The dataset is currently empty. Upload or create new data files. Then, you will be able to explore them in the Dataset Viewer.
AutoEnvScaling — research artifacts
Supporting artifacts for the AutoEnvScaling study of agentic environment authoring: the environments three authoring arms produced, the episodes that produced them, and the runs used to evaluate them. Analysis code, figures and reports live in the companion repository simonucl/auto-envscaling; this dataset holds the material that is too large for Git.
Archived 2026-09-29 from the working storage, which is being decommissioned.
Contents
| Archive | Uncompressed | What it is |
|---|---|---|
authored-environments-20260929.tar.gz |
~1 GB | The corpus the paper rests on. 7,897 authored environment packages — instruction, Dockerfile, verifier tests, reference solution and task metadata for each — produced by eight author models across three authoring arms (prompt, prompt + feedback, harness). Text files only; binaries are excluded. Paths match the package column of analysis/harness_quality_20260918/records.csv in the repository, relative to the storage root. |
failure-study-e4-20260915.tar.gz |
1.3 GB | The raw workspace of the e4 failure study: 72 authoring episodes over 12 cases with two authors, their shipped environments, and the failing base-solver runs they were seeded from. A readable, redacted extract of the same study is in the repository under analysis/qualitative/. |
fig4-proposer-lanes.tar.gz |
~3.5 GB | Raw authoring episodes for the base vs distilled proposer comparison (Qwen3.6-35B-A3B). The per-episode gate reports and the analysis are in the repository; these are the underlying run outputs. |
hilbench-raw.tar.gz |
690 MB | HiL-Bench trajectories and run logs. Metrics, results and training records are in the repository under evals/hilbench/. |
Provenance and caveats
- Environments were admitted by a fixed gate sequence (structure, build, reference solution reaches
reward >= 1.0, no-op scores < 0.5). Admission status per package is in the repository's
records.csv, not in these archives. - The August cohorts' container images were lost with their registry; the repository documents how their base images were reconstructed from TMax recipes and re-validated.
- Archives contain agent transcripts and tool output. They were scanned and scrubbed for credential-shaped strings before upload; if you find anything that looks like a live secret, please open a discussion and it will be removed.
- Model weights are not included. The RL training checkpoints are ~510 GB per iteration in torch distributed-checkpoint format and were not published.
Citation
See the companion repository for the paper and the analysis that reads these artifacts.
- Downloads last month
- 41