JEE-SFT is a multimodal instruction-tuning dataset designed to teach Vision Language Models step-by-step reasoning for complex STEM problems. The dataset focuses on the solution process and includes both Multiple Choice Questions and Numerical Value Questions, with a key feature being its 'Text-Only Reasoning' filter. It was created by author farhananis005 and was last updated on February 4, 2026.
Use Cases
- Instruction-tuning VLMs for step-by-step problem solving based on the described focus on the solution process.
- Training models on STEM reasoning tasks based on the inclusion of Multiple Choice and Numerical Value Questions.
- Filtering model outputs for text-based reasoning based on the dataset's 'Text-Only Reasoning' feature.
Strengths
- Dataset is specifically designed for teaching step-by-step reasoning, unlike versions focused only on final answers.
- Includes a 'Text-Only Reasoning' filter as a key feature for model training.
Limitations
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count is unknown, which may limit suitability assessment.
- Description metadata is limited; actual data quality requires manual inspection after download.
Provenance
- Source
- farhananis005 on Hugging Face
- Freshness
- Last updated 2026-02-04 04:40:54; freshness should be verified.