grskv/kontur-countdown-gemma
The grskv/kontur-countdown-gemma is a 1 billion parameter model based on the Gemma architecture, fine-tuned specifically for generating Countdown-style arithmetic equations. This model excels at creating mathematical expressions that solve for a target number using a given set of digits. Its specialized training pipeline, including Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO), focuses on arithmetic equation generation. It is designed for applications requiring precise and creative numerical problem-solving.
Loading preview...
Overview
The grskv/kontur-countdown-gemma is a 1 billion parameter language model, derived from the Gemma architecture, that has undergone a specialized multi-stage fine-tuning process. Its primary function is to generate arithmetic equations in the style of the "Countdown" game, where a target number must be reached using a set of given numbers and basic mathematical operations.
Key Capabilities
- Countdown-style Equation Generation: Specifically trained to construct valid arithmetic expressions from a given set of numbers to achieve a target result.
- Specialized Fine-Tuning: The model's development involved several stages:
- Stage 03A: Supervised Fine-Tuning (SFT) using data from
verified_Qwen3-4B-Instruct-2507. - Stage 03B: Continued SFT on new 4-number datasets.
- Stage 03C: Further SFT on synthetic data involving 3, 4, 5, and 6 numbers.
- Stage 04: Direct Preference Optimization (DPO) using correct/incorrect data derived from
Qwen3-8BandQwen3-4B-Instruct-2507.
- Stage 03A: Supervised Fine-Tuning (SFT) using data from
Good For
- Arithmetic Problem Generation: Ideal for applications that require the creation of mathematical puzzles or educational tools focused on numerical reasoning.
- Specialized Numerical Tasks: Suitable for use cases where generating specific types of arithmetic equations is crucial, rather than general language understanding.