Expert Reasoning SFT
Multi-step solutions written and verified by domain experts.
- Examples
- 250,000+
- Domains
- 24
- Expert authors
- 1,100+
- Avg. tokens
- 1,850
Illustrative rendering of annotation layers.
01Overview
Original problems with step-by-step solutions authored by PhD-level experts in mathematics, physics, chemistry, biology, finance and software engineering. Every item is independently verified by a second expert and checked for contamination against public benchmarks.
Modalities
- Text
- LaTeX
- Code
Use cases
- Supervised fine-tuning for reasoning
- Reward model and verifier training
- Held-out evaluation sets
02Specification
Content
- Domains
- Math, physics, chemistry, biology, finance, law, software and more
- Difficulty
- Graduate to research level, calibrated on 5-point scale
- Languages
- English (primary); Chinese, Japanese, German on request
- Format
- Problem, reasoning trace, final answer, rubric
Quality
- Authoring
- Original, human-written; no LLM-generated solutions
- Verification
- Independent second-expert verification
- Contamination
- N-gram and embedding checks vs. 40+ public benchmarks
03Schema
| Field | Type | Description |
|---|---|---|
| id | string | Unique example identifier. |
| domain | string | Top-level domain and subfield. |
| difficulty | int (1–5) | Calibrated difficulty. |
| prompt | string | Problem statement (Markdown + LaTeX). |
| reasoning | string | Step-by-step expert solution. |
| answer | string | Canonical final answer. |
| rubric | list<Criterion> | Grading criteria for partial credit. |
04File structure
expert-reasoning-sft/
├── data_card.md
├── train/
│ └── part-0000.jsonl
└── eval/
└── heldout.jsonl05Sample
Evaluate before you license.
- 500 examples across all domains
- Data card and annotation guidelines
- Size
- ≈ 6 MB
- Formats
- JSONL, Parquet
06License
Graviti Commercial Data License
Commercial training rights. Exclusive licensing available per domain.
How we license data →Related datasets
Human Preference Pairs
Pairwise and graded comparisons for RLHF, DPO and reward modeling.
- Comparisons
- 400,000+
- Avg. turns
- 4.2
Agentic Tool-Use Trajectories
Human-demonstrated multi-step tasks across browsers, terminals and APIs.
- Trajectories
- 30,000+
- Environments
- 70+
Get started
Need more than Expert Reasoning SFT?
We extend catalog datasets with new environments, tasks, languages and annotation layers — or collect from scratch to your spec.