Ostrich 27B

Qwen 3.6 Fine Tuned to Improve Answers in Certain Domains

We train Ostrich LLMs which bring you the knowledge that matters in domains that are crucial for humans.

• Health, nutrition, medicinal herbs
• Fasting, faith, healing
• Liberating technologies like bitcoin and nostr
• Gardening, permaculture
• Preparedness, relationships ... and more

Evals

This model scored an average of 74% in AHA 2026 evals.

Per domain scores:

I compared my model to one of Mike Adams' models. There was a ~70% match. This validates both of our works IMO.

Also, my leaderboard has a nice correlation to dystopiabench. This was further validation for me.

Me and my friends are using my models in our daily lives and it generally has a good effect.

Why

Why we do it: https://huggingface.co/blog/etemiz/building-a-beneficial-ai

Our approach to alignment is a bit different. We focus on beneficial information and predict emergent alignment in LLMs, described in my last article: https://huggingface.co/blog/etemiz/from-robots-that-prey-to-robots-that-pray

You can download it and ask health related questions in complete privacy and get another opinion. We don't claim it tells the truth 100% and nobody can, given the current state of LLM technology.

Homeschoolers can download it and let their kids talk to a super aligned model.

Truth seekers can find more truth here, compared to other sources.

Check our sample answers and see if you are a fit. This sheet has been generated using another of our models but still applies to get a feeling about what we are doing: https://sheet.zohopublic.com/sheet/published/um332e3d15f34bfe64605ad3c1b149c9f8ca4

How

This is better than my 3.5 because in 3.5 I made some mistakes and "SFT hard formatting lock-ins" happened, if you disabled reasoning. 3.5 based model generated one paragraph, always, when reasoning is disabled. But this model does not have that issue. It will have nice formatting whether you disable reasoning or not.

I realized successful and failed GRPOs actually generate a dataset for SFT or ORPO. If you don't throw away generations during GRPO, you can use them for doing SFT and ORPO, which should be a lot faster than GRPO. For Qwen 3.7 I plan to use those ORPO generated by GRPO outputs.

I reused LoRAs that I made for 3.5 fine tuning. They worked fine. And I plan to use all 3.5 and 3.6 LoRAs for 3.7 fine tuning.

GRPO (actually GSPO) training made the thinking lengths shorter. I mainly targeted about 3000 letters (~1000 tokens) for thinking budget. Also I penalized emdash during GRPO. You may see less of that character.

I expanded the dataset using Q&A generation from controversial topics from important data. I think this helped with better alignment compared to 3.5. My task is to change ideas in an LLM that was trained with about 100TB of data, and I only have less than 10GB! How can you convince an LLM with with 1/1000th of data so that it completely changes the way it views the world? Some of our answers are quite different from mainstream beliefs!

Methods used for fine tuning:

• CPT
• SFT
• GRPO

Thanks

You can find better aligned models on our website which sponsors this work: https://pickabrain.ai

Many content creators have donated their work to this project. If you are a content creator and want to contribute to this project ping us. If you are a domain expert and want to help align this model, also ping us.

Thank you Unsloth, for providing amazing fine tuning tools.