Shaik1903/ThinkLess-2B-SFT
Shaik1903/ThinkLess-2B-SFT is a 2.3 billion parameter instruction-tuned causal language model, fine-tuned from Qwen3.5-2B. It specializes in efficient reasoning, achieving significantly higher accuracy and reduced reasoning length (43-72%) compared to its base model, particularly in mathematical and general reasoning tasks. This model is optimized for high-accuracy problem-solving, including competition-level math, by leveraging shortest correct solutions from a 9B teacher model.
Loading preview...
ThinkLess-2B-SFT Overview
ThinkLess-2B-SFT is the Supervised Fine-Tuning (SFT) stage of the ThinkLess-2B family, built upon the Qwen3.5-2B base model. This 2.3 billion parameter model is specifically fine-tuned on its own shortest correct solutions, with assistance from a same-family 9B teacher model for more challenging problems. Its primary goal is to deliver high accuracy while drastically cutting down reasoning length.
Key Capabilities
- Enhanced Accuracy: Achieves significant accuracy improvements across various benchmarks, including +4.9% on GSM8K, +6.1% on MATH-500, and +10.6% on GPQA-Diamond compared to its Qwen3.5-2B base.
- Efficient Reasoning: Reduces reasoning token length by 43-72% (e.g., GSM8K reasoning tokens cut by 72%, MATH-500 and GPQA by 43%) while maintaining or improving accuracy.
- Competition-Level Math: Demonstrates strong performance in complex mathematical problems, matching or exceeding the base model on benchmarks like HMMT Feb 2025.
- Optimized for Tight Budgets: Excels under hard thinking token limits, outperforming the base model significantly at 2k, 4k, 8k, and 16k token budgets for benchmarks like GSM8K and MATH-500.
Good For
- High-Accuracy Problem Solving: Ideal for applications requiring the most accurate answers, especially in mathematical and general reasoning contexts.
- Resource-Constrained Environments: Suitable for scenarios where efficient reasoning and reduced token usage are critical, due to its ability to provide accurate solutions with shorter reasoning paths.
- Competitive Math and Science: A strong candidate for tasks involving complex math and science problems where precision and efficiency are paramount.