namkoong-lab/LatentGym_Qwen3-8B_10episodes_4Envs_LOO_number_guessing

TEXT GENERATIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 13, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The namkoong-lab/LatentGym_Qwen3-8B_10episodes_4Envs_LOO_number_guessing 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 solving held-out number guessing tasks within the LatentGym testbed, demonstrating meta-learning capabilities across various environments. This model is designed to generalize to new tasks by learning from a limited number of episodes, making it suitable for adaptive problem-solving scenarios.

Loading preview...

Model Overview

This model, namkoong-lab/LatentGym_Qwen3-8B_10episodes_4Envs_LOO_number_guessing, 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 for in-context meta-learning on held-out number guessing tasks. The model is part of the LatentGym testbed, which focuses on evaluating and developing meta-learning capabilities in LLMs.

Key Capabilities

  • In-context Meta-learning: Demonstrates the ability to learn new tasks from a few examples within the prompt itself.
  • Generalization to Held-out Tasks: Trained on a diverse set of environments (hangman, wordladder, secretary) to generalize to unseen number guessing scenarios.
  • GRPO Fine-tuning: Leverages a reinforcement learning approach to optimize for cumulative reward across episodes.

Training Details

The model was fine-tuned from Qwen/Qwen3-8B over 20 epochs with a learning rate of 3e-07, using AdamW optimizer. Key hyperparameters include a KL coefficient β of 0.04 and a clip range ε of 0.2. It processes 8 rollouts per prompt and 10 episodes per trajectory, aiming to maximize cumulative reward (Σᵢ Gᵢ).

Good for

  • Research into meta-learning and in-context learning for LLMs.
  • Developing agents that can quickly adapt to new, unseen tasks with limited demonstrations.
  • Exploring the effectiveness of GRPO in enhancing generalization capabilities of large language models.

Finetunes
1 models