yuki99981/codetrace-agent-1000-instances

CodeTrace-Agent dataset

English

This release contains 1,000 executable Python software-engineering environments and one hash-linked execution record per environment. The fixed source revision is b3142c640f1766efa368dec2d1421dc1d90b210f. The sanitized version-of-record archive is published at https://doi.org/10.5281/zenodo.22701046.

Validation status

Observation	Result	Interpretation
Construction-time eligibility	1,000/1,000	All released tasks met the original nop=0 and oracle=1 gates during construction
Later registered primary-host replay	993/1,000	Recommended subset for analyses that require a valid later dual gate
Retained later task-level failures	7/1,000	Preserved failure records; excluded from analyses requiring a valid later dual gate

These are observations made at different times. The construction result does not guarantee that all tasks reproduce on every later host. The later replay had no missing receipts, unresolved infrastructure failures, or evidence-binding errors.

Record types

Type	Count	Meaning
agent_interaction	273	Model messages, tool actions, edits, observations, and linked evaluation records
environment_validation	727	Nop/oracle environment checks without a model decision process

Do not treat the 727 environment-validation records as model-behaviour traces. Stored rewards do not estimate a general model success rate.

Integrity and reuse

Join environments and execution records by instance_id through instance-trajectory-index.json. Check file hashes before analysis. Metadata and JSON records can be studied without Docker; task execution requires Docker and a compatible Harbor environment. The versioned verification tools and later-replay status table are maintained at https://github.com/yuki9984/codetrace-agent-data.

Task environments contain source snapshots from ten upstream open-source projects. Their retained licences and notices govern those files. CC BY 4.0 applies only to team-created metadata, indices, annotations, and documentation; the archive is therefore labelled other at dataset level.

中文

本数据集包含 1,000 个可执行的 Python 软件工程环境，并为每个环境提供一条由哈希绑定的执行记录。固定源版本为 b3142c640f1766efa368dec2d1421dc1d90b210f，脱敏后的正式归档已发布至 Zenodo：https://doi.org/10.5281/zenodo.22701046。

构建阶段共有 1,000/1,000 个任务满足原始 nop=0、oracle=1 门禁；之后登记的主机复跑结果为 993/1,000。需要“后续复跑双门有效”的分析应使用这 993 个任务，另外 7 条任务级失败记录继续保留。两组数字对应不同时间的验证，不能混为同一结论。

数据包含 273 条 agent_interaction 和 727 条 environment_validation。后者记录环境的 nop/oracle 检查，不是模型行为样本。请通过 instance_id 和 instance-trajectory-index.json 连接环境与记录，并在分析前核验清单和文件哈希。