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 βββΊ normalize βββΊ merge + exact dedup βββΊ VLM-judge cleaning βββΊ materialize βββΊ SFT pool
(download) dataengine-mm-v1 merged_dedup_v4 Qwen-27B, v1think sft_pool_v1 β
βββΊ rejection sampling (RS)
self-labelled lanes (Bing crawl, Wikipedia) ββ annotator VLM ββ direct admit βββββββββββββββ€
human-labelled lanes (PixMo) βββββββββββββββββββββββββββββββββ direct admit ββββββββββββββββ€
text-only lane (Nemotron subsets) ββ sample + relabel ββββββββββββββββββββββββββββββββββββββ€
βΌ
sft_pool_v1.2 / v1.3 mix βββΊ publish to the Hub
scanned-document lane (NARA, LOC, Internet Archive) ββ download + page manifests βββΊ publish raw (unlabelled)
- 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.mdanddocs/data_pool.md. - Normalize everything into one schema (
dataengine-mm-v1): sharded parquet, conversationalmessages/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. SeeDataEngine/README.md,DataEngine/docs/tutorial.mdandDataEngine/scripts/standardize_*.py. - 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. - Clean with a VLM judge. Qwen-27B reads each image with its QA and marks every pair
correct,minor(rewritten) orwrong(dropped). It also appends new QA where the image supports more than the source asked. The full prompt isvlm_cleaning_qa_audit/prompts/v1think.txt. The client and fleet scripts are invlm_cleaning_qa_audit/kimi_clean_*.py, the conversion back to training parquet ismaterialize_clean.py, andgen_registry.pywrites the admission registry (docs/sft_v12_config/sft_pool_registry_2026-08-18_v2.md, with every cleaning step pinned to a version). - 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*_longformand*_rejectedalongside. SeeDataEngine/scripts/rejection_sampling/(README inside) andDataEngine/docs/sft_pool_v12.mdΒ§3 for the verifier bugs that were found and fixed. - Self-labelled lanes.
wordnet_annot_qaandeval_annot_qaare Bing image crawls, driven by WordNet concepts or by the topic of each evaluation benchmark respectively, labelled by Qwen-397B.wiki_captionandwiki_qaare Wikipedia / Commons images labelled by Kimi-K2.6 and Qwen-397B, split by image resolution. The prompts are indocs/sft_pool_v12/annotation_prompts/and the crawl query tooling ispool_build_and_publish/scan_queries.pyandprobe_wncrawl.py. The Wikimedia download side is described inDataEngine/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. - 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. - Text-only lane.
text6m_build/samples and relabels NVIDIA Nemotron post-training subsets so the VLM keeps its text ability. - Compose the mix.
pool_build_and_publish/stage_v12.py,v13.pyandwrite_mix.pystage 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 inDataEngine/docs/sft_pool_v12.md. - 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.mdand the run logDataEngine/reports/scanned_doc_download_run_20260702.md. - 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 declareddatasets.Imageso 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.pypublish the scan pool across several orgs by shard range;publish_train_extra.pypublishes the training families the first campaign missed.
Which code made which dataset
| Hub repos | Stages | Start reading at |
|---|---|---|
*_cleaned (single-image families) |
1β4, 11 | vlm_cleaning_qa_audit/, docs/sft_v12_config/sft_pool_registry_2026-08-18_v2.md |
*_RS_think, *_RS_nothink, *_longform, *_rejected |
1β5, 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 |
1β4 (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