You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

ElliotVL data pipeline: code and design notes

This repo holds the code, prompts and design documents that produced the ElliotVL datasets published on the Hub under elliot-mllm, Elliot-Data, VL-Data and VL-Data-2.

It is research code, kept as it was run. It is not a packaged library, and many scripts in pool_build_and_publish/ were written for one run, so expect to adapt data locations and job launchers to your own environment. Secrets were scanned for before upload and none were found.

How the data was built

 sources ──► normalize ──► merge + exact dedup ──► VLM-judge cleaning ──► materialize ──► SFT pool
 (download)  dataengine-mm-v1   merged_dedup_v4      Qwen-27B, v1think      sft_pool_v1      β”‚
                                                                                            β”œβ”€β–Ί rejection sampling (RS)
 self-labelled lanes (Bing crawl, Wikipedia) ── annotator VLM ── direct admit ───────────────
 human-labelled lanes (PixMo) ───────────────────────────────── direct admit ────────────────
 text-only lane (Nemotron subsets) ── sample + relabel ──────────────────────────────────────
                                                                                            β–Ό
                                                          sft_pool_v1.2 / v1.3 mix ──► publish to the Hub
 scanned-document lane (NARA, LOC, Internet Archive) ── download + page manifests ──► publish raw (unlabelled)
  1. Pick sources by benchmark coverage. Every candidate training set is mapped to the evaluation benchmarks it serves, with leakage notes. See docs/vlmevalkit_benchmark_training_sets.md and docs/data_pool.md.
  2. Normalize everything into one schema (dataengine-mm-v1): sharded parquet, conversational messages / qa, explicit media records, and provenance and dedup hashes. One row is one image (or image set) with all of its QA kept in order as a conversation. See DataEngine/README.md, DataEngine/docs/tutorial.md and DataEngine/scripts/standardize_*.py.
  3. Merge and deduplicate exactly by image SHA-256 across all sources (merged_dedup_v4). QA attached to the same image from different sources are merged, not dropped.
  4. Clean with a VLM judge. Qwen-27B reads each image with its QA and marks every pair correct, minor (rewritten) or wrong (dropped). It also appends new QA where the image supports more than the source asked. The full prompt is vlm_cleaning_qa_audit/prompts/v1think.txt. The client and fleet scripts are in vlm_cleaning_qa_audit/kimi_clean_*.py, the conversion back to training parquet is materialize_clean.py, and gen_registry.py writes the admission registry (docs/sft_v12_config/sft_pool_registry_2026-08-18_v2.md, with every cleaning step pinned to a version).
  5. Rejection sampling on 44 benchmark-style families: the model answers, a string verifier accepts or rejects, and accepted traces become *_RS_think / *_RS_nothink, with *_longform and *_rejected alongside. See DataEngine/scripts/rejection_sampling/ (README inside) and DataEngine/docs/sft_pool_v12.md Β§3 for the verifier bugs that were found and fixed.
  6. Self-labelled lanes. wordnet_annot_qa and eval_annot_qa are Bing image crawls, driven by WordNet concepts or by the topic of each evaluation benchmark respectively, labelled by Qwen-397B. wiki_caption and wiki_qa are Wikipedia / Commons images labelled by Kimi-K2.6 and Qwen-397B, split by image resolution. The prompts are in docs/sft_pool_v12/annotation_prompts/ and the crawl query tooling is pool_build_and_publish/scan_queries.py and probe_wncrawl.py. The Wikimedia download side is described in DataEngine/docs/wikipedia_wikimedia_pipeline.md. The one-off runners that drove these annotation fleets were not kept as standalone files, so the prompts and docs here are the record of how they were labelled. Nobody judged these rows.
  7. Direct admit. Human-labelled PixMo and GT-verified CoT sets are admitted with their original answers (vlm_cleaning_qa_audit/materialize_direct.py) so human labels are not second-guessed.
  8. Text-only lane. text6m_build/ samples and relabels NVIDIA Nemotron post-training subsets so the VLM keeps its text ability.
  9. Compose the mix. pool_build_and_publish/stage_v12.py, v13.py and write_mix.py stage families into groups A (RS), B (judge-cleaned), C (text) and D (self-labelled / direct), and write the Megatron-Energon blend (docs/sft_pool_v12/sft_pool_v12_mix.yaml). The reasoning and the bugs found along the way are in DataEngine/docs/sft_pool_v12.md.
  10. Scanned documents (raw, unlabelled). Public scans from NARA (Catalog digital objects and the 1940 Census), the Library of Congress and Internet Archive page-image zips are downloaded with resumable per-source state and turned into page-level manifests without unpacking the zips. See DataEngine/docs/scanned_doc_pipeline.md, DataEngine/reports/scanned_doc_manifest_v1.md and the run log DataEngine/reports/scanned_doc_download_run_20260702.md.
  11. Publish. DataEngine/scripts/pool_publish/ (README inside). Every repo is made public and gated with manual approval before any bytes go up. Image columns are declared datasets.Image so the viewer renders them, row groups are about 32 MB, and uploads are committed in resumable batches. scanned_docs_to_hub.py / scanned_docs_publish.py publish the scan pool across several orgs by shard range; publish_train_extra.py publishes the training families the first campaign missed.

Which code made which dataset

Hub repos Stages Start reading at
*_cleaned (single-image families) 1–4, 11 vlm_cleaning_qa_audit/, docs/sft_v12_config/sft_pool_registry_2026-08-18_v2.md
*_RS_think, *_RS_nothink, *_longform, *_rejected 1–5, 11 DataEngine/scripts/rejection_sampling/
wordnet_annot_qa, eval_annot_qa, wiki_caption, wiki_qa 6, 11 docs/sft_pool_v12/annotation_prompts/
pixmo_* 7, 11 vlm_cleaning_qa_audit/materialize_direct.py
nlvr2_multi_cleaned, mimic_cgd_multi_cleaned, mmra_multi_cleaned 1–4 (multi-image rebuild), 11 DataEngine/scripts/rejection_sampling/assemble_rs_splits_multi_image.py
text6m 8, 11 text6m_build/
scanned_docs_* 10, 11 DataEngine/docs/scanned_doc_pipeline.md, DataEngine/scripts/pool_publish/scanned_docs_to_hub.py

Layout

  • DataEngine/: the pipeline repo, covering download, standardization, relabel and annotation clients, rejection sampling, publishing, job launchers, docs and reports (markdown only; the galleries with sample images are not included).
  • vlm_cleaning_qa_audit/: the VLM-judge cleaning fleet, its prompts, materialization and registry generation.
  • pool_build_and_publish/: pool assembly, RS bookkeeping, the v1.2 / v1.3 staging and the first publishing campaign. Many of these are single-run scripts kept for provenance.
  • text6m_build/: the text-only lane.
  • docs/: the planning documents, which are the benchmark-to-training-set map, the data pool, the ElliotVL loop plan, the v1.2 pool docs and mix, and the pinned admission registry.

Access is gated with manual approval, like the datasets.

Downloads last month
250