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

Babel Tower evaluation archive

Shared filesystem for the base-model-evals project. This repository stores the full result for each evaluated question: model responses, available scores, original/rephrased variants, and the settings needed to interpret each run.

The archive is initialized and ready for runs. No measured model results have been added yet. The downloadable toolkit includes a separately labeled synthetic demo; those demonstration records are not published as evaluation runs here.

See an item-level comparison: Download the interactive demo report and open it in your browser. It shows original and rephrased questions, model answers, correctness, and available scores side by side, with search and filters. The demo uses synthetic results and works offline after downloading.

Start here

1. Download the filesystem toolkit and extract it.
2. Open START_HERE.md, or run ./babel from the extracted folder to check local setup.
3. Follow the workflow below to import real JSONL outputs, verify them, and add a sealed run.

The toolkit requires Python 3.10+ on macOS/Linux. Local import, validation, comparison, and CSV/JSONL export work without additional Python packages. Hugging Face transfer support is an optional dependency.

What lives here

File or location	Purpose
store.json	Archive identity, format version, and destination binding.
SOURCES.json	Pinned input datasets, revisions, identities, and interpretation notes.
FORMAT.md	Normalized record schema, provenance, validation, and recovery rules.
RUN_TEMPLATE.json	Example run metadata; replace placeholders with actual evaluation settings.
babel-tower-filesystem.zip	Runnable tools, tests, documentation, and a separate synthetic example.
item-comparison-demo.html	Downloadable, self-contained item comparison using synthetic results.
runs/<run-id>/manifest.json	Created for each real run: settings, file hashes, counts, and status.
runs/<run-id>/raw/	Original JSONL outputs, preserved byte for byte.
runs/<run-id>/records/	Normalized records for comparisons and summaries.
runs/<run-id>/attachments/	Original configs, aggregate harness output, and input snapshots.

Run folders appear when the first run is uploaded. Empty local working folders are recreated by the toolkit on initialization or download.

Project inputs

The existing inputs are versioned separately:

Benchmark input	Pinned revision	Configuration
Global MMLU rephrased	cd6952c6df3e52034886603e36c5ce4796f0777f	all, high_quality, or low_quality; split train
Belebele rephrased	e926d6cc0c50196e0d8bb0d018d9fbf1283dd2e8	all, high_quality, or low_quality; split train

Use each input's id as the persistent question identity and retain language. The all configuration already combines high- and low-quality rows. In these source cards, correct_index refers to the rephrased choices: check the original answer key before comparing correctness. judge_overall_score measures rewrite quality and must not be treated as an evaluated model's accuracy.

Source license declarations are recorded per dataset in SOURCES.json. This archive does not replace the sources' licenses or assign one blanket license to future model outputs.

Add an evaluation run

Enable --log_samples and --output_path when using lm-evaluation-harness. Keep the input's persistent id/question_id in the task document. For example, from the extracted toolkit:

./babel new-run --run-id experiment-001 \
  --model YOUR-MODEL-ID --benchmark belebele \
  --metadata /path/to/completed-run-settings.json

./babel import experiment-001 /path/to/original-samples.jsonl \
  --format harness --variant original
./babel import experiment-001 /path/to/rephrased-samples.jsonl \
  --format harness --variant rephrased

./babel attach experiment-001 /path/to/evaluation-config.yaml /path/to/results.json
./babel validate --run experiment-001
./babel compare experiment-001 --items comparison.html
./babel seal experiment-001
./babel hub-plan --run experiment-001

The paired original/rephrased JSONL format is also supported with --format paired. A canonical format is available for custom producers. Record actual model/dataset revisions, prompts, seeds, harness version, and metric definitions. Missing data remains missing; logging cannot invent scores that the evaluation did not calculate.

Install transfer support and authenticate the command-line client when using it:

python3 -m pip install '.[hub]'
hf auth login
./babel hub-push --run experiment-001

The bundled real archive is bound to this existing public dataset. The upload command sends only registered files for selected sealed runs. The preview lists exact paths and hashes. Browser sign-in and command-line sign-in are separate; the Hugging Face website's file uploader also works with an existing browser session.

Download and verify

python3 -m babel_fs hub-pull \
  --repo base-model-evals/off-the-shelf-model-evals \
  --destination /path/to/new-local-archive

The download pins a commit, verifies registered files, restores local working folders, and refuses to overwrite an existing destination. Add --revision to select a specific commit or tag. Documentation and the toolkit remain available from this repository's Files tab.

Compare individual responses

After importing original and rephrased outputs, generate a report from the extracted toolkit:

./babel compare experiment-001 --items comparison.html
./babel compare experiment-001 --items comparison.csv
./babel compare experiment-001 --items comparison.jsonl

The HTML report opens in a browser and works offline. Search or filter by outcome and language to inspect each question's inputs, answers, correctness, scores, and source file/line side by side. CSV provides one row per paired or unmatched stable item; JSONL retains both full normalized records. File extensions select the format, or supply --format html, --format csv, or --format jsonl with --items. Existing report files are never overwritten.

The command still prints the summary and report metadata. Use --baseline and --variant to select other variants, or add --other-run RUN-ID to compare runs. Reports are local derived files and are not automatically sent by hub-push.

Analysis and integrity

Summaries report total and scored counts separately. Comparisons join stable question identities, language, benchmark, filter, and repeat identifier. They report improvements, regressions, unscored pairs, and missing matches. Item-level reports include unmatched stable items; run-local identities are excluded and counted because their pairing cannot be established reliably. Known conflicting dataset revisions are rejected for cross-run comparisons.

Correctness uses each version's own answer key. Missing correctness stays unscored. The reports display recorded log probabilities without inventing score differences; choice indices alone are not treated as semantic answer agreement.

Raw files are retained alongside normalized files. SHA-256 hashes, duplicate checks, transactional imports, sealed runs, and explicit version metadata protect the archive from accidental corruption. CSV and optional Parquet are derived analysis exports. The automatic Dataset Viewer is disabled because this repository is a heterogeneous experiment archive; metadata files are not benchmark rows.