namkoong-lab/LatentGym_Qwen3-8B_10episodes_MultiLatent_hangman
The namkoong-lab/LatentGym_Qwen3-8B_10episodes_MultiLatent_hangman model is an 8 billion parameter Qwen3-based language model, fine-tuned using GRPO for in-context meta-learning on the hangman game. It is specifically optimized to learn and adapt to latent game properties like 'vowel_count_4', 'has_double_letter', and 'starts_with_S' over 10 episodes. This model is designed for research into meta-learning capabilities within LLMs, particularly for tasks requiring adaptive strategy based on implicit environmental cues.
Loading preview...
Model Overview
This model, namkoong-lab/LatentGym_Qwen3-8B_10episodes_MultiLatent_hangman, is an 8-billion parameter variant of the Qwen3 architecture, specifically fine-tuned for in-context meta-learning within the hangman environment. It is a component of the LatentGym testbed, which focuses on evaluating and developing LLMs' ability to infer and adapt to latent environmental variables.
Key Capabilities
- Meta-learning in Hangman: The model has been trained to adapt its strategy over 10 episodes of the hangman game.
- Latent Variable Adaptation: It demonstrates adaptation to specific latent game properties, including
vowel_count_4,has_double_letter, andstarts_with_S, which were present during its GRPO fine-tuning. - GRPO Fine-tuning: Utilizes the GRPO (Generalized Reinforcement Learning with Policy Optimization) algorithm for training, with specific hyperparameters like a learning rate of 5e-07 and a KL coefficient of 0.04.
Good For
- Research in Meta-learning: Ideal for researchers exploring how LLMs can infer and leverage hidden environmental states for improved performance.
- Adaptive Agent Development: Useful for understanding and building language models that can adapt their behavior based on limited in-context information.
- LatentGym Testbed Exploration: Serves as a specific example within the LatentGym collection for studying meta-learning in game environments.