Ruler138/CodeAnything-1.835M

CodeAnything 1.835M SFT and evaluation release

Gated public release containing the clean 1,835,476-sample training set, the 800-sample/16-domain evaluation set, paper-model predictions and rendered outputs, and raw per-sample rating records.

Layout

training/
  manifest/
    all_training_v5.jsonl
    all_training_v5.jsonl.idx
    all_training_v5.jsonl.true_lengths.u32
  shards/<domain>/           exact media/code closure (tar shards)
evaluation/
  benchmark/                 800 GT code + image/video samples
  model_outputs/v4/<model>/  predictions.json + rendered/
  model_outputs/v5/<model>/  predictions.json + rendered/
  ratings/v3|v4|v5/          raw judge JSON records

The training manifest contains domain, language, code_file, image_file/video_file, and code. Paths are relative to the extracted training shard root. The .idx file stores JSONL byte offsets and the .u32 file stores precomputed token lengths used by the training fast path.

Evaluation

The benchmark contains 50 examples for each of 16 domains. Model-output directories preserve both generated code and the final artifact shown to the visual judge. Raw rating files include six rubric dimensions, overall score, and rationale per sample.

Model

The corresponding best Base-init checkpoint is available at Ruler138/CodeAnything-Qwen3.5-9B-SFT-1.835M.

Code

Training, download, inference, rendering, and rating code is published at yuweiyang-anu/Code_Anything on branch release/sft-eval-history-20260910.

Access and use

This repository is public gated. Request access on the repository page, then authenticate with a Hugging Face user token. Users are responsible for reviewing the licenses and terms of upstream sources represented in the dataset before redistribution or commercial use.