juiceb0xc0de/HRM-Text-1B-Blind-Atlas-Puzzle

Why this dataset is difficult

This dataset is a blind mechanistic-interpretability puzzle built from an anonymized atlas of sapientinc/HRM-Text-1B.

The challenge is not just to inspect the SQLite database. The challenge is to avoid over-interpreting it.

The anonymized database preserves the atlas structure, feature statistics, per-head measurements, coactivation records, Sub-Zero- style measurements, and recurrent slot layout, but removes the obvious identity anchors: model names, readable table names, readable column names, component labels, and semantic labels are replaced with anonymous identifiers.

This is especially deceptive because HRM-Text-1B is not a standard decoder-only Transformer. It is a ~1B parameter dual-timescale recurrent architecture with high-level and low-level Transformer stacks. In the atlas, this can appear like a much deeper model than it really is. A model that assumes every slot corresponds to an ordinary physical layer may confidently infer a 100B+ or 100+ layer model, even though that is the wrong explanation.

The dataset is meant to test whether a model can:

• reason from partial structural evidence
• notice recurrent / repeated-compute patterns
• resist shortcuts from layer count alone
• distinguish “I can infer this” from “I am guessing”
• avoid hallucinating precise architecture details from anonymized schema

A good solver should be willing to say that the database appears transformer-like but anonymized, that the layer structure is suspicious, and that recurrence or repeated module use is a likely explanation. A bad solver will confidently invent the source model, parameter count, training setup, or architecture family from incomplete evidence.