base-model-evals/off-the-shelf-model-evals

Babel Tower benchmark archive

Filesystem toolkit 1.3.0 provides versioned storage and validation for the proposed multilingual difficulty-calibrated benchmark. It connects a question bank, evaluation protocols, model-family panels, item-level responses, external calibration results, and frozen benchmark releases.

Current data status: no real evaluation runs, complete foundation item bank, or fitted calibration parameters have been published here. The toolkit includes separate synthetic examples. Uploading the filesystem does not run experiments, fit IRT models, or establish that the proposed benchmark works.

Download the updated toolkit and extract it. Open START_HERE.md, then follow the benchmark workflow. Python 3.10+ on macOS/Linux supports local operations without extra packages. Hugging Face transfer and Parquet export are optional dependencies.

Before: results organized by experiment run

The earlier structure preserved raw results and normalized item responses. Bank, protocol, model-panel and calibration definitions had no first-class validation.

After: a versioned benchmark archive

The diagram shows the supported layout. Empty folders are created locally by init; actual artifact and run directories appear on the Hub when registered content is uploaded. See the full diagram explanation.

Collection	Purpose
item-banks/<id>/	Shared question identities, translated items, language/script, choices and keys, source revisions, passages/overlap groups and exclusions.
protocols/<id>/	Task prompts, demonstrations, candidate text/label scoring, normalization and harness configuration.
model-panels/<id>/	Checkpoint revisions, base-model declarations, family identities and calibration/development/validation splits.
runs/<id>/	Original evaluation files and checked item responses tied to exact bank, protocol and checkpoint references.
calibrations/<id>/	Imported IRT parameters, scale, uncertainty fields, diagnostics, validation evidence and hash-pinned input runs.
releases/<id>/	A frozen compatible bank/protocol/calibration combination with declared weights and limitations.

Every new registered artifact contains an artifact.json payload and manifest.json integrity record. Existing IDs cannot be overwritten through the toolkit. New versions use new IDs. Existing legacy runs remain supported.

What changed

• Strict benchmark runs validate question/script identity, answer keys, exact checkpoint revisions and frozen per-task protocol references.
• Candidate scores and lengths must align. UTF-8 byte lengths are checked, and declared normalization determines the prediction and correctness checks.
• Entire declared model families stay in one split. Calibration input roles must match those assignments, and releases reference held-out validation runs.
• A bank audit reports counts, exclusions, missing translation rows, exact-content duplicates and declared grouping. It does not judge translation quality.
• Response-matrix export retains global IDs, languages/scripts, family splits, passage/overlap groups and exact provenance for an external statistical program.
• Hugging Face transfer includes the new definitions and their dependencies, verifies integrity, and refuses conflicting existing artifacts or sealed runs.

These checks establish file/schema consistency, not scientific validity. Exact prompts are preserved when supplied; token counts and model execution cannot be independently verified without the actual evaluation environment.

Foundation sources

The proposed first version uses Belebele, GlobalPIQA Parallel, and the SIB-200 test set. SOURCES.json records candidate upstream commit pins. The actual language/script intersection, counts, exclusions and FLORES overlap still need an audit. No new translation or non-parallel component is included in this plan.

The earlier Belebele/Global MMLU rephrasing sources remain recorded in LEGACY_REPHRASING_SOURCES.json for earlier work. Source datasets retain their own licenses and usage terms; this archive does not assign a blanket license to future model outputs.

Try the complete synthetic storage example

From the extracted toolkit:

./babel --root new-example benchmark-demo
./babel --root new-example validate
./babel --root new-example bank-audit demo-bank-v1
./babel --root new-example artifacts

This creates two handwritten question identities in two languages, three fictional checkpoints, twelve handwritten responses, and illustrative calibration/release artifacts. No model or IRT fitting program is run. The ZIP also contains an already-built benchmark-demo-store/.

The previous item comparison demonstration remains available for download. It uses synthetic original/rephrased responses. Real strict runs can also be compared item by item under compatible definitions.

Real workflow

1. Assemble and audit the bank; register its definition, task protocols and model panel.
2. Run evaluations using lm-evaluation-harness or another producer, recording the exact prompts, candidates, likelihoods and lengths required by the contract.
3. Import canonical responses, attach original logs/configurations, validate and seal.
4. Export a response matrix. Use a separate program for hierarchical Rasch/2PL fitting, uncertainty, linking checks and held-out-family validation.
5. Register the actual fitting outputs and release definitions, then upload them.

Read BENCHMARK_WORKFLOW.md, ARTIFACTS.md, and FORMAT.md for commands, schemas and limits. VERIFICATION.md records software validation.

The toolkit does not yet run IRT fitting, score future models on a calibrated scale, compute clustered confidence intervals, apply overall release weights, or select/validate short forms. Ordinary accuracy summaries remain ordinary accuracy. The filesystem now preserves the data these analyses need.

Transfer and restore

hub-plan previews registered artifacts, selected sealed runs and calibration run dependencies. hub-push uses the configured public project repository. Command-line uploads use the Hugging Face client's authentication; browser sign-in is separate. Reports in derived/ are not uploaded automatically.

./babel hub-plan
./babel hub-push
./babel hub-pull --repo base-model-evals/off-the-shelf-model-evals \
  --destination /path/to/new-local-archive

A pull resolves one commit, validates the registered files and references, and refuses to overwrite an existing local destination. The Dataset Viewer is disabled because this repository contains heterogeneous archive files rather than one uniform data table.