namkoong-lab/LatentGym_Qwen3-8B_1episode_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

This model is a Qwen3-8B variant, fine-tuned by namkoong-lab 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 adaptation to new environments based on limited training examples. The model leverages its 8 billion parameters and 32768 token context length to perform meta-learning across diverse latent environments like hangman, wordladder, and secretary tasks.

Loading preview...

Model Overview

This model, LatentGym_Qwen3-8B_1episode_4Envs_LOO_number_guessing, is a specialized variant of the Qwen3-8B base model, developed by namkoong-lab. It has been fine-tuned using the GRPO (Generalized Reinforcement Learning with Policy Optimization) algorithm to excel at in-context meta-learning, particularly for held-out number guessing tasks. This model is part of the larger LatentGym testbed, designed to evaluate and develop meta-learning capabilities in large language models.

Key Capabilities

  • In-context Meta-learning: Demonstrates the ability to adapt to new, unseen tasks (held-out number guessing) with only one episode (N=1) of training data.
  • GRPO Fine-tuning: Utilizes a specific reinforcement learning algorithm for optimization, focusing on cumulative reward (Σᵢ Gᵢ).
  • Diverse Environment Exposure: Trained on a variety of latent environments, including hangman, wordladder, and secretary tasks, exposing it to different problem structures and latent variables.
  • Qwen3-8B Base: Built upon the robust Qwen3-8B architecture, providing a strong foundation for language understanding and generation.

Training Details

The model was trained with specific hyperparameters including an AdamW optimizer, a learning rate of 5e-07, and a constant learning rate schedule with warmup. Key parameters like a KL coefficient β of 0.04 and a clip range ε of 0.2 were used, alongside a max generation length of 64 tokens and sampling temperature of 0.8 with top-p 0.95 during training.

Finetunes
1 models