AdarshSingh7647/Eklav-4B-Math-CotGen

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 26, 2026Architecture:Transformer Featherless Exclusive Cold

AdarshSingh7647/Eklav-4B-Math-CotGen is a 4 billion parameter language model based on Qwen/Qwen3-4B, fine-tuned for mathematical reasoning tasks. It utilizes a CotGen (Chain-of-Thought Generation) baseline, specifically a standard full trace CoT SFT method, to enhance its problem-solving capabilities. This model is designed to excel in math-related benchmarks, demonstrating strong performance across various mathematical datasets. Its primary use case is complex mathematical problem-solving and reasoning.

Loading preview...

Eklav-4B-Math-CotGen Overview

AdarshSingh7647/Eklav-4B-Math-CotGen is a 4 billion parameter model built upon the Qwen/Qwen3-4B base architecture, specifically optimized for mathematical reasoning. This model employs a Chain-of-Thought Generation (CotGen) baseline, which involves standard full trace CoT Supervised Fine-Tuning (SFT).

Key Capabilities and Features

  • Mathematical Reasoning: The model is fine-tuned to handle complex mathematical problems, focusing on generating coherent reasoning steps.
  • CotGen Baseline: It serves as a baseline for the Eklav training approach, which aims to teach models to continue partial reasoning traces rather than imitating full traces end-to-end.
  • Performance on Math Benchmarks: The model demonstrates competitive results on various mathematical benchmarks, including:
    • GSM8K: 93.8%
    • MATH-500: 88.7%
    • AIME 1983-2024: 52.6%
    • Omni-MATH: 34.7%
  • General Knowledge: It also shows solid performance on general knowledge benchmarks like MMLU (78.7%) and GPQA-Diamond (46.0%).

When to Use This Model

This model is particularly well-suited for applications requiring robust mathematical problem-solving and the generation of reasoning steps. It can be a strong candidate for tasks involving:

  • Automated math tutoring systems.
  • Generating solutions for competitive programming math problems.
  • Research into improving mathematical reasoning capabilities of LLMs.

It provides a strong foundation for tasks where a model needs to not just provide an answer, but also demonstrate the logical steps to reach it.

Finetunes
1 models