HyperbeeAI/Tulpar-7b-v0

TEXT GENERATIONConcurrency Cost:1Model Size:7BQuant:FP8Ctx Length:4kPublished:Aug 23, 2023License:llama2Architecture:Transformer0.0K Open Weights Cold

HyperbeeAI's Tulpar-7b-v0 is a 7 billion parameter language model built upon the Llama2-7b architecture, trained on a curated instruction-finetuning dataset including GPT-4 generated data. This model is optimized for general instruction-following tasks, demonstrating capabilities across various benchmarks like MMLU, HellaSwag, and BigBenchHard. It is primarily intended for English-language applications requiring robust conversational and reasoning abilities.

Loading preview...

HyperbeeAI/Tulpar-7b-v0: An Instruction-Tuned Llama2-7b Model

Tulpar-7b-v0 is a 7 billion parameter language model developed by HyperbeeAI, based on the Llama2-7b architecture. It has been instruction-finetuned using a carefully filtered and preprocessed dataset, which incorporates high-quality, GPT-4 generated data alongside established datasets such as Airoboros and Platypus.

Key Capabilities & Performance

This model demonstrates general instruction-following capabilities, with its performance evaluated across several benchmarks:

  • HF Leaderboard Evaluation: Achieved an average score of 0.5979, with specific results including 0.5614 on arc_challenge, 0.7901 on hellaswag, 0.5242 on mmlu, and 0.5160 on truthfulqa_mc.
  • GPT4All Evaluation: Showed an overall average of 0.6468, with scores like 0.8306 on boolq, 0.7905 on piqa, and 0.7159 on winogrande.
  • BigBenchHard: Recorded an average score of 0.3754 across various complex reasoning tasks.

Intended Use and Limitations

Tulpar-7b-v0 is designed for a broad range of instruction-based applications. However, it is important to note that the model is exclusively finetuned in English, and its performance in other languages or multilingual scenarios is not covered. Users are advised to conduct thorough safety tests for their specific use cases before deployment, as HyperbeeAI does not guarantee ethical, accurate, unbiased, or objective responses.