README.md

---
title: Centauri Cognitive Simulator
emoji: 🧠
colorFrom: indigo
colorTo: green
sdk: gradio
sdk_version: 6.22.0
app_file: app.py
pinned: false
license: apache-2.0
short_description: Small cognition models predict your choices, trial by trial
python_version: "3.12"
startup_duration_timeout: 1h
models:
  - socius/Qwentaur-8B-LoRA-r16
  - socius/Qwentaur-0.6B-LoRA-r16
datasets:
  - marcelbinz/Psych-101
---

# 🧠 Centauri Cognitive Simulator

Play a real psychology experiment while three models predict, trial by trial, what a
**human** would do next:

| predictor | what it is |
|---|---|
| **Qwentaur-8B** | [`socius/Qwentaur-8B-LoRA-r16`](https://huggingface.co/socius/Qwentaur-8B-LoRA-r16) — LoRA r=16 on `unsloth/Qwen3-8B-Base` |
| **Qwentaur-0.6B** | [`socius/Qwentaur-0.6B-LoRA-r16`](https://huggingface.co/socius/Qwentaur-0.6B-LoRA-r16) — 13× smaller |
| **Qwen3-8B-Base** | the same 8B weights with the adapter switched off (no fine-tuning) |

From *Small Foundation Models of Human Cognition and Behaviour* (Oh & Gobet, 2026) —
[paper](https://huggingface.co/papers/2608.05224) ·
[code](https://github.com/socius-org/Centauri). The paper's headline claim is that
sub-billion-parameter models already match a 70B Centaur at fitting human choices
in-distribution; the side-by-side bars and the running loss table let you watch that.

## How it works

Everything the models see is plain [Psych-101](https://huggingface.co/datasets/marcelbinz/Psych-101)
text, exactly as during fine-tuning: human choices are wrapped in `<<...>>`, and the
choice distribution is read straight off the next-token logits after `You press <<`,
renormalised over the available option letters. No sampling, no prompt engineering,
no chat template.

Four experiments are implemented with the dataset's verbatim instruction wording:

- **Horizon task** (Wilson et al., 2014) — explore or exploit two slot machines
- **Iowa Gambling Task** (Steingroever et al., 2015) — learning under risk
- **Intertemporal choice** (Ruggeri et al., 2022) — patience
- **Multi-attribute inference** (Hilbig & Moshagen, 2014) — cue integration

The second tab scores the models the way the paper does — mean negative log-likelihood
of a **real participant's** actual choices — on human sessions bundled verbatim from
Psych-101 (Binz et al., 2025, Apache-2.0).

## Credits

- Adapters: `socius/*` (Apache-2.0), Oh & Gobet, *Small Foundation Models of Human
  Cognition and Behaviour*, 2026.
- Base models: `unsloth/Qwen3-8B-Base`, `unsloth/Qwen3-0.6B-Base`.
- Task wording and bundled human transcripts: [Psych-101](https://huggingface.co/datasets/marcelbinz/Psych-101)
  (Binz et al., 2025, Apache-2.0).