namkoong-lab/LatentGym_Qwen3-8B_10episodes_4Envs_LOO_hangman
The namkoong-lab/LatentGym_Qwen3-8B_10episodes_4Envs_LOO_hangman model is an 8 billion parameter Qwen3-based language model, fine-tuned using the GRPO algorithm for in-context meta-learning. It is specifically optimized for tasks within the LatentGym testbed, demonstrating proficiency in adapting to held-out hangman environments. This model excels at learning from limited episodes to solve new, related problems, making it suitable for research in meta-learning and few-shot adaptation.
Loading preview...
Model Overview
This model, LatentGym_Qwen3-8B_10episodes_4Envs_LOO_hangman, is an 8 billion parameter variant of the Qwen3 architecture, developed by namkoong-lab. It has been specifically fine-tuned using the GRPO (Generalized Reinforcement Learning with Policy Optimization) algorithm to enhance its capabilities in in-context meta-learning. The primary focus of this training was on adapting to held-out hangman environments, making it a specialized tool within the LatentGym testbed.
Key Capabilities
- In-context Meta-learning: Demonstrates the ability to learn and adapt to new tasks with minimal examples, particularly within the hangman game context.
- GRPO Fine-tuning: Utilizes a sophisticated reinforcement learning approach for optimization, allowing for robust policy learning.
- LatentGym Integration: Designed as part of a research testbed, providing a platform for exploring meta-learning and generalization across various environments.
Training Details
The model was trained on 10 episodes across 4 environments, with specific latents seen during training for wordladder, secretary, and number_guessing tasks. Key hyperparameters include a learning rate of 5e-07, a KL coefficient β of 0.04, and a max generation length of 64 tokens, indicating a focus on concise and effective responses for task completion.