5-Shot): 64.84
• TruthfulQA (0-shot): 69.81
• Winogrande (5-shot): 86.66
• GSM8k (5-shot): 68.16

An evaluation of its GGUF version using EQ-Bench reported a score of 75.15. The model supports a context length of 8192 tokens.

Usage and Integrations

The model is available in various quantized formats, including GGUF and ExLlamav2 (3.5 through 8 bpw) provided by bartowski. It is designed to work with a ChatML-aligned prompt template, although the README notes that ChatML templating may not function as expected directly within the transformers library, providing a specific code example for proper multi-turn conversation handling.