English POLAR experiment runs — Team Ugh
Public experiment archives for Structure over Scale: Hierarchy-Constrained Modeling for English Online Polarization Detection.
This release stores completed cross-validation results, predictions by ID, metrics, configurations, logs, analyses and provenance. The completed CV runs were trained without retaining checkpoints. These archives do not recover those discarded model weights. New final-fit checkpoints can be archived separately using the same lossless format when trained.
Studies
The machine-readable archive index records each verified
archive's exact Hugging Face commit and path. Each study is stored under
studies/<study>/<content-fingerprint>/ with artifacts.tar.gz and a public
manifest.json listing every included file, its size and SHA-256 checksum.
ordered_v2: 125 trained decoder-extension, parent-to-child M5 and retained-backbone head runs.loss_controls_v1: 125 new supervision-scope/class-weight/lambda controls, one baseline shadow, and reused baseline results.robustness_v1: independent metric audit, statistical sensitivity, score-only controls and threshold stability.rq3_v1: the separate, earlier 75-run fixed-M4 head study.non_rq4_completion: report tables, review codebook, provenance and RQ4 handoff.
These are distinct studies, not one pooled experiment. OOF scores are training-set cross-validation estimates, not official-test or leaderboard performance. The recorded selected configuration is pooled M5-soft with weighted conditional TYPE loss and lambda=1. Its mean OOF DET macro-F1 is 0.7937 and all-text TYPE macro-F1 is 0.3999; superiority over M4 was not established.
Archive integrity and use
Archives retain original file bytes and relative directory layout. Uploads are followed by a full, commit-pinned download, archive checksum comparison, and verification of every included member. The index includes only verified uploads. Token/credential/cache files and raw-text CSV/JSONL review files are excluded, with reasons recorded in the manifest. The public filtered archive is not a byte-for-byte copy of every local study directory; original completion markers may reference deliberately excluded files.
Download the listed archive at its pinned revision and unpack it to a fresh
directory after verifying its SHA-256. The project's
analysis_tools.hub_archive restore command uses a local upload receipt to
perform these checks and restore the original layout. Future retained FP32
checkpoints keep checkpoints/best/training_state.pt, tokenizer, configuration
and threshold metadata, compatible with the project's checkpoint prediction
CLI. These archives are project artifacts, not a standard Transformers
from_pretrained model export.
Scientific source snapshots and exact experiment configuration/protocol files are included where recorded. RQ4 external evaluation remains separate from this storage release. The qualitative review is assistant-assisted, not independently human-verified annotation. Refer to each study's result files for uncertainty and limitations rather than selecting the highest inspected mean across unrelated protocols.
Verified archive links
| Study | Included files | Commit-pinned archive directory |
|---|---|---|
| ordered_v2 | 3,375 | Files and checksum manifest |
| loss_controls_v1 | 3,602 | Files and checksum manifest |
| robustness_v1 | 15 | Files and checksum manifest |
| rq3_v1 | 1,124 | Files and checksum manifest |
| non_rq4_completion | 13 | Files and checksum manifest |
Storage utility
The storage tool source and project storage instructions are included for inspecting the archival/verification/restore workflow. The training implementation is supplied in the studies' recorded source snapshots.