cstr/Spaetzle-v60-7b

Spaetzle-v60-7b: A Bilingual 7B Merged Model

Spaetzle-v60-7b is a 7 billion parameter language model developed by cstr, created through a progressive merge using the dare_ties method. It combines abideen/AlphaMonarch-dora and cstr/Spaetzle-v58-7b to achieve a suitable compromise for both English and German local tasks.

Key Capabilities & Performance

• Bilingual Optimization: Intended for balanced performance across English and German language tasks.
• Competitive Benchmarking: Achieves a score of 60.95 on the DE benchmark and 71.65 on the EN benchmark on the Occiglot Euro LLM Leaderboard, outperforming several other 7B models like meta-llama/Meta-Llama-3-8B and mistralai/Mistral-7B-Instruct-v0.2 in certain German metrics.
• Quantized Performance: The int4-inc quantized version demonstrates strong performance on the Low-bit Quantized Open LLM Leaderboard with an average score of 68.01, ranking closely behind Intel/SOLAR-10.7B-Instruct-v1.0-int4-inc while being a smaller model.
• Low Contamination: Contamination checks against Mistral instruct 7b v0.1 show very low results (e.g., MMLU < 0.1%, TruthfulQA < 0.1%), indicating minimal data overlap.

Ideal Use Cases

• Bilingual Applications: Excellent for scenarios requiring robust performance in both English and German.
• Resource-Constrained Environments: The 7B parameter size and strong quantized performance make it suitable for deployment where computational resources are limited.
• General Text Generation: Capable of handling a variety of text generation tasks in both supported languages.