README.md

# Dolma3 Data Attribution — Index

Entry point for the data attribution artifacts produced by the HCAI-Lab Dolma3 project. Use this dataset as the lookup table for "where do I find X?". All artifacts live under the `HCAI-Lab` org on Hugging Face or in the `soc127-dedup` Cloudflare R2 bucket.

If you only read one file: `inventory.json` has every artifact catalogued with location, scale, schema reference, and consumer use case.

## HF Collections (grouped views)

The HCAI-Lab org has 8 themed collections covering 41 dataset items. Pick the one closest to your use case:

| Collection | Items | Browse on HF |
|---|---|---|
| Dolma3 — Source Corpus + Manifest | 7 | `huggingface.co/collections/HCAI-Lab/dolma3-source-corpus-manifest-6a13f27b4ec80bdf68467093` |
| Dolma3 — Working Samples + Preconditioner | 6 | `huggingface.co/collections/HCAI-Lab/dolma3-working-samples-preconditioner-6a13f27ee2d04d7730f25849` |
| Dolma3 — Query Data | 3 | `huggingface.co/collections/HCAI-Lab/dolma3-query-data-6a13f281a8f0f4aacc1276b8` |
| TrackStar — Indices + Training Shards | 1 | `huggingface.co/collections/HCAI-Lab/trackstar-indices-training-shards-6a13f28219af7da31d666589` |
| TrackStar — Scores + Analysis | 2 | `huggingface.co/collections/HCAI-Lab/trackstar-scores-analysis-6a13f283e8a01cc6bb365928` |
| OLMES Evaluations | 7 | `huggingface.co/collections/HCAI-Lab/olmes-evaluations-6a13f2857c48958852e5a714` |
| Archive (pre-6T and legacy) | 14 | `huggingface.co/collections/HCAI-Lab/archive-pre-6t-and-legacy-6a13f288a645c069d72d8aae` |
| Other projects under HCAI-Lab | 1 | `huggingface.co/collections/HCAI-Lab/other-projects-under-hcai-lab-6a13f28d5a929010e1fd4874` |

The original master collection (`HCAI-Lab/dolma3-data-attribution-6a0fe8c1ae78751740458be4`) still exists and lists 22 items as a flat view; the sub-collections above are the recommended navigation path.

**Note**: HF Collections accept only datasets, models, spaces, papers, and other collections — **not buckets**. So the TrackStar gradient index (1.3 TB bucket), score matrices, query gradients, preconditioners, and figures are NOT in any collection. They're catalogued in `inventory.json` only. The "TrackStar" collections above contain only the dataset-addressable items in each family.

## Navigation by use case

### "I want the raw deduplicated 6T Dolma3 corpus"

Source shards on Cloudflare R2 at `soc127-dedup/soc127/phase{1_pool_shared,2_nonpool_final}/`. 58,621 `.jsonl.zst` shards, ~5 TB. Schema is `{id, text, metadata}` per line. See `docs/DATA_INVENTORY.md` §Source corpus.

### "I want stratified working samples for training or evaluation"

Pre-materialized samples on HF, six sizes:

- `HCAI-Lab/dolma3-6t-sample-500-docs` (288K docs, 539M tokens)
- `HCAI-Lab/dolma3-6t-sample-1000-docs` (575K docs, 1.1B tokens)
- `HCAI-Lab/dolma3-6t-sample-5000-docs` (2.86M docs, 5.3B tokens)
- `HCAI-Lab/dolma3-6t-sample-10000-docs` (5.68M docs, 10.5B tokens)
- `HCAI-Lab/dolma3-6t-sample-50000-docs` (26.2M docs, 62.8B tokens)
- `HCAI-Lab/dolma3-6t-sample-100000-docs` (49.7M docs, 118.4B tokens; 130/576 bins underfilled at this scale)

All stratified across 576 topic×format bins, seed=42. Manifest schema in `docs/WORKING_SAMPLE_DATA_ACCESS.md`. Each `dolma3-6t-sample-*-docs` repo exists both as an HF dataset (lookup above) and an HF bucket of the same name (S3-style access via `hf buckets sync`); bit-identical data on both surfaces.

