# MIREL training data The release contains processed parquet files, not the original image archives. Image references in the parquet files are repository-relative and contain no machine-specific usernames or absolute internal paths. Put the downloaded image archives under `final/data/images/` (or rewrite the paths as described below) before training. ## SFT data `SFT/verified_cot.parquet` contains 2,698 unique, verified structured-CoT examples. It is deduplicated by `task_uid` from the historical 5,396-row file; no example is intentionally repeated in this release. The source tasks are: - `M4-MIT-States_PropertyCoherence`: 1,054 rows; - `M4-MIT-States_StateCoherence`: 594 rows; - `Mantis-dreamsim`: 638 rows; - `Mantis-imagecode`: 412 rows. The messages include the system prompt used for CoT-SFT and set `enable_thinking=false`. ## PG-OPD data The two model directories intentionally contain model-specific data; they are not interchangeable. ### Qwen3.5-4B `Qwen35-4B/pg_opd.parquet` contains 2,394 rows: `Qwen35-4B/pg_opd_8k.parquet` contains 8,000 rows for the data-scaling ablation (4,000 verified privilege rows and 4,000 native-control rows). This is an auxiliary scaling artifact. - 1,197 verified evidence-privileged rows; - 1,197 native fallback rows; - evidence was selected from the 4B Stage-A filtering run. ### Qwen3.5-9B `Qwen35-9B/pg_opd.parquet` contains 3,000 rows: - 1,500 verified privilege rows generated and blind-verified with the 9B SFT-LoRA checkpoint; - 1,500 native-control rows matched from the same P1 pool; - source Stage-A pool: 8,000 P1 samples, with 1,735 verified privilege rows available before the deterministic 1,500-row selection. For both recipes, privilege text contains evidence content only: no GT answer, answer label, or `support=[...]` locator is passed to the privileged teacher. Rows retain the native prompt/images and the teacher-side evidence fields. The teacher advantage is computed online by comparing fixed native and privileged teachers on the same student response tokens. There is no GRPO answer reward in PG-OPD. ## Original data to download ### M4-Instruct Download the M4-Instruct release described by the official M4-Instruct data card and place the following archives under one local `M4-Instruct-Data` directory: - `MIT-States_PropertyCoherence.zip` - `MIT-States_StateCoherence.zip` - `COMICS_Dialogue.zip` (needed by the P1 objective-data construction) - `DocVQA.zip` - `OCR-VQA.zip` The M4 data card and license are included in the upstream M4-Instruct release; the project copy records CC BY 4.0. The processed SFT file uses the two MIT-States subsets. The historical P1 objective construction additionally uses the COMICS, DocVQA and OCR-VQA subsets. ### Mantis-Instruct Download the upstream **TIGER-Lab/Mantis-Instruct** release and its image archives. The processed files use these Mantis subsets: - `Mantis-dreamsim` for verified SFT; - `Mantis-imagecode` for verified SFT and objective PG/GRPO construction. Use the upstream dataset's `script` revision when downloading image archives, as its dataset card explains that the images are provided separately. Follow the upstream license and citation requirements. Official references: - https://github.com/TIGER-AI-Lab/Mantis - https://huggingface.co/datasets/TIGER-Lab/Mantis-Instruct - https://llava-vl.github.io/blog/2024-06-16-llava-next-interleave/ (M4-Instruct) ## Image layout and path rewriting The sanitized parquet paths use this layout relative to the `PG-OPD` working directory: ```text ../../data/images/m4/... ../../data/images/mantis/... ``` Place the M4 and Mantis image archives below `final/data/images/m4/` and `final/data/images/mantis/`, preserving the suffixes recorded in the parquet files. Alternatively, use `code/PG-OPD/scripts/rewrite_image_paths.py` to map `../../data/images` to another local image root. The helper changes only image path fields; row order and textual fields remain unchanged. Run its `--check-only` mode before training.