namkoong-lab/LatentGym_Qwen3-8B_1episode_4Envs_LOO_wordladder

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_1episode_4Envs_LOO_wordladder model is an 8 billion parameter Qwen3-based language model fine-tuned using the GRPO algorithm. It is specifically optimized for in-context meta-learning on held-out word ladder tasks, forming part of the LatentGym testbed. This model demonstrates specialized adaptation to meta-learning environments, particularly for tasks involving latent variable inference and sequential decision-making. Its primary application is in research contexts exploring meta-learning and generalization across varied environmental latents.

Loading preview...

Model Overview

namkoong-lab/LatentGym_Qwen3-8B_1episode_4Envs_LOO_wordladder is an 8 billion parameter model built upon the Qwen/Qwen3-8B base architecture. It has been fine-tuned using the GRPO (Generalized Reinforcement Learning with Policy Optimization) algorithm, specifically for in-context meta-learning on a held-out word ladder environment. This model is a component of the LatentGym testbed, designed to evaluate meta-learning capabilities.

Key Characteristics

  • Base Model: Qwen3-8B, a powerful large language model.
  • Fine-tuning Algorithm: GRPO, indicating a focus on reinforcement learning-based optimization for meta-learning.
  • Meta-learning Focus: Trained to generalize to new tasks by learning from a small number of examples within a single episode (N=1).
  • Environmental Latents: The training involved environments like hangman, secretary, and number_guessing, each with specific latents (e.g., vowel_count_4, inverse_order, set_of_3) to foster robust meta-learning.
  • Hyperparameters: Utilizes AdamW optimizer, a learning rate of 5e-07, and a KL coefficient of 0.04, among other specific settings tailored for GRPO training.

Ideal Use Cases

This model is particularly suited for:

  • Research in Meta-Learning: Exploring how LLMs can adapt to new tasks with minimal examples.
  • Evaluation of Generalization: Testing the ability of models to infer and utilize latent environmental properties.
  • Development of Adaptive Agents: Investigating agents that can quickly learn and perform in novel, but related, environments.

Finetunes
1 models