LegumMagister/chsm8-labeler-gemma3-1b
LegumMagister/chsm8-labeler-gemma3-1b is a 1 billion parameter language model, distilled from Gemma-4-31B and fine-tuned from `google/gemma-3-1b-it`. It specializes in generating five natural-language style descriptions for chess game sides, based on moves and measured board features. This model achieves high throughput at 154 games/s on a GH200, significantly faster than its 31B teacher, while maintaining comparable label quality and minimizing contradictions with game data.
Loading preview...
Overview
LegumMagister/chsm8-labeler-gemma3-1b is a specialized 1 billion parameter language model, fine-tuned from google/gemma-3-1b-it. Its primary function is to generate five natural-language style instructions describing how a chess player played, given one side of a chess game (moves and measured board features). This model was distilled from a larger Gemma-4-31B teacher model, leveraging a dataset of 2.40 million clean label sets.
Key Capabilities
- High-throughput Chess Style Labeling: Generates five distinct style descriptions (terse, short, medium, long) for chess game sides at approximately 154 games/s on a single GH200, an 80x speedup over its 31B teacher.
- High Fidelity Distillation: Achieves 94% distillation score, closely matching the teacher's performance in terms of label informativeness and claim grounding.
- Reduced Contradictions: The default 'clean' version of the model contradicts measured board data in only 0.5% of labels, a 9x improvement over the teacher model's rate.
- Specialized Input/Output: Processes plain text input containing chess moves and 16 measured board features, outputting five distinct style labels.
Good for
- Automated Chess Analysis: Ideal for applications requiring rapid, automated generation of descriptive chess style labels.
- Integrating with Chess Engines: Provides meaningful, game-specific labels that can enhance the predictive power of style-conditioned chess models.
- High-Performance Labeling: Suitable for scenarios where high throughput and efficiency are critical, such as processing large datasets of chess games.
- Replacing Larger Models: Offers a significantly faster and more efficient alternative to larger, slower models for this specific task, without substantial loss in quality.