Centauri is a research codebase for training and evaluating small foundation models of human cognition and behaviour. It fine-tunes small pretrained LLMs with LoRA to predict human choices on the Psych-101 benchmark, extending Centaur-style SFT (Binz et al., 2025) down to sub-billion-parameter models.
Centaur (Binz et al., 2025) showed a fine-tuned 70B LLM can predict human behaviour across
psychology experiments. Our study asks how small that model can be, and what it actually learns.
We LoRA-fine-tune four families of base models (Llama, Qwen, SmolLM, OLMo) — 135M to 14B
parameters — on Psych-101 (transcribed human behaviour across 160 psychological experiments),
varying adapter rank (r = 4, 8, 16, 32, 64) and training-set size (6.25, 12.5, 25, 50, 100%)
independently of scale.
How small? In-distribution, scale barely matters — all fourteen models sit in a narrow band, and 0.6B to 1B parameters suffice to match a reproduced Centaur-70B on held-out participants. Out of distribution that band opens into a markedly steeper gradient, so scale is worth buying for transfer rather than for fit.
What do they learn? Stripping the four prompt channels — instruction, stimuli, feedback and choice history — shows the predictions rest on what the participant saw and was told, not on choice-history shortcuts. Masking stimulus and feedback content destroys 75.7% of learned information and drives models below chance, while permuting trial order changes predictions only where the task makes order informative.
| Family | Base | Sizes |
|---|---|---|
 |
Llama 3.2 / 3.1 | 1B · 3B · 8B |
 |
Qwen3 | 0.6B · 1.7B · 4B · 8B · 14B |
 |
OLMo 2 / 3 · fully open | 1B · 7B |
 |
SmolLM2 / 3 | 0.1B · 0.4B · 1.7B · 3B |
Exact base checkpoints are pinned per size in each family’s train_*.py.
git clone https://github.com/socius-org/Centauri.git && cd Centauri
pip install torch transformers datasets peft unsloth trl wandb pandas numpy huggingface_hub
pip install matplotlib scienceplots scipy scikit-learn # analysis / figuresAll scripts run from the repo root.
# Train one cell
python qwentaur/train_qwentaur.py --size 0.6B --lora_rank 16 --wandb
# Schedule a whole family sweep (LoRA-rank sweep + dataset-size ablation)
python build_dataset_fractions.py # one-time: stratified subset indices
python schedule_training_runs.py --family smollm # add --dry_run to preview the plan
python schedule_training_runs.py --family qwen --sizes 1.7B --ranks 8 # just a subset
# Evaluate in-distribution (Psych-101, resumable)
python schedule_eval_runs.py --family qwen --resume
# Evaluate out-of-distribution (Psych-201 held-out experiments)
python schedule_eval_runs.py --family smollm --eval_script ./eval_model_ood.py \
--output_dir ./results/psych201 --resume
# A single model, directly
python eval_model.py --model socius/Qwentaur-8B-LoRA-r16 --backend unsloth \
--output_dir ./results/psych101The experiment grid (which size/rank/fraction cells exist per family) lives in one place —
utils.py — shared by both schedulers, the uploader and the figure scripts, so the
design stays in sync.
utils.py shared grid (FAMILIES) + cell helpers + HF naming + fraction loader
schedule_training_runs.py enumerate + launch training cells per family
schedule_eval_runs.py enumerate + launch eval per cell (against the socius HF repos)
build_dataset_fractions.py one-time: stratified subset indices for the data ablation
eval_model.py Psych-101 evaluator (unsloth / transformers backends)
eval_model_ood.py Psych-201 held-out evaluator
upload_adapters.py push trained adapters to socius/<RepoName>
organise_collections.py group HF adapter repos into collections
llama-centaur/ qwentaur/ olmotaur/ smoltaur/ each: train_<x>.py + assets/
results/
psych101/ flat per-task eval CSVs + psych101_aggr.csv + figures/ + tables/
psych201/ flat per-task OOD eval CSVs + figures/ + tables/
4bit/ 4-bit base + finetuned eval CSVs
lm_eval/ standard-benchmark results (cogsoc, metabench) + model-card generation
prompt_decomposition/ prompt-ablation studies
Experiments: (1) main scale curve at r=16; (2) LoRA-rank sweep r ∈ {4, 8, 16, 32, 64} at full
data; (3) dataset-size ablation at fixed r=16 over {6.25, 12.5, 25, 50, 100}%; (4) 4-bit QLoRA vs
bf16; (5) out-of-distribution eval on 18 held-out Psych-201 experiments; (6) prompt-channel ablation
and order permutation.
Each figure/table generator resolves its data relative to its own location — no arguments needed:
python results/psych101/figures/generate_scaling_plots.py # scale + precision curves
python results/psych101/figures/generate_ablation_plots.py --family qwen # rank / data-size
python results/psych101/figures/cognitive_baseline_comparison.py # vs domain-specific models
python results/psych101/tables/generate_heatmap_tables.py # LaTeX tables
python results/psych201/figures/generate_scaling_plots.py # out-of-distributionPer-task CSVs live flat in results/psych101 and results/psych201, named
socius-<Family>-<size>-LoRA-r<rank>[-f<fraction>].csv; psych101_aggr.csv is the aggregate the
figures read.
Everything is hosted under the socius HuggingFace organisation:
- 117 LoRA adapters —
socius/<Family>-<size>-LoRA-r<rank>[-f<fraction>], spanning the rank sweep and dataset-size ablation, grouped into per-size collections under the top-level Centauri collection. - Datasets —
marcelbinz/Psych-101(training),marcelbinz/Psych-101-test(46 held-out tasks),socius/Psych-201-RT.
@inproceedings{oh2026smallcogfm,
title = {Small Foundation Models of Human Cognition and Behaviour},
author = {Oh, Nick and Gobet, Fernand},
booktitle = {Third Conference on Language Modeling (COLM)},
year = {2026},
note = {arXiv:2608.05224}
}Built on the Centaur methodology and the Psych-101 dataset (Binz et al.).
Released under the MIT License. Fine-tuned model weights inherit the licenses of their respective base models (Llama, Qwen, OLMo, SmolLM).