namkoong-lab/LatentGym_Qwen3-8B_10episodes_MultiLatent_number_guessing

Model Overview

This model, LatentGym_Qwen3-8B_10episodes_MultiLatent_number_guessing, is an 8-billion parameter variant of the Qwen3 architecture, developed by namkoong-lab. It has been fine-tuned using the GRPO (Gradient-based Reinforcement Learning with Policy Optimization) algorithm, specifically for in-context meta-learning within the LatentGym testbed.

Key Capabilities

• In-context Meta-learning: Designed to adapt and learn new tasks or variations within a given environment without explicit retraining.
• Number Guessing Environment: Specialized in the number_guessing environment, which involves inferring numerical patterns or rules.
• Multi-Latent Handling: Trained to manage scenarios with three distinct latent variables (set_of_2, set_of_3, dynamic_range) during its training phase, enhancing its ability to generalize to similar latent structures.
• Episodic Learning: Optimized for tasks that unfold over 10 episodes, focusing on cumulative reward (Σᵢ Gᵢ).

Training Details

The model was fine-tuned from Qwen/Qwen3-8B using specific hyperparameters including an AdamW optimizer with a learning rate of 5e-07, a constant learning rate schedule with warmup, and a KL coefficient β of 0.04. Training involved 20 epochs with a max generation length of 64 tokens and a sampling temperature of 0.8.

Good for

• Research into meta-learning and in-context adaptation for LLMs.
• Experiments within the LatentGym framework, particularly for the number_guessing task.
• Developing agents that need to quickly infer and adapt to underlying latent structures in sequential decision-making problems.