### "I want WebOrganizer topic / format labels per document"

Three options, pick the one that fits your access pattern:

- `HCAI-Lab/dolma3-olmo3-corpus-manifest` — unified manifest, 1.1B rows, 32 columns (topic + format + quality + token count + source shard path). Single repo, the easiest entry point.
- `HCAI-Lab/soc91-labels` — HF mirror of the raw R2 sidecars: 2,719 parquet chunks (~60 MB each), 169.94 GB total. Each row carries `source_shard_path`. Use when you want only the topic/format columns and not the full manifest.
- R2 prefix `soc127-dedup/soc91-labels/` — original per-shard parquets (58,465 files) if you need per-source-shard granularity.

### "I want quality scores per document"

- `HCAI-Lab/soc139-quality-sidecars` — HF mirror, 1.26B rows, 41.8 GB, 80 parquet files. Columns: `doc_id`, `quality_label_id`, `quality_score`, `quality_high_prob`, `quality_low_prob`, `quality_confidence`, `source_shard_path`.
- R2 prefix `soc127-dedup/soc139-quality-sidecars/` — original per-shard parquets if needed.

Includes the SOC-142 label-fix markers — any local cache from before commit `3342baf` had inverted high/low quality labels and should be re-pulled.

### "I want attribution scores (influence) for the four OLMES benchmarks"

Per-query score matrices on HF Buckets, one bucket per run-and-model-variant:

- `HCAI-Lab/trackstar-scores-base-olmes-4bench` — OLMo-3-7B base, 4 benchmarks (gsm8k, mmlu_social_science, mmlu_stem, socialiqa), 2532 `.npy` files, ~396 GB
- `HCAI-Lab/trackstar-scores-instruct-cot-olmes-4bench` — same shape but for the instruct-cot variant
- `HCAI-Lab/trackstar-scores-{base,instruct-base}-bbh` — BBH attribution
- `HCAI-Lab/trackstar-scores-{base,instruct-base}-gsm8k-arc` — GSM8K + ARC attribution

Score matrix format: `shard_NNNN.npy` (float32, shape `[shard_docs, n_queries]`) + `shard_NNNN_doc_ids.json` (ordered positional doc IDs `shard_NNNN:INDEX`) + `query_ids.json` (ordered query IDs). See `docs/TRACKSTAR_DATA_ARTIFACTS.md` §1.

### "I want to resolve those positional doc IDs back to document text"

Required companion: `HCAI-Lab/dolma3-6t-sample-10000-docs-trackstar-shards` — 316 plain JSONL files mapping `shard_NNNN:INDEX` to `{id: <Dolma UUID>, text: <full document>}`. Without this you can't map score matrices back to source text.

### "I just want the top-K most influential docs per query — don't make me load 400 GB"

`HCAI-Lab/dolma3-trackstar-influence-scores` (private dataset) has:
- `influence_scores_full.parquet` — aggregated influence scores
- `top2k_{gsm8k,mmlu_socsci,mmlu_stem,socialiqa}.{csv,parquet}` — top-2K per query, rank-ordered

`HCAI-Lab/trackstar-top2k-{base,instruct-base}-gsm8k-arc` — same shape for the GSM8K + ARC benchmark set.

### "I want to score new query sets against the existing training corpus"

You need the Bergson training-gradient index: `HCAI-Lab/trackstar-gradient-index-base` (private bucket, 1.2 TB, 316 shard subdirs). Each shard has `gradients.bin` + `normalizers.pth` + `preconditioners*.pth` + configs. CPU-only scoring against new queries; no GPU rebuild required. ~156 GPU-hours to reproduce from scratch.

