Datasets:
DOI:
License:
File size: 5,567 Bytes
9b7bd53 0fa8531 82343f1 0fa8531 82343f1 0fa8531 9b7bd53 82343f1 0fa8531 9b7bd53 82343f1 9b7bd53 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 | ---
license: cc-by-4.0
pretty_name: AtlasOpenCoder
---
# AtlasOpenCoder — v1.0
Calibrated concept profiles for **OpenCoder-1.5B-Base**, read on 169 Python constructs across
61 pretraining and annealing checkpoints. This dataset is the battery, the licence tables and
the artefacts every figure and table of two papers reads:
* **"Calibrated Concept Profiles: Making the Reading of a Model's Neurons a Measurement"**
(Wilam, 2026) — the method. Its code: `piotrwilam/AtlasCalibratedConceptProfiles`.
* **"Representational Drift in a Code Model: Behaviour Settles, Which Neurons Carry It Does
Not"** (Wilam, 2026) — the finding, read with that instrument on this dataset.
Its code: `piotrwilam/AtlasRepresentationalDrift`.
`MANIFEST_PAPER1.md` and `MANIFEST_PAPER2.md` below are named for the first and the second of
those, in that order.
## What is here
| | |
|---|---|
| `phase_battery/` | the battery: 557,990 prompts in sets A, B and C, two twins and the nu arm |
| `battery/` | the earlier pilot battery |
| `validation/` | every artefact a figure or table in either paper reads |
| `inventory/`, `calibration/` | the 206 constructs and the corpus rate tables |
| `probe/companions.json` | the logged companion shares |
| `MANIFEST_PAPER1.md`, `MANIFEST_PAPER2.md` | every number in each paper, to its file, field and script |
**The battery cannot be regenerated.** Its per-prompt seeds came from `abs(hash(...))`, salted
per process, so the seeds recorded in `CONTENT_HASHES.json` do not reproduce the prompts they
name. That is why it is published.
## What is not here
**The raw record (2,022 GB over 61 references)** — per-core accumulators and per-prompt
activations. It is regenerable from OpenCoder's released checkpoints, this battery and the
code. One reference of it is planned for the next version.
**The pilot recording.** Two artefacts were once computed from a pilot tree and later found to
differ from the phase recording at every reference; everything derived from it is excluded and
named in the papers.
## Two batteries, two hashes
The **recorded** battery — the one the model was run on, and the one the papers' numbers come
from — carries prompt-content hash `570b42db…`, and its per-set hashes are the figures the
papers cite.
The **released** battery differs from it in **1,927 of 557,990 prompts (0.35%)**: those prompts
drew padding containing one of six third-party email addresses, which are replaced here with
`user@example.com`. The token offsets each prompt stores are recomputed to match, and the
recomputation is verified per prompt. See `phase_battery/REDACTION.json` and
`phase_battery/RELEASED_HASHES.json`.
Only a reader who re-runs the model on those 1,927 prompts sees a difference. Every artefact in
`validation/` was computed on the recorded battery.
## Licence
The dataset's own content (generator text, profiles, tables, artefacts) is **CC BY 4.0** (`LICENSE`); the padding snippets in the battery remain under their original licences, listed in `ATTRIBUTION.md`.
**The padding is not ours.** Padding blocks were drawn from `codeparrot/codeparrot-clean` and
carry the licences of the repositories they came from: MIT, Apache-2.0, BSD, GPL, AGPL, LGPL
and others. Roughly **29% of prompts** contain at least one such fragment, each one to three
lines and a median of 45 characters. Every block, its source repositories and their licences
are listed in `ATTRIBUTION.md`.
## Citing
Cite **the paper that owns this dataset, and the DOI together with the tag** — the method
paper, named above. The dataset is its instrument and its realisation on OpenCoder; the
second paper reads it.
> Wilam, P. (2026). *Calibrated Concept Profiles: Making the Reading of a Model's Neurons a
> Measurement.* Dataset: AtlasOpenCoder, tag `v1`.
**The DOI is shown on this page, above the file list, and belongs in the citation beside the
tag.** It is not repeated here: a DOI names the revision it was minted on, so a card cannot
carry its own without pointing at a later revision than the one the DOI names. The paper's
arXiv identifier is in the paper's own listing.
A Hugging Face DOI is assigned **to the revision it is generated on**, so a later version takes
a new DOI and the earlier one is marked outdated. Name the tag as well as the DOI: the tag says
which content a number came from in words a reader can check against this repository's history.
**Version 1 is the 20-cell yardstick** (4 token classes x 5 fifths of the prompt). A later
version computed under a different yardstick names different neurons for the same constructs,
so a figure from one version cannot be compared with a figure from another.
## What the directory names mean
`validation/paper1/` and `validation/paper2/` record **which paper's script wrote a file, not
which paper reads it**. Both papers read across them: the first paper's numbers come from 31
files under `validation/paper2/` and `validation/r2/`, and the second paper's from 14 under
`validation/paper1/`. `MANIFEST_PAPER1.md` and `MANIFEST_PAPER2.md` in this repository give, for every number
printed in each paper, the file and the field it is read from and the script that
rebuilds that file -- 407 numbers and 495, including the few that no released file holds,
which are marked and collected in a section of their own.
## Provenance
The model is `OpenCoder-LLM/OpenCoder-1.5B-Base-Checkpoints` (MIT). The reference corpus is
`codeparrot/codeparrot-clean`. The yardstick is the 20-cell split (4 token classes × 5 fifths
of the prompt), version 1.
|