alfalfa-w/self-conditioning-deep-reasoning

Self-conditioning for deep visual reasoning

This repository contains the available main-experiment checkpoints for the self-conditioning reproduction repository. The checkpoint inventory identifies each task, method, training step, base model, and file checksum.

Experiment	Available checkpoints	Selection
Line tracing	2 transformers	Vanilla at 60000, SelfCond at 53000
Long multiplication	2 LoRA adapters	Vanilla at 325000, SelfCond at 125000
TinyStories rendering	2 LoRA adapters	Vanilla at 190000, SelfCond at 74000
Controlled DiffThinker	10 LoRA adapters	Vanilla at 40000 and SelfCond at 16000 for VSP, Maze, TSP, Sudoku, and Jigsaw
Large-scale DiffThinker	6 LoRA adapters	The final five-epoch SelfCond checkpoint for each of the five tasks and the four-task unified model

The VanillaFM/SelfCondFM checkpoint pairs are matched by cumulative training FLOPs. Each pair uses the nearest evaluated VanillaFM checkpoint whose budget is no smaller than its SelfCondFM counterpart.

Line tracing uses a transformer initialized from scratch with the 9M_LONG configuration. FLUX LoRA adapters require black-forest-labs/FLUX.2-klein-base-4B. QwenImage LoRA adapters require Qwen/Qwen-Image-Edit-2511. The checkpoint configuration files are included; the base model, VAE, and text encoder remain separate dependencies.

Download a selected checkpoint with huggingface_hub:

from huggingface_hub import snapshot_download

root = snapshot_download(
    repo_id="alfalfa-w/self-conditioning-deep-reasoning",
    allow_patterns=[
        "01_line_tracing/line_tracing/selfcond/checkpoint-53000/*",
        "MANIFEST.json",
    ],
)
transformer_path = (
    f"{root}/01_line_tracing/line_tracing/selfcond/checkpoint-53000/transformer"
)

Use the reproduction evaluator with --transformer_path for line tracing and --lora_path for adapters. The method router uses EXPERIMENT_METHOD=vanilla or EXPERIMENT_METHOD=proj_condition for SelfCond. Use the experiment's configuration to set its remaining sampling parameters.

MANIFEST.json records the code and dataset revisions, observed training state, and SHA-256 digest of each weight and configuration file. The accompanying dataset repository is alfalfa-w/self-conditioning-deep-reasoning-data.