namkoong-lab/LatentGym_Qwen3-8B_10episodes_4Envs_LOO_secretary
The namkoong-lab/LatentGym_Qwen3-8B_10episodes_4Envs_LOO_secretary 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 tasks involving held-out secretary environments within the LatentGym testbed, demonstrating meta-learning capabilities across diverse latent conditions. This model is designed to perform well in scenarios requiring adaptive problem-solving based on limited in-context examples, particularly for tasks like hangman, wordladder, and number guessing.
Loading preview...
LatentGym_Qwen3-8B_10episodes_4Envs_LOO_secretary Overview
This model 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 training focused on a "held-out-secretary" setup, utilizing 3 training environments over 10 episodes.
Key Capabilities & Training Focus
The model's training regimen emphasizes meta-learning across various latent conditions, enabling it to adapt to new tasks based on in-context examples. The training involved specific latent features for different environments:
- Hangman:
vowel_count_4,ending_ABLE - Wordladder:
hub_word_3letter,hub_word_4letter,order_outside_in - Number Guessing:
set_of_3,range_100
Training hyperparameters included an AdamW optimizer with a learning rate of 5e-07, a constant learning rate schedule with warmup, and a KL coefficient \u03b2 of 0.04. The model was trained for 20 epochs with a max generation length of 64 tokens, using a seed of 263.
Good For
- In-context meta-learning research: Particularly for understanding how LLMs adapt to new tasks with limited examples.
- LatentGym testbed experiments: As a component within the LatentGym collection for evaluating meta-learning performance.
- Adaptive problem-solving: In environments similar to hangman, wordladder, or number guessing, where the model needs to infer rules or strategies from context.