Why this matters
Agentic systems fail in two common ways: invisible chain-of-thoughts that make failures hard to debug, and synthetic training traces that lack runnable verification. This dataset addresses both by shipping reference trajectories with visible per-turn reasoning plus deterministic programmatic verifiers and real sandbox tool executions—so every accepted row is replayable and machine-checkable.
What Sets It Apart
- Verifiable-first design: every accepted row passed deterministic programmatic verifiers and was re-verified by deterministic replay; verifier smoke tests cover positive and negative cases.
- Visible reasoning and tooling: 100% of rows include visible chain-of-thought; 5,997 rows include real sandbox tool calls (fs/sqlite/json/python subprocess) with final-state equality checks and recovery/parallel-call lanes recorded.
- RL/NeMo readiness: dataset ships four loadable configs (canonical, sft_openai, responses_api, rl_tool_prompts) and programmatic reward_contract_json per-row for NeMo Gym / TRL-style rollouts.
- Structured, diverse families: 13 verifiable families (tool calls, constraint instructions, structured outputs, stateful dialogue, verifiable math/science, long-context aggregation, planning, delegation, citations, memory management, code execution), with rows ordered easy→hard within families to support curriculum-style training.
- Practical auditability: full audit view (lossless JSONL), parquet shards, manifest SHAs, and explicit provenance metadata (campaign run ids, seed/attempt data, acceptance attempts).
Who it's for — fit and tradeoffs
Great fit if you train or evaluate agentic LLMs and need replayable, machine-checkable supervision for SFT or RL (GRPO/RLVR/NeMo Gym). The dataset accelerates agent-tool-chain testing, reward-contract development, and debugging thanks to visible CoT and deterministic verifiers.
Look elsewhere if you need large-scale real-user conversational logs, non-synthetic PII-containing data, or long-context corpora dominated by >32K native documents—the long-context family has lower yield and is a small share of rows.
Where it sits in a workflow
Use the sft_openai view for supervised fine-tuning (assistant turns only), responses_api for full reference trajectories and verifier metadata, and rl_tool_prompts for rollout-ready prompts + reward contracts. The raw canonical JSONL is intended for audit, replay, and building custom verifiers or environment servers.
