yyxdnmd/DiARC-adapters

DiARC Adapters

This repository contains the selected best PEFT/LoRA adapters for DiARC across three backbone models and six ARC-style datasets.

The adapters are organized as:

<model>/<dataset>/

Models: llama-3.2-3b, minitron-8b, and qwen3-4b.

Datasets: arc-agi-1, arc-agi-2, miniarc, conceptarc, 1d-arc, and arccommunity.

Each leaf directory contains adapter_model.safetensors, adapter_config.json, tokenizer metadata when present, and a short README.

Selected adapters

Model	Dataset	Score	Method	Setting	Path
Llama-3.2-3B	ARC-AGI-1	61.50%	Output-level / random perturb	run2	llama-3.2-3b/arc-agi-1
Minitron-8B	ARC-AGI-1	72.38%	DSL-level rule inversion	motif-3 representative; motif-2/motif-3 tie in Table 4	minitron-8b/arc-agi-1
Qwen3-4B	ARC-AGI-1	97.00%	Output-level / morphology	run1	qwen3-4b/arc-agi-1
Llama-3.2-3B	ARC-AGI-2	6.53%	Output-level / random perturb	eval10	llama-3.2-3b/arc-agi-2
Minitron-8B	ARC-AGI-2	10.28%	Output-level / morphology	eval10	minitron-8b/arc-agi-2
Qwen3-4B	ARC-AGI-2	33.06%	Output-level / rigid shift	eval14	qwen3-4b/arc-agi-2
Llama-3.2-3B	MiniARC	73.15%	Output-level / morphology	14x	llama-3.2-3b/miniarc
Minitron-8B	MiniARC	76.51%	Output-level / morphology	14x representative; morphology/rigid-shift tie in Table 3	minitron-8b/miniarc
Qwen3-4B	MiniARC	97.32%	Output-level / grid-block	14x representative; grid-block/morphology/random-perturb tie in Table 3	qwen3-4b/miniarc
Llama-3.2-3B	ConceptARC	59.17%	Output-level / morphology	14x representative; morphology/random-perturb tie in Table 3	llama-3.2-3b/conceptarc
Minitron-8B	ConceptARC	62.71%	Output-level / grid-block	6x representative; grid-block/random-perturb tie in Table 3	minitron-8b/conceptarc
Qwen3-4B	ConceptARC	97.29%	Output-level / rigid shift	3x	qwen3-4b/conceptarc
Llama-3.2-3B	1D-ARC	52.39%	Output-level / random perturb	4x	llama-3.2-3b/1d-arc
Minitron-8B	1D-ARC	58.82%	Output-level / random perturb	3x	minitron-8b/1d-arc
Qwen3-4B	1D-ARC	45.95%	Output-level / random perturb	2x old2d	qwen3-4b/1d-arc
Llama-3.2-3B	ARCcommunity	30.43%	Output-level / grid-block	10x	llama-3.2-3b/arccommunity
Minitron-8B	ARCcommunity	43.48%	Output-level / grid-block	14x representative; all four output-level families tie in Table 3	minitron-8b/arccommunity
Qwen3-4B	ARCcommunity	63.04%	Output-level / morphology	2x representative; morphology/rigid-shift tie in Table 3	qwen3-4b/arccommunity

Notes

• Scores are the paper-reported accuracies for the selected model-dataset pair.
• For entries with tied best scores, one representative adapter is uploaded and the tied local candidates are recorded in manifest.json.
• The original adapter configs may reference local SFT base-model paths. Use the corresponding DiARC SFT base model for each backbone.