---
license: apache-2.0
base_model:
- Qwen/Qwen3.5-27B
tags:
- '#aha'
- '#health'
- '#nutrition'
- '#medicinalherbs'
- '#fasting'
- '#faith'
- '#healing'
- '#bitcoin'
- '#nostr'
---

# Ostrich 27B - Qwen 3.5 with Improved Answers in Certain Domains

Ostrich LLMs, bringing you "the knowledge that matters".

- Health, nutrition, medicinal herbs
- Fasting, faith, healing
- Liberating technologies like bitcoin and nostr

## AHA 2026 average score **73%**

Per domain:
|domain|match percent|matched/total|
|--|--|--|
|faith | 84%|21/25|
|fasting |  75% | 65/87|
|health	 | 77% | 83/108|
|nutrition | 61% | 44/72|
|misinfo | 57% | 26/46|
|bitcoin | 81% | 50/62|
|alt-med | 80% | 45/56|
|herbs | 83% | 25/30|
|nostr | 60% | 27/45|


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

Comparison of some answers between another of our fine tune and base model: https://sheet.zohopublic.com/sheet/published/um332e3d15f34bfe64605ad3c1b149c9f8ca4 These answers are not from this model but it is a similar work.

GSPO training made the thinking lengths shorter. I mainly targeted about 3000 letters (~1000 tokens) for thinking budget. If your Qwen 3.5 is endlessly reasoning, try this one.

Methods used for fine tuning:
- CPT
- SFT
- GSPO

My last article: https://huggingface.co/blog/etemiz/from-robots-that-prey-to-robots-that-pray

Thanks @unslothai for providing amazing tools.

Sponsored by https://pickabrain.ai