Nitral-AI/Captain_BMO-12B
Nitral-AI/Captain_BMO-12B is a 12 billion parameter instruction-tuned language model based on the Nemo 12B instruct architecture, offering a context length of 32768 tokens. It was trained on a specialized dataset including GU_instruct-Remastered-1.1 and hathor/poppy, making it suitable for general instruction-following tasks. This model uses Mistral formatting and is provided as a one-off release for internal testing purposes.
Loading preview...
Nitral-AI/Captain_BMO-12B Overview
Nitral-AI/Captain_BMO-12B is a 12 billion parameter instruction-tuned language model built upon the Nemo 12B instruct base architecture. It supports a substantial context length of 32768 tokens, allowing for processing longer inputs and generating more coherent responses.
Key Characteristics
- Base Model: Nemo 12B instruct.
- Training Data: Fine-tuned on a unique blend of datasets, including a 200k randomized subset of
GU_instruct-Remastered-1.1and 25khathor/poppydata, trained over 3 epochs. - Formatting: Utilizes Mistral formatting for prompts and responses.
- Quantizations: Available in various quantized formats, including GGUF from Bartowski and Exl2 quantizations (4bpw and 6bpw) from Nitral-AI.
Intended Use and Support
This model was primarily developed for internal testing and is released as a "one-off train." Users should note that extended support for this model is not planned, and performance may vary depending on specific use cases and context size. It is best suited for experimentation with instruction-following tasks where its specific training data might offer unique response characteristics.