Complete multi-hop research trajectories capturing expert reasoning from plan to resolution.
Available Now · Sample Available
Deep Research
End-to-end expert research trajectories: plans, searches, source conflicts, and resolutions.
Volume
0
Volume
Scalable Volume
Coverage
Finance, Coding, Physics, Legal, Medical, Mechanical Engineering
0.0
%
Opus 5
Targets
BrowseComp, GAIA, DeepResearch Bench
Modality & Delivery
Text→Text
Formats
JSONL (.jsonl).sft
What You Receive
Coverage
Package Structure
Differentiators
- Captures the plan, searches, source evaluation, conflict resolution, and unresolved gaps as one end-to-end trajectory.
Record Schema
| Field | Description |
|---|---|
action_mask | Optional mask identifying trainable or scoreable actions in the trajectory. |
answer_aliases | Accepted alternative forms of the answer. |
applicable_time_window | Time period within which the answer and evidence are applicable. |
citation_dpo | Citation-level preference data and sentence attribution. |
competencies_exercised | Research and reasoning competencies evaluated by the task. |
conflicts | Conflicting claims, resolution strategy, and unresolved conflicts. |
contamination_check | Assessment of whether the task can be answered without external search. |
cross_document_requirement | Score indicating how strongly the task requires evidence across documents. |
difficulty_calibration | Estimated steps and baseline accuracy measures used to calibrate difficulty. |
dpo | Chosen and rejected trajectories with preference source and margin. |
evidence | Evidence sources with titles, URLs, excerpts, credibility, primary-source status, and conclusion usage. |
evidence_utility | Aggregate utility score for collected evidence. |
format_penalty | Penalty applied for format violations. |
format_valid | Whether the response satisfies required output formatting. |
gain_beyond_rag | Estimated performance gain beyond a retrieval-augmented generation baseline. |
grading_rubric | Rubric items and search-based, evolving, and variance settings used for grading. |
grpo | Group-relative optimization metadata and reward statistics. |
id | Stable identifier for the RL research record. |
multi_step_necessity | Score indicating how strongly the task requires multi-step reasoning. |
noise_info | Noise sources and severity associated with the research task. |
outcome_reward | Overall numerical reward assigned to the completed trajectory. |
process_rewards | Step-level rewards covering information gain, query alignment, redundancy, execution success, and value estimates. |
question | Deep-research question presented to the agent. |
reward_type | Reward formulation used for the RL record. |
safety_labels | Risk, harmful-content, PII, browsing, disclaimer, and refusal annotations. |
task_type | Research task classification. |
timestamp | Timestamp associated with record creation or execution. |
tokenizer | Tokenizer name, version, and rendering configuration. |
tool_definitions | Tools available to the agent, including schemas, limits, and timeouts. |
trajectory | Ordered research steps containing thoughts, actions, and observations. |
wrong_paths | Known unproductive queries or reasoning paths and explanations of their errors. |
Results
| Model | Metric | Result |
|---|---|---|
| Opus 5 | Pass@1 |
Methodology & Quality
Methodology
SME involvement
Domain-expert curated with up to L2 review
Stressors
Multi-hop search, conflicting sources, evidence synthesis, and unresolved uncertainty.
Design criteria
- Preserve complete multi-hop research traces rather than final answers alone.
Build assumptions
Quality approach
- Legal sample trajectories include rubric-based rewards and retained comparison groups.
Known limitations
Customization & Scalability
Custom sets available. Additional volume in progress