AdarshSingh7647/Eklav-14B-Math

TEXT GENERATIONPricing:Input $0.48 / Output $0.96Concurrent Unit Cost:1Model Size:14BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 26, 2026Architecture:Transformer Featherless Exclusive Cold

AdarshSingh7647/Eklav-14B-Math is a 14 billion parameter language model based on Qwen3-14B, specifically fine-tuned for mathematical reasoning tasks. It utilizes the Eklav training method, where the model learns to complete partial reasoning traces provided by a teacher model. This approach allows it to generate its own reasoning conditioned on hints, demonstrating improved performance on various math benchmarks compared to standard CoT distillation.

Loading preview...

Eklav-14B-Math: Enhanced Mathematical Reasoning

AdarshSingh7647's Eklav-14B-Math is a 14 billion parameter model built upon the Qwen3-14B base, uniquely designed for advanced mathematical reasoning. Its core innovation lies in the Eklav training method, which teaches the model to complete a teacher's reasoning process from partial hints, rather than simply imitating full end-to-end traces. This "hint-conditioned SFT" approach enables the model to develop its own reasoning capabilities, conditioned on provided guidance.

Key Capabilities & Differentiators

  • Novel Training Objective: Unlike standard Chain-of-Thought (CoT) distillation, Eklav-14B-Math learns to continue reasoning from an incomplete trace, fostering more robust problem-solving.
  • Improved Math Performance: The model shows an average +2.5% pass@1 improvement across six math benchmarks (AIME, GPQA-Diamond, GSM8K, MATH-500, Omni-MATH) compared to traditional full-trace CoT SFT using the same base model and training data.
  • Specialized for Math: Explicitly trained and optimized for mathematical tasks, making it a strong candidate for applications requiring precise numerical and logical deduction.

Performance Highlights

Eklav-14B-Math demonstrates notable gains in specific math benchmarks:

  • AIME 2025: Achieved 33.3% pass@1, a significant increase over CotGen's 26.7%.
  • GPQA-Diamond: Scored 59.1% pass@1, outperforming CotGen's 55.9%.
  • GSM8K: Reached 96.4% pass@1, slightly better than CotGen's 95.9%.
  • MATH-500: Posted 92.0% pass@1, an improvement over CotGen's 90.5%.

This model is particularly well-suited for tasks demanding sophisticated mathematical problem-solving and reasoning generation.

Finetunes
2 models