MaYiding/OracleProto-Experiment-Results

OracleProto Experiment Results

Raw experimental observations, model answers, retrieval audits and analysis artifacts from OracleProto. The companion benchmark dataset contains the fixed evaluation questions; this repository contains the experimental records.

Code · Benchmark · Leaderboard · Paper · 中文说明

Release contents	Count
Run directories	30
SQLite snapshots	155
Source artifacts represented	1,495
Answered samples in the viewer	35,120
Included source size	101.33 GB
Compressed archives and other artifacts	18.06 GB

The collection contains exploratory runs, partial collection attempts, paper panels and information-boundary experiments. Do not pool every run into one benchmark score. Run identifiers, stored prompt/configuration fingerprints and source databases define separate experimental strata. The fixed paper scoring contract is documented in the companion benchmark and harness.

Load the answer index

from datasets import load_dataset

answers = load_dataset(
    "MaYiding/OracleProto-Experiment-Results", "predictions", split="test"
)
print(answers.column_names)

The viewer includes answered run_results samples from the model databases. It exposes answer letters, stored parsing/correctness flags, token counts and latency. It does not recompute scores, treat missing attempts as answers, or flatten the auxiliary audit databases. Full observations and request/retrieval traces remain in the SQLite archives.

Answer-index fields	Meaning
run_directory, database, model	Run, source-relative database and stored model identifier
question_id, sample_index	Original question ID and one-based sample slot
answer_letters, parse_ok	Stored normalized answer and parser flag; an unparsed answer can have null letters
correct_stored	Original correctness flag; null means no stored correctness value, not an incorrect answer
prompt_tokens, completion_tokens, reasoning_tokens	Stored token counts
latency_ms, error	Stored latency in milliseconds and nullable error field

Rows are included when either normalized answer letters or a raw final answer is present. The index is a browsing aid; scoring requires the matching question set and run-specific protocol.

Restore a run

Install huggingface_hub and the zstd command, then download this repository's restore.py:

python restore.py --run 20261002-153314-95cd --output restored-results

Restore one database with a source-relative glob:

python restore.py --match '20261002-153314-95cd/db/gpt-5.3-codex*' --output restored-one

Use --revision <commit> to pin a Hub revision. The helper verifies the compressed SHA-256 and restored-content SHA-256, recreates the run-relative layout, and refuses to overwrite existing files. Gzip source observations are stored with zstd and restored to gzip after their uncompressed content is verified.

Files and provenance

Path	Contents
runs/<run>/db/*.db.zst	Model SQLite snapshots with questions, prompts, run configuration and observations
runs/<run>/*.sqlite.zst	Information-boundary inputs, requests, trials, retrieval caches and audit records
runs/	Sanitized manifests, analysis tables, experiment scripts, logs and scientific plots
data/predictions/	Compact answer index used by the viewer
metadata/artifacts.jsonl	Browseable artifact catalog
metadata/runs.jsonl	Per-run counts and sizes
metadata/release_manifest.json	Source/archive/restored hashes, representations, table row counts and redaction counts
SHA256SUMS	Checksums of published files

Each SQLite snapshot is rebuilt from a private local clone, including committed changes present in its WAL. Table row counts are checked and SQLite quick_check must pass before compression. The original source files are not modified. Source-file hashes identify original bytes; restored hashes identify the sanitized logical export. They need not match, and redacted configuration strings can differ from the original configuration fingerprints preserved in the records.

The referenced workspace revision is a harness reference, not a claim that all runs used that exact code revision. Use each run's stored protocols, manifests and hashes when interpreting comparisons. Credentials must be supplied separately to run new model requests.

Experiment billing and invoices

The invoices viewer configuration contains 42 reported invoice rows. Full billing evidence is available in:

• runs/collection_300/billing/requests.sqlite3.zst: request, usage, batch, claim and historical-usage records.
• runs/paper_metrics_20261006/reported_invoices.csv and .json: reported spend, inclusion flags and invoice status.
• runs/paper_metrics_20261006/ledger_usage_snapshot.jsonl.zst: usage-ledger evidence.
• Analysis workbook: eight original sheets and 601 formulas, preserved byte for byte after a complete OOXML privacy check.

Reported invoices and ledger-based cost estimates are distinct measurements. Duplicate flags, unresolved days, currency units and invoice-confirmation notes are retained. No invoice totals or scientific scores are recomputed in this archive.

Publication scope

API credentials, private service hosts and local filesystem paths are redacted. Experiment billing databases, reported invoices, usage ledgers, the original analysis workbook and sanitized reproduction handoffs are included. Redundant packages, regenerable previews, Git bundles, manuscript editing workspaces and transient SQLite sidecars are excluded. Committed SQLite sidecar data is incorporated into each snapshot. Exclusions are summarized by category without publishing private filenames.

Redaction statistics count text replacements, including repetitions and credential-like snippets in retrieved pages. They do not identify how many distinct or active secrets existed.

This public release is not a byte-for-byte backup of every local working file. Excluded private material remains outside the release. Original-file deletion should be limited to artifacts whose verified public representation meets the owner's preservation needs.

Project-generated code and experimental artifacts follow the project's MIT license. Retrieved third-party text, URLs and model-generated content retain any applicable source/provider terms; the project license does not grant rights owned by third parties. Retrieval traces can quote source material and can contain event outcomes; they are experimental evidence, not a contamination-free training corpus.

Citation

@article{OracleProto,
  title={OracleProto: A Reproducible Framework for Benchmarking LLM Forecasting via Knowledge Cutoffs and Temporal Masking},
  author={Yiding Ma and Chengyun Ruan and Kaibo Huang and Jin Tan and Zhongliang Yang and Linna Zhou},
  journal={arXiv preprint arXiv:2605.03762},
  year={2026}
}