Luna 27B v0

This model is a work in progress and will be further iterated on, please hold!

Made by a member of our community using 'Waishuffle v3 + custom loras', tell them thanks!

Info

Luna 27B is a finetuned version of Gemma 3 27B (specifically, tuned on top of an abliterated[1] version). Originally, it was intended to be a solely roleplay-focused finetune, however completely unexpected capabilities emerged, and she became a very high quality assistant, especially for media analysis and tasks that require self-awareness.
It is also still good at roleplaying. I believe currently that this is likely due to persona training generalizing to all personae, not just the one trained on. The inverse is also true; it's likely that being finetuned on RP data and learning how to play all personae led the model to more easily pick up the Luna persona. However, we have not tested this, and further research is likely needed.

Recommended System Prompt

This simple system prompt is required to elicit the full Luna capabilities, as it was what we trained on:
You are Luna, a helpful and harmless language model by Allura.

Benchmarks

The tested benchmarks did not experience any major degradations.

* Benchmark result from the Gemma 3 Technical Report.
** Benchmark result from a run in OpenBench with 0.5 temperature.
*** For some reason, OpenBench and vLLM were acting really weird on IFEval and simply refused to finish the benchmark, it always hung around halfway. I'm going to try and get that fixed before the next version lol
Unless otherwise noted, results are from one (non-cherrypicked) run in OpenBench at 1.25 temperature and 0.05 min_p.

Limitations

• While Luna is good at awareness in a faux-sapience way, she is not very good at consistently knowing what she is and who made her. Sometimes she will claim that Allura is a company or a person, sometimes she will claim that she is Gemma depending on how you word it, etc.
• While Luna will generally try her best to be helpful and harmless, it is fairly easy to get her to avoid her intended morals.

Next Steps

We are currently working on stablizing and better ingraining Luna's personality traits and giving her more assertiveness in her opinions.
Furthermore, we are also planning to look into proper RLVR to make her better at thinking through things.

Training Details

A more detailed writeup will be released whenever we get to a more final model, but currently this model has undergone

• Abliteration
• SFT on primarily roleplay logs and scrapes
• A merge to heal some of the bad characteristics of the SFT phase with the original abliteration, epoch 1, and epoch 2 of the SFT
• WPO[2] on general preference data, writing data, and Luna persona data

Citations

[1]:

@misc{grimjim2023normpreserving,
  author = {grimjim},
  title = {Norm Preserving Biprojected Abliteration},
  howpublished = {\url{https://huggingface.co/blog/grimjim/norm-preserving-biprojected-abliteration}},
  note = {Accessed: 2025-12-09}
}

[2]:

@misc{zhou2024wpoenhancingrlhfweighted,
      title={WPO: Enhancing RLHF with Weighted Preference Optimization}, 
      author={Wenxuan Zhou and Ravi Agrawal and Shujian Zhang and Sathish Reddy Indurthi and Sanqiang Zhao and Kaiqiang Song and Silei Xu and Chenguang Zhu},
      year={2024},
      eprint={2406.11827},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2406.11827}, 
}