oof-baroomf/image-models-are-emergent-reasoners

Image diffusion arithmetic research artifacts

Preservation archive for experiments on arithmetic with pretrained image diffusion models and comparison models. It includes unsuccessful experiments, controls, intermediate checkpoints, synthetic images, and evaluation records. These artifacts do not establish a pretraining advantage or general mathematical reasoning ability.

Contents and recovery

checkpoint-manifest.json maps original project paths to gzip-compressed PyTorch checkpoint files under checkpoints/. Verify the compressed file's byte length and SHA-256 hash, then decompress it to recover the saved state.

weights-manifest.json maps original paths to individual PyTorch or safetensors files under weights/sha256/. Identical weight files share one stored object. These are experiment states, not standalone model pipelines.

original-archives-manifest.json maps larger original experiment archives to objects under original-archives/sha256/. The original archive bytes are preserved.

artifact-manifest.json lists compressed parts under artifacts/. Each part has a companion JSON file under artifact-manifests/ listing its member paths, byte lengths, SHA-256 hashes, and executable modes. Find the original project-relative path in these manifests, download the corresponding part, verify the archive hash, and extract the desired member. The part manifests contain no local absolute paths.

Use an immutable repository revision when downloading files. The preservation records accompanying local removals record the exact revision used for verification. archive-summary.json records the completed release's counts.

Verification and exclusions

Every released object was downloaded anonymously from its immutable Hub revision into a new local file and checked against the original hash. Every member of a newly packed artifact archive was additionally restored to a real file and checked for path, size, hash, and mode parity. Local originals were eligible for removal only after these checks passed.

Selected text and serialized metadata were inspected for personal information and credentials. A separate secret scanner checked the selected text. Checkpoint pickle metadata was inspected without executing pickle payloads. Every selected image underwent format and metadata inspection; representative images from 232 experiment groups were visually reviewed as synthetic arithmetic, glyph, or rule-task outputs. This is not a claim that every raster received OCR or an individual visual review.

Files identified as credentials, Codex transcripts, personal records, unresolved privacy cases, or unclassified binary formats were excluded. Current experiment dependencies, Git-tracked files, working source, and text records remain local even when archived here. This release covers a frozen selection of project files; it is not a live mirror of ongoing experiments.

Weights and other material derived from upstream projects remain subject to their upstream and file-specific terms. This archive grants no new blanket license for third-party material. Only load serialized models from sources you trust.