The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Error code: FileFormatMismatchBetweenSplitsError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Seer Agent Teacher Traces
Research dataset for namezhenzhang/seer-oracle-replay. Contains compact replay inputs, immutable source contracts, complete-schedule reference results and original Teacher trace archives.
Scope: TMax-27B on Terminal-Bench 2.1 (267 attempts, 253 normal); DeepSWE-Preview 32B on SWE-bench Verified (2,000 attempts, 1,998 generation-complete). The two infrastructure failures and 14 execution exceptions remain in the archive and are explicitly separated from replay.
The default compact download uses one compressed bundle (approximately 11 MB unpacked) and supports all scheduling/plot reproduction without GPUs, Docker, Slurm, GCS credentials or base model weights. Optional raw shards total approximately 10.34 GB of original compressed archives; RAW_ARCHIVE_INDEX.json identifies every archive and its immutable source hash. MANIFEST.json provides file-level hashes.
Results estimate reserved resources under fixed trace timings and zero pause/offload/reload cost. CPU/RAM are sandbox quotas, GPU-hours are shared-service reservation time, and storage is compressed OCI cache only. They are not measured deployed speedups or physical utilization savings. CPU/RAM minima are analytical in this model; GPU/cache values are best-found feasible schedules, not global optima. DeepSWE excludes downstream official grading.
The HF provenance and reference metadata may contain original machine paths as historical identifiers; no original machine is needed to replay. Upstream benchmark/harness terms remain applicable. No model weights, Docker image payloads, Python environments, credentials or unrelated chat/session export are included.
See the GitHub README and METHODS/HANDOFF documents for exact commands and next experiments. Always pin the HF commit for reported results.
Reproduce
Code: https://github.com/namezhenzhang/seer-oracle-replay
Use the full HF revision in GitHub DATA_RELEASE.json. seer-download fetches REPLAY_BUNDLE.json and one compact archive, verifies SHA256 and safely unpacks it. Full raw shards are optional via --with-raw. MANIFEST.json covers replay/reference/raw payloads; REPLAY_BUNDLE.json covers the compact delivery archive; CODE_MANIFEST.json covers the code mirror.
- Downloads last month
- 191