README.md

metadata

license: mit
language:
  - en
base_model:
  - Salesforce/codet5-large
tags:
  - ARC-AGI
  - ARC
  - code
datasets:
  - mindware/arc-mega
  - Open-Orca/SlimOrca
  - camel-ai/math
  - skeskinen/TinyStories-GPT4
  - rajpurkar/squad_v2
  - garage-bAInd/Open-Platypus
  - Sharathhebbar24/arxiv-math-instruct-50k
  - AlgorithmicResearchGroup/arxiv-physics-instruct-tune-30k
  - TIGER-Lab/MathInstruct
  - neoneye/histogram-comparisons-small-v1
  - ise-uiuc/Magicoder-Evol-Instruct-110K
  - PrimeIntellect/INTELLECT-MATH-SFT-Data
  - PrimeIntellect/verifiable-math-problems
  - sethapun/arithmetic_2md_1to1000
  - EleutherAI/proof-pile-2
  - MMInstruction/M3IT
  - stingning/ultrachat
  - timdettmers/openassistant-guanaco
  - Dahoas/instruct-synthetic-prompt-responses
  - pankajmathur/WizardLM_Orca

This checkpoint is the primary CodeT5-based solver we used for the MindsAI @ Tufa Labs entry in the ARC Prize 2025 competition. It shares the same architecture as mindware/arc-codet5-660m-scr (a 16-layer decoder variant of Salesforce/codet5-large), but does not include the Span-Corruption Refinement (SCR) auxiliary training stage. Instead, it represents the best non-refinement checkpoint obtained during long-horizon pretraining on TPU-v4 systems.

• No SCR stage: this model was trained purely with the original span-corruption + instruction fine-tuning curriculum + ARC fine tunining.
• Decoder-only pruning: the original decoder depth (24) was reduced to 16 layers after experiments showed encoder pruning harmed sample efficiency, while decoder pruning could be recovered through extended training.
• Long-run TPU training: training spanned roughly two years on a V4-64 TPU, made possible by Google’s TPU Research Cloud program.

📚 ARC-Related Datasets & Frameworks

• RE-ARC — procedurally generates examples for the 400 ARC training tasks (we also include RE-ARC eval + ARC 1.5).
• ConceptARC
• 1D-ARC
• ARC_gym, Sort-of-ARC
• Andreas Koepf’s generator suites (includes RE-ARC-style grids, code generation targets, and solution graphs).
• Jack Cole’s custom generators covering ~70 tasks plus larger concept sets (cellular automata, math-derived boards, etc.).

Several auxiliary datasets predict task metadata (graphs, heuristics, explanations) rather than final boards; they are part of the broader instruction mixture this model saw during pretraining.

ARC Data Formatting

• ARC tasks ship as JSON where each task_id contains train pairs and test inputs; every grid is a rectangular list of lists with integers 0-9. Dimensions follow the original 1×1–30×30 spec, though the evaluator accepts up to 50×50.
• Example task payload:

  {
    "task_id": {
      "train": [
        {"input": [[0,0],[1,1]], "output": [[1,1],[1,1]]}
      ],
      "test": [
        {"input": [[0,0,0],[0,1,0],[0,0,0]]}
      ]
    }
  }
  

• Model prompts (prompt column during training/TTT/inference) are serialized text strings: solve: train input1 <train_input> output1 <prefix><train_output>. … test tinput1 <test_input> toutput1 . Each grid token <train_input> / <train_output> / <test_input> is produced by grid_to_string, so rows are concatenated digits separated by spaces. Multiple train examples increment the index (input2, output2, etc.).
• Prompt example:

  solve: train input1 000 010 000 output1 11 3 3 10 111 101 111. input2 00 02 output2 5 2 2 20 22 20. test tinput1 0000 0300 0000 0000 toutput1 
  

• Model targets (correct_answer column and expected decoder output before post-processing) follow output_prefix semantics: {total_chars} {height} {width} {symbols} {row_strings}. Here total_chars = height*width + (height - 1) and symbols is the deduplicated sequence of colors as they are first encountered when scanning the board row-major; that rule applies to every output grid we emit (training outputs inside the prompt and the predicted test toutput). Example target string for a 3×3 donut:

   11 3 3 10 111 101 111.