cstr/Spaetzle-v60-7b

Overview

cstr/Spaetzle-v60-7b is a 7 billion parameter language model developed by cstr. It is a progressive merge, primarily using the dare-ties method, combining abideen/AlphaMonarch-dora and cstr/Spaetzle-v58-7b. The primary goal of this merge is to achieve a suitable compromise for tasks requiring proficiency in both English and German.

Key Capabilities

• Bilingual Performance: Designed for balanced performance in English and German language tasks.
• Competitive Benchmarking: Achieves a score of 65.08 on EQ-Bench (v2_de) and shows competitive results on the Occiglot Euro LLM Leaderboard against models like Mixtral-8x22B-v0.1 and Llama-3-SauerkrautLM-8b-Instruct, particularly in German and English benchmarks.
• Quantized Performance: The cstr/Spaetzle-v60-7b-int4-inc version demonstrates strong performance in low-bit quantized leaderboards, scoring 68.01 on average.

Good For

• Bilingual Applications: Ideal for use cases requiring robust language understanding and generation in both English and German.
• Local German Tasks: Specifically tuned to perform well on German-centric tasks.
• Research and Development: Provides a merged model for exploring progressive merging techniques and their impact on bilingual performance.

Important Considerations

As a model merge, cstr is considered the 'provider' under the EU AI Act Art. 53, carrying obligations related to copyright policy and training content, even though no new training data was used in its creation. The training content is inherited from its constituent models.