elemental-tasks/paper-artifacts

ElementalTask artifacts, v0.1

Task snapshots, observed evaluation trajectories, and function vectors for What Do Language Models Learn and When? The Implicit Curriculum Hypothesis. Use the BibTeX entry in CITATION.bib when citing the paper.

Some third-party task materials are omitted from this release pending redistribution permission. See the coverage manifest for included tasks.

This release contains 88 tasks, 19,863 reference examples, 38,889 observed accuracy measurements across 10 models, and 806 function vectors.

Download the Hugging Face snapshot or the GitHub release. Individual ZIPs are listed in DOWNLOADS.md.

Files

• tasks-v0.1.zip: frozen reference inputs, complete prompts, saved targets, per-task provenance, and current scoring source for reference.
• trajectories-v0.1.zip: observed accuracy CSV and a checkpoint/token table.
• fvs-<model>-v0.1.zip: one archive per model, with unchanged NumPy arrays, original extraction metadata, and an index with checksums.
• manifest.json: artifact counts, archive hashes, and hashes of source inputs.
• task_membership.csv: appendix, fixed-26, and expanded FV-analysis membership, with release availability kept separate.
• coverage.csv: task/model coverage, unavailable vectors, and actual FV splits.
• SHA256SUMS: hashes of the ZIPs, metadata, and unpacked files.
• olmo13b_replay.json: verification of the expanded OLMo-2 13B experiment and comparisons with the older predictions under different loading variants.

The appendix suite has 86 explicitly listed tasks; the expanded analysis adds 21 compositions, giving 107 task IDs across the analysis scopes. The manifest records the included and withheld subsets. This release does not redefine the paper's task count or claim that every analysis used the same tasks.

Loading and verification

Python 3.10+ and NumPy are sufficient; model weights and GPU access are unnecessary.

To download the complete versioned snapshot using huggingface_hub:

from huggingface_hub import snapshot_download
snapshot_download("elemental-tasks/paper-artifacts", repo_type="dataset",
                  revision="v0.1", local_dir="elementaltask-v0.1")

Then run the verifier from the downloaded directory:

python load_artifacts.py --root . --verify

To load an archive directly:

import csv
import io
import json
import zipfile
import numpy as np

with zipfile.ZipFile("tasks-v0.1.zip") as archive:
    examples = [json.loads(line) for line in
                archive.read("examples/basic_arithmetic.jsonl").splitlines()]
    prompt, target = examples[0]["prompt"], examples[0]["target"]

with zipfile.ZipFile("trajectories-v0.1.zip") as archive:
    rows = list(csv.DictReader(io.StringIO(archive.read("trajectories.csv").decode())))
    # Keep missing values missing; zero is a measured accuracy.
    observed = [r for r in rows if r["status"] == "observed"]

with zipfile.ZipFile("fvs-olmo2_13b-v0.1.zip") as archive:
    index = list(csv.DictReader(io.StringIO(archive.read("index.csv").decode())))
    vector = np.load(io.BytesIO(archive.read(index[0]["path"])), allow_pickle=False)

Provenance and limits

Task examples come from one archived evaluation run per task, preferring a final checkpoint. Complete prompts preserve the demonstrations actually used in that run. This snapshot does not cover all prompt variants across all model/checkpoint runs. Per-task metadata flags example counts that differ from the appendix. Current scoring sources are included for inspection; their historical revisions were not recorded, and they are not a standalone evaluation package.

Trajectories preserve saved pivot values, including missingness. They are not smoothed, rounded into token buckets, or reconstructed from raw responses. Pythia main aliases step143000, with 299,892,736,000 training tokens; preserve the raw labels but avoid counting both as independent checkpoints. Other token positions are labeled estimates or unknown. Immutable model revision hashes were not recovered. Historical paper analyses also used metrics JSON and other preprocessing; the pivot export alone does not reproduce that preprocessing.

FV arrays are copied without renormalizing or changing their split. In the inspected extractor, residual_stream means full block output, not selected attention-head outputs. OLMo-2 13B's saved metadata differs from the paper's head-selection entry. Some OLMo-2 1B vectors use incorrect or unsplit examples. These facts are recorded in the indexes. Withheld tasks also have their derived artifacts excluded from this release, so the bundles do not contain the full FV prediction basis used in the paper.

The local layer-10 residual vectors reproduce all 44 OLMo-2 13B task scores in the expanded camera-ready baseline experiment to numerical precision. This check uses the complete local prediction basis, including tasks withheld here. The older 26-task predictions are not exactly reproduced: three later basis tasks and changes to trajectory loading explain much of the difference. With the 95-task basis restored and pivot-only loading, 25 targets differ by 0.022 percentage points on average (maximum 0.141); the remaining target's pivot has 15 checkpoints versus 51 in its saved result. See olmo13b_replay.json for the separate checks and their limits. This verifies the downstream FV set without establishing why it differs from the paper's head-selection entry. The ignoring_context reference snapshot contains one example, while the appendix lists five; the release preserves the archived record and flags the difference.

Licenses

Original project data and artifact documentation: CC BY 4.0; see LICENSE-DATA.txt. Project code: MIT; see LICENSE-CODE.txt. IOI examples retain the upstream MIT license; see THIRD_PARTY_NOTICES.md. Existing permissions for previously distributed versions remain available.

Rebuilding from the research checkout

The builder reads only the reviewed task allowlist in release/spec.json and checks the inventory hashes before exporting. It requires the archived source results referenced by that spec. Run from the repository root:

python scripts/build_release_artifacts.py --output artifacts/v0.1
python release/load_artifacts.py --root artifacts/v0.1 --verify

The output directory must not already exist. Use a new path for a rebuild.