cstr/Spaetzle-v12-7b

TEXT GENERATIONConcurrent Unit Cost:1Model Size:7BQuant:FP8Context Size:4kTool Calling:SupportedPublished:Mar 11, 2024License:cc-by-sa-4.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

cstr/Spaetzle-v12-7b is a 7 billion parameter merged language model, built upon mayflowergmbh/Wiedervereinigung-7b-dpo-laser and incorporating flemmingmiguel/NeuDist-Ro-7B, Blizado/discolm-mfto-7b-german-v0.1, and ResplendentAI/Flora_DPO_7B. This model is specifically optimized for German language tasks, showing improved performance on benchmarks like EQ-Bench (de) compared to its predecessors, while maintaining general English capabilities. It is designed for applications requiring strong German language understanding and generation.

Loading preview...

Spaetzle-v12-7b Overview

Spaetzle-v12-7b is a 7 billion parameter language model created by cstr through a merge of several specialized models using LazyMergekit. Its foundation is mayflowergmbh/Wiedervereinigung-7b-dpo-laser, augmented with flemmingmiguel/NeuDist-Ro-7B, Blizado/discolm-mfto-7b-german-v0.1, and ResplendentAI/Flora_DPO_7B. This specific merge strategy aims to enhance performance in German language tasks.

Key Capabilities & Performance

  • German Language Optimization: While showing a slight decrease in general English tasks compared to cstr/spaetzle-v8-7b, Spaetzle-v12-7b demonstrates improved performance on German benchmarks, achieving an EQ-Bench (de) score of 64.81.
  • General Benchmarks: The model's overall performance includes an average score of 69.36 across various benchmarks such as HellaSwag (86.16), MMLU (63.48), and GSM8k (62.70).
  • AGIEval: Achieves an average of 42.64% on AGIEval tasks, with notable scores in SAT English (76.21%) and LSAT Logical Reasoning (48.82%).
  • GPT4All: Scores an average of 74.3% on GPT4All, including 87.83% on BoolQ and 83.67% on ARC Easy.
  • TruthfulQA: Records a TruthfulQA mc2 score of 58.44%.

Licensing and Provider Obligations

As a model merge, cstr is considered the provider under the EU AI Act Art. 53. The model is licensed under cc-by-sa-4.0, derived from the most restrictive license of its constituent models, specifically ResplendentAI/Flora_DPO_7B.