Atlas v21. This version was to test the 31B Dense model. This is the full merged model.
-
This is an experimental model.
-
The safety refusals have been altered, do not use without proper safety layers or system prompt.
-
Trained on curated dataset comprising 60% Opus 4.6/Sonnet 4.6 thinking traces across Math, Science, General Knowledge, 40% Mental Health conversations 50/50 split between synthetic and authentic*
-
*authentic conversations have been cleaned to remove any personal information and I have approval from the sources
-
Developed by: senaro
-
License: apache-2.0
-
Finetuned from model : senaro/atlas-gemma4-31b-trm7
This gemma4 model was trained 2x faster with Unsloth and Huggingface's TRL library.
License & Attribution
This model is a derivative work fine-tuned from the base Gemma 4 architecture. Use of this model is subject to the Gemma Terms of Use.
Gemma is a trademark of Google LLC.
Citation
BibTeX:
@misc{gemma2026gemma4,
title={Gemma 4 Technical Report},
author={Gemma Team},
year={2026},
eprint={2607.02770},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
APA 7th Edition: Gemma Team. (2026). Gemma 4 Technical Report (arXiv:2607.02770). arXiv. https://arxiv.org