### "I want pre-built preconditioners for new attribution runs"

`HCAI-Lab/trackstar-preconditioners` — three subdirs (`olmo-3-1025-7b`, `olmo-3-7b-instruct`, `olmo-3-7b-think`), 78 files total, ~885 MB. Mixed preconditioners per SOC-152 / SOC-168. **The preconditioner must match the gradient-build model** (SOC-162 finding) — mixing base preconditioner with instruct gradients collapses benchmark-specific signal.

### "I want pre-built query gradient indices (instead of rebuilding from queries)"

`HCAI-Lab/trackstar-query-gradients-base` (private bucket, 17 GB). Three subdirs: `base/`, `instruct_base/`, `instruct_cot/`, each containing the four OLMES benchmark query gradient builds.

### "I want OLMES evaluation results / per-query model predictions"

Datasets under `HCAI-Lab/olmes-eval-olmo3-7b-{base,instruct-base,instruct-cot,thinking,think-mc,think-cot}` plus the SOC-166 bucket `HCAI-Lab/trackstar-olmes-eval-artifacts`.

### "I want dedup state for reproducing the corpus"

Bloom filter at `HCAI-Lab/dolma3-6t-bloom-index`, doc IDs at `HCAI-Lab/archive-dolma3-6t-doc-ids-pershard` (private, gzip-compressed JSONL per shard), unique docs materialized at `HCAI-Lab/dolma3-6t-unique` (1.258B docs).

## Schema references (where to find canonical definitions)

| Schema | Where it's defined |
|---|---|
| Source shard JSONL (id, text, metadata) | `docs/DATA_INVENTORY.md` §Document format |
| Working sample manifest (6 cols: doc_id/token_count/shard_path/bin_id/bin_topic/bin_format) | `docs/WORKING_SAMPLE_DATA_ACCESS.md` §Manifest schema |
| Unified corpus manifest (32-col PyArrow) | `src/data_attribution/recipes/corpus_manifest.py:15-52` |
| Sidecar row construction (joins source + WebOrganizer + quality) | `src/dolma/sidecar_manifest_fields.py:43-124` |
| Score matrix format | `docs/TRACKSTAR_DATA_ARTIFACTS.md` §1 |
| Bergson gradient index layout (per shard) | `docs/TRACKSTAR_DATA_ARTIFACTS.md` §5 |
| Bin ID formula (1-576) | `topic_idx * len(FORMATS) + format_idx + 1` — `src/dolma/manifest_fields.py:60-67` |

## Access

- **Code**: `github.com/eilab-gt/social-data-attribution`. CLI entry points in `pyproject.toml [project.scripts]`.
- **HF**: set `HF_TOKEN` in env, or store at `~/.hf_token`. For private repos in HCAI-Lab, request collaborator invite from the lab admin.
- **R2**: read-only token issued via 1Password Share. Set `R2_ACCESS_KEY_ID` and `R2_SECRET_ACCESS_KEY`. Repo provides `scripts/bootstrap/with_r2_credentials.sh` as a wrapper.
- **PACE**: not available to external consumers.

## Known gotchas

- Set `HF_HUB_DISABLE_XET=1` and point `HF_HOME` at local storage if running uploads on PACE NFS / Lustre. Xet finalization hangs there.
- SOC-142: an earlier batch of `soc91-labels` sidecars had inverted high/low quality labels. The fix is in commit `3342baf`. Refresh any local cache from before that.
- Underfilled bins in stratified samples (e.g., 17 underfilled bins in `sample_10000_docs`) are real, not a bug — the source corpus genuinely has fewer documents in that topic×format combination than requested.
- Don't mix instruct query files into a `base` query directory before launching attribution scoring — the reduce job will silently consume them and produce nonsense (see `docs/ATTRIBUTION_RUNBOOK.md`).

## Lifecycle

This inventory was first built 2026-05-22 as part of the external-team handoff migration. Source preservation report: `PRESERVATION_GAP_REPORT.md` in the source repo. Regenerated when artifacts change.

## Contact

HCAI-Lab data attribution work. Contact Glenn Matlin via the lab Slack channel.