ScaRE multilingual reasoning model suite

This public repository contains final inference checkpoints used in the multilingual ScaRE experiments. It includes German- and Korean-trained final models for the three hint schedules and the evaluated full-training and rank-1 extrapolation conditions, plus two additional ScaRE checkpoints used in supplemental evaluations. Every model folder contains the inference weights, model/tokenizer configuration, and a SHA-256 manifest.

Checkpoints

Training setting	Method/condition	Checkpoint	Subfolder
German	Hint-free (RELEX control)	Full 250-step model	de/none/full_step250
German	Hint-free (RELEX control)	Development-selected rank-1 extrapolation	de/none/selected_rank1
German	Fixed hints (ScaRE)	Full 250-step model	de/fixed/full_step250
German	Fixed hints (ScaRE)	Development-selected rank-1 extrapolation	de/fixed/selected_rank1
German	Hint decay (ScaRE decay)	Full 250-step model	de/decay/full_step250
German	Hint decay (ScaRE decay)	Development-selected rank-1 extrapolation	de/decay/selected_rank1
Korean	Hint-free (RELEX control)	Full 250-step model	ko/none/full_step250
Korean	Hint-free (RELEX control)	Development-selected rank-1 extrapolation	ko/none/selected_rank1
Korean	Fixed hints (ScaRE)	Full 250-step model	ko/fixed/full_step250
Korean	Fixed hints (ScaRE)	Development-selected rank-1 extrapolation	ko/fixed/selected_rank1
Korean	Hint decay (ScaRE decay)	Full 250-step model	ko/decay/full_step250
Korean	Hint decay (ScaRE decay)	Development-selected rank-1 extrapolation	ko/decay/selected_rank1
Mixed German/Korean	ScaRE rank-1, selected on development data	32-step supplemental checkpoint	mixed_de_ko/fixed/selected_rank1
German	ScaRE alpha=0.5	Magnitude-evaluation checkpoint	de/fixed/alpha05

none is the hint-free RELEX control. fixed and decay are the ScaRE and ScaRE-decay training schedules. Rank-1 candidates were selected on hint-free development accuracy in the training language. The alpha=0.5 model is the separately evaluated magnitude-sensitivity checkpoint.

Loading

from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "Owenboman/scare"
subfolder = "de/fixed/selected_rank1"  # choose a subfolder from the table
tokenizer = AutoTokenizer.from_pretrained(repo_id, subfolder=subfolder)
model = AutoModelForCausalLM.from_pretrained(
    repo_id, subfolder=subfolder, device_map="auto", torch_dtype="auto"
)

Limitations and license

Experiments use a single training seed per setting. These artifacts do not include training/evaluation data, logs, optimizer states, or intermediate training checkpoints. Extrapolated model selection used development data, not test results. The checkpoints derive from Qwen/Qwen3-4B-Base; the base model's Apache-2.0 license and terms apply.