Elfsong/Mystery-Benchmark-Full

Mystery Benchmark full evaluations and GPT-OSS ablations

This gated dataset contains the complete Kimi K2.5 and GPT-OSS 120B full evaluations, plus the GPT-OSS 120B guarded ablation experiment. All runs use the same 1,000 fixed nominal cases. There are 7,000 completed episodes: 4,000 original baseline episodes and 3,000 new GPT-OSS ablation episodes. All 7,000 final trajectories passed completion validation and step replay checks.

The ablation policies are full_minus_accuse, delay_control, and prompt_ledger. See run/gpt_oss_120b/ablations_20261003/RESULTS.md for the results and comparison with the original GPT-OSS vanilla and guarded baselines. The baselines were not rerun in the ablation job.

Contents

• data/episodes.csv: previewable combined index with per-episode metrics and raw/shard paths.
• data/cases.jsonl.gz: all 1,000 source case files used by both models, including ground truth.
• data/trajectories/: 35 shards, including the ablations in gpt_oss_120b/ablations_20261003/; one row per episode with every ordered trajectory record (header, steps, footer).
• raw/{kimi_k25,gpt_oss_120b}/trajectories.tar.gz and raw/gpt_oss_120b/ablations_20261003/trajectories.tar.gz: byte-for-byte original JSONL trajectory files, with original relative paths.
• raw/*/executions/, raw/*/failed_attempts/, and ablation execution/failed-attempt archives: execution ledgers and failed retry artifacts. Kimi also includes the original queue archive.
• run/*/: original summaries, validation, run configurations, performance and source manifests.
• SHA256SUMS: SHA-256 checksums for all published files.

Join episodes to source cases by instance_id and verify using source_sha256. The trajectory value is the path inside trajectory_archive; the corresponding sharded row is in trajectory_shard and can be matched by model, policy, and instance ID.

Interpretation

The nominal 1,000 case files comprise 874 byte-distinct worlds because the benchmark generator can retry seeds. Do not treat all nominal cases as independent world samples. See each model's run/*/final_summary.json for scoring details. Model decoding settings differ between the two runs; paired comparisons should use the supplied fixed case identities and note those settings.

Access to this dataset is gated because raw trajectories and run metadata are included.