metadata
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