namkoong-lab/LatentGym_Qwen3-8B_10episodes_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_10episodes_4Envs_LOO_wordladder 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 word ladder tasks within the LatentGym testbed, demonstrating adaptation to new environments based on learned latent patterns. This model excels at tasks requiring generalization from a limited set of training environments to novel, unseen scenarios.

Loading preview...

Model Overview

This model, LatentGym_Qwen3-8B_10episodes_4Envs_LOO_wordladder, 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 within the LatentGym testbed. The primary focus is on demonstrating the model's ability to generalize to held-out word ladder environments after training on a limited set of related tasks.

Key Capabilities

  • In-context Meta-learning: Designed to learn and adapt to new tasks or environments based on patterns observed during training, without explicit re-training.
  • Generalization to Unseen Environments: Optimized to perform well on "held-out" word ladder tasks, indicating its capacity to apply learned strategies to novel scenarios.
  • GRPO Fine-tuning: Leverages a reinforcement learning approach for fine-tuning, which is particularly effective for sequential decision-making and adaptation tasks.

Training Details

The model was fine-tuned from Qwen/Qwen3-8B using specific hyperparameters including an AdamW optimizer, a learning rate of 5e-07, and a constant learning rate schedule with warmup. Training involved 20 epochs with 10 episodes per trajectory, a max generation length of 64, and sampling parameters of T=0.8 and top-p=0.95. The reward function was based on the cumulative sum of Gᵢ.

Good For

  • Research into meta-learning and generalization in large language models.
  • Experiments requiring models to adapt to new, unseen tasks with minimal examples.
  • Applications within sequential decision-making or problem-solving domains where in-context learning is crucial.

Finetunes
1 models