eulogik/fugusashi-preferences

Fugusashi Preferences — Model Router Training Dataset

What is This?

Community-curated prompt → model preference dataset for Fugusashi router training. Each entry maps a prompt to the optimal model, used to train the CMA-ES evolved routing weights.

Part of the Fugusashi project — Like Sakana Fugu. But Free.

Dataset Format

JSONL with fields:

Field	Type	Description
prompt	string	The user prompt
preferred_model	string	Optimal model for this prompt
source	string	Data source (e.g., eulogik-seed)
category	string	Prompt category (e.g., code, creative, reasoning)
score	float	Preference confidence (0-1)

Quick Load

from datasets import load_dataset

dataset = load_dataset("eulogik/fugusashi-preferences")
for sample in dataset["train"]:
    print(f"{sample['prompt'][:50]}... → {sample['preferred_model']}")

Data Sources

Source	Samples	Description
eulogik-seed	20	Initial seed data curated by eulogik

How to Contribute

We welcome community contributions! To add your own preference samples:

1. Fork this dataset on HuggingFace
2. Add your entries in JSONL format
3. Submit a pull request

Contribution Guidelines

• Prompts should be diverse (code, creative, reasoning, analysis)
• Models available: gpt-oss-120b, nemotron-3-ultra, nemotron-3-super, hermes-3-405b, lfm-2.5-1.2b
• Set score based on how confident you are in the preference (0.5-1.0)
• Include realistic prompts that users would actually send

Project Links

Resource	Link
Source Code	github.com/eulogik/fugusashi
PyPI Package	pypi.org/project/fugusashi
Live Demo	huggingface.co/spaces/eulogik/fugusashi
Router Model	huggingface.co/eulogik/fugusashi-router
Website	eulogik.github.io/fugusashi
eulogik	eulogik.com

Citation

@dataset{fugusashi-preferences2026,
  title={Fugusashi Model Router Preferences},
  author={{eulogik}},
  year={2026},
  url={https://huggingface.co/datasets/eulogik/fugusashi-preferences}
}

License

MIT — use it however you want.

---

Built with ❤️ by eulogik

Fugu Sashi. Served Free.