maverickrzw/myopd

Vision-OPD v6 - 9,875-Row Research Dataset

An unchanged copy of the local vision-opd-diverse-v6-evidence-9875-20261031 research freeze, with its images and provenance. The original training table, image bytes, prompts, answers, row order and metadata are preserved under v6/. This dataset is intended for experiments in privileged-vision distillation: a Student and a Teacher can receive different visual inputs.

Keep this repository private. This upload is not a public redistribution grant, a semantic quality certification or a native Visual-CoT certification. The added context images have not been certified as useful evidence for each individual question. The underlying media rights remain under review.

Composition

Component	Rows
Original Vision-OPD	6,241
Retained earlier expansion	1,686
v5 expansion	948
New retrieved visual-context pool	1,000
Total, train split	9,875

There is no new validation/test split. The first 8,875 rows are identical to the previous v5 research release. The final 1,000 rows are identical to the separately frozen evidence pool; no original dataset or row was overwritten to make this merge.

Structure	Rows
Student receives at least two images	1,658
Student has image(s); Teacher has at least two extra images	1,378
Student has no image; Teacher has at least two images	60

These structure counts overlap and must not be summed as disjoint categories. The inherited 368 process-image candidates are not semantically certified. The new 1,000 context records are NOT an additional 1,000 process-image examples.

The Added 1,000 Records

Source	Selected split	Rows
TQA	train	400
SPIQA	train	579
SciVQA	train	21

Their content groups are life science 133, earth science 134, physical science 133, research plots 166, photographs/visualizations 145, schematics 145 and tables 144. These are heuristic content labels, not proof of benchmark coverage. School science is not college-level expertise; the research component is predominantly CS/ML and does not cover all MMMU subjects.

For each added record:

• Student retains one native question figure and the native question, options and primary caption where applicable.
• Teacher receives the same input, followed by two distinct native figures retrieved from the same lesson or paper and their native captions.
• The figures are retrieved visual contexts, not annotated gold support, not sequential thinking states and not native Visual-CoT.
• Native-source checks do not prove that Student has sufficient information, the answer is correct, or both additional figures are useful for that question.

Other retained sources include MathCanvas, Zebra, Molmo2 multi-image tasks, VisuLogic-Train and XVR, alongside earlier expansion sources. Precise versions, selection policies, native media roles and limitations are preserved in v6/provenance/; do not infer source-specific licensing from this card.

Files

README.md
upload_integrity.json              # Upload-time file sizes and SHA256 values
v6/
  train.parquet                    # Original canonical table, 9,875 rows
  media/                           # 23,363 referenced images, original bytes
  manifest.json
  DATA_CARD.md
  checksums.sha256
  artifact_hashes.json
  build_audit.json
  runtime_audit.json
  viewer_audit.json
  provenance/
    v5/                            # Frozen component metadata and receipts
    evidence_1k/                   # Including pending per-row review template
    ...                            # Merge/runtime/viewer source snapshots

The local-only absolute-path training copy, runtime caches, logs, credentials and environment variables are not uploaded. Original provenance contains local paths for traceability; these are not portable download URLs.

Canonical v6/train.parquet SHA256:

32e7a4097699687ad74f15165903808aa6e69898953005735a24c16b010c7cea

Schema And Image Handling

Important existing columns include source_id, prompt, images, teacher_prompt, teacher_images, reward_model and extra_info. The canonical Arrow schema is unchanged; inspect the Parquet for its exact types and any additional fields.

prompt and teacher_prompt contain chat messages with ordered <image> markers. Image references contain relative image/path values. Answers are in reward_model.ground_truth, with source/provenance details in extra_info. The images are external files, NOT embedded Arrow image bytes and NOT automatically cast Hugging Face Image features.

Resolve image paths relative to the downloaded v6/ directory. Calling load_dataset alone loads the table but does not resolve those image strings. Keep the original order and all entries of teacher_images; do not reconstruct or truncate them. In particular, original Vision-OPD records retain their native Student image and Teacher crop arrangement, which is not universally a Student-image prefix followed by extra images.

Download And Inspect

Authenticate with your own Hugging Face account/token through the usual Hub login or HF_TOKEN environment variable. Never put tokens in the dataset. Use a recent PyArrow; older versions had compatibility problems with these files.

from pathlib import Path

import pyarrow.parquet as pq
from huggingface_hub import snapshot_download

snapshot = snapshot_download(
    repo_id="maverickrzw/myopd",
    repo_type="dataset",
    token=True,
    revision="main",  # Pin the uploaded commit SHA for reproducible experiments.
    allow_patterns=["v6/**"],
)
root = Path(snapshot) / "v6"
table = pq.read_table(root / "train.parquet")
assert table.num_rows == 9875
sample = table.slice(0, 1).to_pylist()[0]
student_paths = [root / image["path"] for image in sample["images"]]
teacher_paths = [root / image["path"] for image in sample["teacher_images"]]

For the original OPD loader, make a separate absolute-path materialization. Do not edit the canonical Parquet:

import pyarrow as pa

rows = table.to_pylist()
for row in rows:
    for field in ("images", "teacher_images"):
        row[field] = [
            {"image": str(root / ref["path"]), "path": str(root / ref["path"])}
            for ref in row[field]
        ]
pq.write_table(
    pa.Table.from_pylist(rows, schema=table.schema),
    "train_absolute.parquet",
    compression="zstd",
)

Validation

• Exact schema and row-value preservation for both components; 9,875 unique IDs.
• All 23,363 referenced media copied, SHA256 checked and decoded during the freeze.
• Source QA, figure/caption membership and existing deterministic source-quality exclusions rechecked for all added 1,000 records.
• Production RLHFDataset reload of all 9,875 rows, without row drops/reordering.
• Fresh production preprocessing for 101 source/stride/extreme cases, zero failures. Unchanged component checks were reused with matching code and media hashes: this is NOT a fresh full 9,875-row tokenization claim.
• Existing Qwen3.5 template with Thinking disabled, Student budget 8,192 and Teacher plus a 1,024-token response within 10,240; retained Teacher images.
• Viewer count/filter tests, source receipts and frozen metadata hashes included.

Limitations And Benchmark Boundaries

semantic_certification=false and semantic_approved_rows=0 remain unchanged. Native-source fidelity is not proof of semantic correctness or useful evidence. The pending 1K review template is at v6/provenance/evidence_1k/semantic_review_template.csv.

Protected-data checks for the added pool use exact/RGB/pHash fingerprints, not a proof of semantic or source-family disjointness. Known original VStar overlap inherited from v5 remains. Independent benchmark evaluation data must not be reused for training. SPIQA's entire archive SHA was not verified locally; the selected member sizes/CRCs and hashes were checked.

V6 has not been trained or evaluated. Earlier v5 model scores do not apply to this dataset. No benchmark improvement, causal benefit or matched-ablation result is claimed for the added 1,000 records.

Rights And Access

This is a mixed-source research collection, not a new blanket license. Underlying figure/photo/textbook rights and attribution remain pending. Dataset-card licenses do not clear every source image. Review source terms and obtain any necessary permissions before use or redistribution.

The original freeze records public_redistribution_authorized=false and unrestricted_media_rights_certified=false. These flags are intentionally preserved. A private Hub upload does not change those flags or establish legal permission for any use. Do not make this repository public without resolving the applicable rights and quality requirements.