---
license: apache-2.0
base_model: Qwen/Qwen3-30B-A3B
library_name: transformers
pipeline_tag: text-generation
datasets:
- EdisonScientific/EDB-Bio
language:
- en
tags:
- biology
- experimental-design
- simulation
---

# Ibis0 BED outcome simulator

[paper](ARXIV_URL) · [code](https://github.com/EdisonScientific/ibis0) · [data](https://huggingface.co/datasets/EdisonScientific/EDB-Bio)

The outcome simulator from the Bayesian experimental design (BED) experiments in **"Do Language Models Know
What Compelling Scientific Evidence Looks Like?"**. Given a candidate experiment with its result hidden, the
hypothesis θ′ that the experiment itself tests, and a requested valence, it generates a plausible outcome:
either compelling support for θ′ or a clear null. During experiment selection the two outcomes are weighted
by the belief model's p(θ′ | evidence) to estimate expected information gain. The belief model is
[Ibis0-30B-A3B-Classifier](https://huggingface.co/EdisonScientific/Ibis0-30B-A3B-Classifier).

## Clone and install ibis0

The prompt builder and BED scripts below live in the
[EdisonScientific/ibis0 code repository](https://github.com/EdisonScientific/ibis0).
Install Python 3.12 and [uv](https://docs.astral.sh/uv/), then clone and install the repository:

```bash
git clone --recurse-submodules https://github.com/EdisonScientific/ibis0.git
cd ibis0
uv sync --frozen --all-groups
```

Run the referenced scripts from this repository with `uv run python <script-path>`.
Local GPU inference requires Linux and a CUDA GPU.

## Use

[`build_theta_prime_sft.py`](https://github.com/EdisonScientific/ibis0/blob/main/scripts/data_pipeline/build_theta_prime_sft.py)
builds the simulator's prompts from an experiment, its hypothesis, and the requested outcome type.
[`bed_thetaprime.py`](https://github.com/EdisonScientific/ibis0/blob/main/scripts/analysis/rq5_bed/bed_thetaprime.py)
uses the simulator and classifier together to select experiments.

The simulator returns JSON with three fields: `supports_hypothesis` (`"yes"` or `"no"`),
`weight_of_evidence_decibans`, and `y`, which contains the simulated measurements for the domain.

## Licence

Apache-2.0. Fine-tuned from [Qwen/Qwen3-30B-A3B](https://huggingface.co/Qwen/Qwen3-30B-A3B) (Apache-2.0).

## Citation

```bibtex
@article{ibis0,
  title   = {Do Language Models Know What Compelling Scientific Evidence Looks Like?},
  author  = {Igoe, Conor and Grand, Gabriel and Koch, Zane and Bou, Albert and Moss, Eli and Hinks, Michaela M. and Rodriques, Samuel G. and White, Andrew and Narayanan, Siddharth},
  journal = {arXiv preprint arXiv:ARXIV_ID},
  year    = {2026}
}
```