README.md

metadata

license: mit
language: en
tags:
  - scientific-discovery
  - research-ideas

Alien Space of Science data

Complete downloadable artifacts for the Alien code repository. These are ordinary archives and JSONL/NumPy files, not a hosted application.

The matching checkpoints are in alien-models.

Downloads

• papers.tar.gz: 16,068 methodological summaries (blog.md) and their extracted/refined atom JSON files; original clustering and the 80%-coverage atom assignments. Extracts to papers/. Original PDFs are referenced by metadata, not redistributed.
• data.tar.gz: SQLite paper and author metadata, original token vocabulary, and coherence train/validation splits. Extracts to data/.
• availability-compact.tar.gz: availability training data. Extracts to data/availability_dataset/; approximately 1.8 GB compressed and 4.3 GB extracted.
• contrastive_availability_train.jsonl.gz and contrastive_availability_val.jsonl.gz: optional exchange-format examples, approximately 3.75 GB and 0.42 GB compressed. The metadata and token mapping are provided alongside them.
• research-assets.tar.gz: fixed atom combinations and Markdown proposal texts from Alien (300), Random (300), Claude Opus 4.7 (300), and Gemini 3.1 Pro (300). Each method folder contains a combinations.json index linking atom IDs to proposal files.

Download only the files needed for your workflow. The code README provides extraction and training commands. The compact representation is recommended for retraining.

hf download alejandrohdez/alien-data papers.tar.gz data.tar.gz --repo-type dataset --local-dir downloads
tar -xzf downloads/papers.tar.gz
tar -xzf downloads/data.tar.gz

Counts and structure

The corpus contains 16,068 papers; 16,020 have mapped atoms. Coherence includes ten serialized variants for each of 15,203 papers with at least two atoms: 136,827 training and 15,203 validation examples.

Availability uses 46,329 author atom profiles, with 44,809 profiles having at least two atoms. Raw JSONL contains 43,098,876 training and 4,788,764 validation records. Filtering empty complements leaves 43,077,647 usable training and 4,786,404 validation records. The training/export bank includes 39,791 authors. The split is example-level, not author-disjoint.

Availability records pair a query_ids token sequence with the disjoint complement token sequence in author_ids; author_id is the researcher identifier. The compact loader and binary record layout are supplied in generation/availability_data.py. author_ids.json explicitly defines profile row order and must be retained.

Provenance and contact identifiers

Paper metadata and author links originate from the authors' OpenReview/DBLP-based corpus; methodological summaries and atoms were generated by the LLM pipeline. Paper source URLs are provided in the SQLite database. The atom catalogue uses 80%-coverage clustering.

Email columns are omitted. The 1,208 email-based author IDs in the metadata are consistently replaced across public profiles, author links, training examples, and model bank metadata. Author display names and non-email service identifiers remain; this is not a fully anonymous dataset. Token sequences, training split membership, and model-bank row alignment are unchanged.

File sizes and SHA-256 checksums are listed in manifest.json. The authors' contributions, including metadata, summaries, atoms, training exports, and proposals, are released under the MIT license; see LICENSE. Original paper content and other third-party material retain their source terms.