CoolFace
Modelpublic

socius/Llama-Centaur-3B-LoRA-r16-f0.0625

sourceHugging Facellama3.2updated 2mo agoView on Hugging Face
0likes11downloads
Model Card

<div align="center"> <img src="Llama-Centaur-3B-LoRA-r16-f0.0625.png" alt="Llama-Centaur-3B-LoRA-r16-f0.0625" width="1000">

![Meta Llama](https://huggingface.co/meta-llama/Llama-3.2-3B)

![socius](https://huggingface.co/collections/socius/llama-centaur-3b-lora-6a377e5b7251293c9fec0a94) ![Paper](https://arxiv.org/abs/2608.05224) ![Parameters](https://huggingface.co/socius/Llama-Centaur-3B-LoRA-r16-f0.0625) ![LoRA](https://huggingface.co/socius/Llama-Centaur-3B-LoRA-r16-f0.0625) ![Dataset](https://huggingface.co/datasets/marcelbinz/Psych-101) ![Data fraction](https://huggingface.co/datasets/marcelbinz/Psych-101) </div>

Llama-Centaur-3B-LoRA-r16-f0.0625

LoRA adapter for Llama-Centaur-3B, fine-tuned on a stratified 6.25% subset of Psych-101 as part of the LoRA-rank sweep and dataset-size ablation for Small Foundation Models of Human Cognition and Behaviour.

fieldvalue
base modelunsloth/Llama-3.2-3B
LoRA rank16 (alpha = rank, rsLoRA)
data fraction6.25% of Psych-101
training1 epoch, completion-only loss, seed 3407

Load with PEFT on top of unsloth/Llama-3.2-3B, or evaluate with the project's eval_model.py --backend unsloth.