Datasets:
IndustrialJudgeBench: trajectory collection
Public release 0.1.1 contains the unchanged version 0.1 engineering-agent trajectories for students building an agentic-error judge benchmark. It contains 5,304 attempt records, with 5,301 usable event streams. Gold TJB annotations have not been assigned.
Contents
artifacts/judge-traces-2026-11-12.zip: the verified 871 MB handoff, including original-format trace exports, normalized event references, prompts, recorded outcomes, configuration context, provenance, prior reviews and reader tools.data/runs.jsonl: one input-index row per attempt. Event and context paths refer to members inside the archive, not separately uploaded Hub files.data/outcomes.jsonl: recorded checker outcomes and validity information, linked by run ID. These are observations and source flags, not gold error labels.annotation/: a blank record template and annotation guidance.metadata/: coverage, release checksum and the credential-pattern scan report.
The runs and outcomes configurations both use the split name corpus.
No train/dev/test assignment has been made. JSON-valued index columns are encoded
as strings to preserve consistent types across different harnesses. Detailed
native records remain inside the archive.
Coverage
| Source | Attempt records |
|---|---|
| Earlier Claude/GPT pilots, C9–C16 | 600 |
| Luna original and two repetitions | 3,138 |
| Sol, three passes | 1,371 |
| Qwen stand runs | 195 |
| Total | 5,304 |
Counts describe recorded launches, not distinct tasks or confirmed inferences. Zero-token launches, interrupted runs and repetitions remain separate. Three launches have empty trace files. Four other files contain a malformed or truncated JSONL line; the archive preserves source text and parse-gap records. Historical control/recheck rows are supplied as reference evidence and are not counted as agent attempts. Unstarted campaign skips are recorded separately.
Six third-Sol attempts are marked INVALID_INFRASTRUCTURE because a restored CLI symlink prevented tool execution. Other source validity and task/oracle warnings remain intact. A failed checker does not establish an agentic error, and a passing run may contain intermediate errors that were later recovered.
Download and inspect
The dataset page is public. File downloads require a Hugging Face account and
manual approval from the repository owner. Students should request access on the
dataset page. After approval, log in with hf auth login, then use:
hf download Lollykonkon/IndustrialJudgeBench --repo-type dataset --include "artifacts/*" --local-dir IndustrialJudgeBench
python -m zipfile -e IndustrialJudgeBench/artifacts/judge-traces-2026-11-12.zip IndustrialJudgeBench/trace-packets
python IndustrialJudgeBench/trace-packets/tools/verify.py
python IndustrialJudgeBench/trace-packets/tools/read_packet.py RUN_ID
Choose RUN_ID from the input index. The reader prints recorded input context and
events without automatically loading outcome labels. Read the archive's README,
CONTEXT_LIMITS.md and TOOLS.md before annotation. The archive SHA-256 is
26a6c30cd1ad84ed9e2a383bd7907275b870970fe194f8a8dd10230ec08e1f02.
All 25,567 packaged file hashes and 25,568 ZIP-member CRCs were verified.
Annotation and evaluation
Confirm the actual TJB schema/version before filling category identifiers. Keep observed events, persistence/recovery, trial validity and responsibility hypotheses separate. Prior judge reports and the proposed engineering taxonomy are reference material, not adjudicated ground truth. Preserve independent reviewer annotations and disagreements before adjudication.
Keep related tasks, source variants, models and repetitions together when forming future data splits. Existing historical-source candidates and conservative groups need review; they do not prove task-text or evaluator equivalence. Embedded task and tool content may reveal evaluation information, so this collection does not establish a label-blind evaluation.
Context limits
Codex and Claude exports contain the available harness messages and tool events. DMCP/Qwen provides tool traces and harness logs; complete model conversations are not established by those artifacts. Missing final artifacts or messages remain evidence gaps. The archive is not a simulator installation or a complete task workspace bundle. See CONTEXT_LIMITS.md for annotation implications.
Privacy, provenance and licensing
Account identity fields were redacted in exported copies. The known-credential text scan has no unresolved findings; it is not image inspection or a guarantee against unknown secret formats. No credential store, environment file or global session directory was collected. Original and exported hashes are preserved. Owned locked traces were copied with original permissions restored and independently checked. No model or judge inference was started to prepare this collection.
Source material retains source-specific licensing terms; a dataset-wide public redistribution license has not been established. The repository page is public for benchmark construction; dataset file access is gated with manual approval. Model outputs, checker verdicts and earlier reviews retain their recorded provenance and are not newly validated gold labels.
- Downloads last month
- 5,528