The dataset captures real, end-to-end GLM 5.2 sessions rather than synthetic summaries, so each row preserves the agent's causal exploration, tool arguments, intermediate results, corrections, and the single next assistant step used as the training target. That makes it especially useful when you need verifiable agent behavior traces for SFT or distillation rather than handcrafted demonstrations.
What Sets It Apart
- Real model trajectories: every row is a cumulative prefix of an authentic GLM 5.2 session (207 accepted trajectories → 1,821 next-step rows), not human-curated pseudo-dialogue. This preserves natural error-recovery and multi-step planning patterns.
- One next step per row: each row supervises only the final assistant message in the prefix, avoiding over-weighting early turns during cumulative-prefix training.
- Verification-first collection: traces pass deterministic runtime, test, and safety checks and an independent Codex review; sessions include deterministic tool-call and protected-file checks.
- Runtime-normalized and scrubbed: plumbing/UI artifacts and credentials are removed or canonicalized while causal context is retained; distributed as parquet and JSONL under CC BY 4.0.
Who It's For and Tradeoffs
Great fit if you are training or evaluating coding/agent SFT pipelines, distillation workflows, or research on multi-step tool use, state tracking, and build-test-fix loops. The dataset is compact and high-fidelity (1K–10K rows), so it complements larger human or synthetic corpora rather than replacing them. Look elsewhere if you need large-scale human-written demonstrations, multi-language natural dialogs, or datasets focused on non-agent NLP tasks. The data reflects behaviors of GLM 5.2 and may carry model-specific biases.
Where It Fits
Use this as a verified trace corpus to: bootstrap agent instruction-following models, analyze tool-selection and parallel-call strategies, or distill reasoning patterns into smaller student models. It pairs well with larger code corpora and human-labeled evaluation sets for robustness testing.
