Blame for README.md

---
license: other
language:
  - en
task_categories:
  - text-generation
tags:
  - memory-layers
  - interpretability
  - synthetic
  - the-pile
pretty_name: memlayers data v1
size_categories:
  - 100M<n<1B
---

# `memlayers-data` v1

The frozen training bundle for a study of **what memory-layer values store and
how they are organized** (PKM / PEER against a kNN-LM store), COMS E6998
Algorithms in LLMs. Code that builds, verifies and reads it:
[RoieGilad7/memory-layers-analysis](https://github.com/RoieGilad7/memory-layers-analysis),
tag `data-v1`.

**One bundle, three parts.**

1. **A 150M-token training slice of the Pile** — `monology/pile-uncopyrighted`
   at revision `3be90335b66f24456a5d6659d9c8d208c0357119`, five subjects at
   **exactly 27,000,000 GPT-2 tokens each**: `Pile-CC`, `Github`,
   `PubMed Central`, `FreeLaw`, `Wikipedia (en)`.
2. **Six held-out subjects, each split into a `label` half and a `test` half** —
   the five above plus **`USPTO Backgrounds`, which never appears in training at
   all**, so it measures a subject the model has never seen. Every half is
   within 1% of 2,500,000 tokens per subject; the halves share no documents and
   no document ids with the training split.
3. **Synthetic biographies of 2,500 fictional people** — five facts each (birth
   date, birth city, university, major, employer) drawn uniformly at random from
   fixed value lists, so a model can only answer a question about one by having
   *stored* it. 2,000 people appear in training at one of three exposure tiers
   (frequent / normal / rare) and 500 are **controls that never appear**, which
   is the guessing floor. 8,000 distinct biographies, 190,000 copies,
   15,142,237 tokens (10.06% of `train.bin`), and 100,000 fact-recall prompts.

`train.bin` is **150,475,264 GPT-2 tokens / 293,897 sequences of 512**.

## Licensing and provenance

- **Upstream license tag: `other`.** This bundle is a derived slice of
  [`monology/pile-uncopyrighted`](https://huggingface.co/datasets/monology/pile-uncopyrighted),
  which is EleutherAI's **The Pile** with the `Books3`, `BookCorpus2`,
  `OpenSubtitles`, `YTSubtitles` and `OWT2` subsets removed — per that dataset's
  card, the subsets "not explicitly allowed to be used in AI training" according
  to the original Pile paper. It was published "in response to authors demanding
  that LLMs stop using their works".
- **Component documents keep their own original licenses.** The Pile is a
  compilation of many sources; nothing here relicenses any of them. Consult the
  upstream dataset card and the Pile paper before redistributing or using the
  Pile-derived files for anything beyond research reproduction.
- Pile text here is **cleaned, deduplicated, split and tokenized**, not
  reproduced verbatim as published: see `reports/` in the bundle and the data
  reports in the code repository for exactly what was done.
- **The biographies are synthetic.** Every person, name and fact is generated by
  `data/bios.py` from a seed. They describe nobody. Candidate names are rejected
  if they occur in the Pile text or if a name word would hand a probe one of its
  answers.
- No part of this bundle is intended for training a model that is deployed; it
  exists so one experiment can be reproduced token for token.

## Files

| path | files | size | what |
|---|---:|---:|---|
| `tokens/` | 2 | 304.1 MB | `train.bin` (uint16 GPT-2 ids) + `train.idx.parquet`, the segment index |
| `heldout/` | 16 | 103.7 MB | `label.bin`, `test.bin`, their indexes, and the per-subject cleaned documents of each half |
| `train/` | 5 | 227.2 MB | the cleaned, deduplicated training documents, one `.jsonl.gz` per subject |
| `heldout_pool/` | 6 | 59.3 MB | the pool each held-out half was drawn from |
| `bios/` | 8 | 58.7 MB | `people.jsonl`, `bios_train.jsonl`, `bios_control.jsonl`, `facts.jsonl` (100,000 prompts), the value lists, the templates, the name check |
| `samples/` | 5 | 0.6 MB | the hand-check pack: Pile samples, biographies with every prompt, the value lists, and **`templates_rendered.md`** — all 120 sentence frames rendered twice |
| `reports/` | 5 | 0.03 MB | what each build stage did: extraction, dedup, split, pack |
| `raw/` | 12 | 398.0 MB | the archived source text a rebuild needs. **Optional** — `data.fetch` skips it unless you ask |
| `manifest.json`, `validation_report.{json,md}` | 3 | 0.05 MB | the file list with a sha256 per file, and the A1–A18 results |

**59 hashed files, 1,151,617,170 bytes** (753,606,445 without `raw/`).

## How to verify your copy

The code repository carries the manifest, so verification does not trust this
repo's copy of it:

```bash
git clone https://github.com/RoieGilad7/memory-layers-analysis
cd memory-layers-analysis
git checkout data-v1
pip install -r requirements.txt

# download (skips raw/ by default; --with-raw includes it) and check every sha256
python -m data.fetch --version v1

# or check a copy you already have, without downloading
python -m data.fetch --version v1 --verify-only
```

`data.fetch` hashes every file against `data/manifests/v1.json` and exits
non-zero on any mismatch. The bundle's identity is its **`data_sha256`**, taken
over the manifest's file list alone:

```
train.bin    sha256 7e3d50f2513d5a1b42c4fbf201ad4811e81e3175f41068277cf708f191bad8ac
bundle  data_sha256 bb1bfd407af2c44e86d020df1c075f8764c5bf6979be2835684f6f0453086fd6
```

Use the `data-v1` tag, here and in the code repository, to get this exact
bundle.

## Checks

`python -m data.validate --config configs/data/v1.yaml` runs 18 checks on the
real build, all of which pass for this bundle: byte-identical rebuild, every
file's hash, exact per-subject token counts, subject shares inside every 1M
window, format and index tiling, decode-back-to-source, train/held-out leakage
(ids, exact hashes, 13-gram overlap), the halves' disjointness, no held-out-only
subject in training, subject labels against the source, a bag-of-words subject
classifier, the biography facts and their uniformity, **no answer leaks in any
of the 100,000 prompts**, exposure counts and control isolation, the probe floor
for a random scorer and an untrained model, the loader's order and resume, a
per-subject unigram baseline, and a 2-hour learnability run.

The learnability run is the one result worth quoting here: after 5,830 steps on
this data, a small PKM model recalls the `frequent`-tier facts **40.5%** of the
time as the top-1 answer against a **0.63%** control floor — 64× — and recall is
monotone across the exposure tiers. A *structurally new* wording of the same
question recovers the fact only ~2% of the time, which is the finding the study
is about.

## Citation

```bibtex
@misc{memlayers_data_v1,
  title  = {memlayers-data v1: a controlled slice of the Pile with synthetic biographies},
  author = {Roie Gilad},
  year   = {2026},
  note   = {COMS E6998 Algorithms in LLMs final project. data\_sha256
            bb1bfd407af2c44e86d020df1c075f8764c5bf6979be2835684f6f0453086fd6},
  url    = {https://huggingface.co/datasets/roieg/memlayers-data}
}
```