Model Overview

cstr/llama3-8b-spaetzle-v20 is an 8 billion parameter language model created by cstr through a merge of two distinct models: cstr/llama3-8b-spaetzle-v13 and nbeerbower/llama-3-wissenschaft-8B-v2. The merging process utilized the dare_ties method, with specific parameters applied to nbeerbower/llama-3-wissenschaft-8B-v2 including a density of 0.65 and a weight of 0.4.

Performance Benchmarks

This model demonstrates competitive performance across several key benchmarks. On the EQ-Bench v2_de, it achieved a score of 65.7 with 171/171 parseable results. According to the Open LLM Leaderboard, cstr/llama3-8b-spaetzle-v20 has an average score of 71.83. Detailed scores include:

• ARC: 70.39
• HellaSwag: 85.69
• MMLU: 68.52
• TruthfulQA: 60.98
• Winogrande: 78.37
• GSM8K: 67.02

Key Characteristics

• Architecture: Merged Llama 3-based models.
• Parameter Count: 8 billion parameters.
• Context Length: 8192 tokens.
• Merging Method: dare_ties for combining model strengths.
• Data Type: bfloat16 for efficient computation.

Usage Considerations

This model is suitable for a range of general-purpose language generation and understanding tasks, particularly where a balance of performance and efficiency for an 8B model is desired. Its benchmark scores suggest proficiency in common reasoning, factual recall, and language comprehension tasks.