namkoong-lab/LatentGym_Qwen3-8B_10episodes_4Envs_full
The namkoong-lab/LatentGym_Qwen3-8B_10episodes_4Envs_full 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 adapting to new tasks across four distinct environments, demonstrating meta-learning capabilities. This model is part of the LatentGym testbed, focusing on learning from limited episodes within diverse latent spaces.
Loading preview...
Model Overview
This model, LatentGym_Qwen3-8B_10episodes_4Envs_full, is an 8 billion parameter variant of the Qwen3 architecture, developed by namkoong-lab. It has been fine-tuned using the GRPO (Generalized Reinforcement Learning with Policy Optimization) algorithm, specifically designed for in-context meta-learning. The model's training focused on adapting to new tasks across four distinct environments, with 10 episodes per trajectory, making it adept at learning from limited interactions.
Key Capabilities & Training
The model's meta-learning ability is demonstrated across environments such as hangman, wordladder, secretary, and number_guessing. During training, it was exposed to specific latent features within each environment, for example, vowel_count_4 in hangman or hub_word_3letter in wordladder. The training utilized an AdamW optimizer with a learning rate of 5e-07, a constant_with_warmup LR schedule, and a KL coefficient β of 0.04. It was trained for 20 epochs with a max generation length of 64 tokens, emphasizing efficient learning and adaptation.
Use Cases
This model is particularly suited for research in meta-learning, reinforcement learning from human feedback (RLHF) applications, and scenarios requiring rapid adaptation to new tasks with minimal examples. Its focus on learning from diverse latent spaces makes it valuable for developing agents that can generalize across different problem types.